The ability to perform multi-modal multi-hop reasoning by iteratively integrating information across various modalities and external knowledge is critical for addressing complex real-world challenges. However, existing Multi-modal Large Language Models (MLLMs) are predominantly limited to single-step reasoning, as existing benchmarks lack the complexity needed to evaluate and drive multi-hop abilities. To bridge this gap, we introduce MMhops, a novel, large-scale benchmark designed to systematically evaluate and foster multi-modal multi-hop reasoning. MMhops dataset comprises two challenging task formats, Bridging and Comparison, which necessitate that models dynamically construct complex reasoning chains by integrating external knowledge. To tackle the challenges posed by MMhops, we propose MMhops-R1, a novel multi-modal Retrieval-Augmented Generation (mRAG) framework for dynamic reasoning. Our framework utilizes reinforcement learning to optimize the model for autonomously planning reasoning paths, formulating targeted queries, and synthesizing multi-level information. Comprehensive experiments demonstrate that MMhops-R1 significantly outperforms strong baselines on MMhops, highlighting that dynamic planning and multi-modal knowledge integration are crucial for complex reasoning. Moreover, MMhops-R1 demonstrates strong generalization to tasks requiring fixed-hop reasoning, underscoring the robustness of our dynamic planning approach.

Masked autoencoders have been extended beyond images, but random masking often fails to capture non-uniform motion regions in inherently disordered and irregular data like point cloud videos. In this paper, we propose a motion-guided disentanglement method (MGD) to improve masked autoencoders for point cloud video representation learning. Specifically, we begin by estimating motion intensity using an optimal transport approach, which guides the separate masking of dynamic and static regions. This motion-guided masking ensures balanced coverage, addressing the limitations of random masking in capturing non-uniformly distributed motion regions. Furthermore, we disentangle the prediction tasks into motion prediction for high-motion point tubes and appearance reconstruction for low-motion ones. This disentanglement enables the model to more effectively capture both motion and appearance in point cloud videos. We conducted experiments on four widely used point cloud video datasets—NTU RGB+D, MSR-Action3D, NvGesture, and SHREC’17—which demonstrate that our approach consistently improves masked autoencoders for point cloud video representation learning, achieving new state-of-the-art results. Code will be publicly available on GitHub.

Balancing training on long-tail data distributions remains a long-standing challenge in deep learning. While methods such as re-weighting and re-sampling help alleviate the imbalance issue, limited sample diversity continues to hinder models from learning robust and generalizable feature representations, particularly for tail classes. In contrast to existing methods, we offer a novel perspective on long-tail learning, inspired by an observation: datasets with finer granularity tend to be less affected by data imbalance. In this paper, we investigate this phenomenon through both quantitative and qualitative studies, showing that increased granularity enhances the generalization of learned features in tail categories. Motivated by these findings, we propose a method to increase dataset granularity through category extrapolation. Specifically, we introduce open-set fine-grained classes that are related to existing ones, aiming to enhance representation learning for both head and tail classes. To automate the curation of auxiliary data, we leverage large language models (LLMs) as knowledge bases to search for auxiliary categories and retrieve relevant images through web crawling. To prevent the overwhelming presence of auxiliary classes from disrupting training, we introduce a neighbor-silencing loss that encourages the model to focus on class discrimination within the target dataset. During inference, the classifier weights for auxiliary categories are masked out, leaving only the target class weights for use. Extensive experiments on three standard long-tail benchmarks demonstrate the effectiveness of our approach, notably outperforming strong baseline methods that use the same amount of data. The project is available at shizhen-zhao.github.io/LT_Neighbors/.

Guangze Gao, Zixuan Li, Chunfeng Yuan, Jiawei Li, Wu Jianzhuo, Yuehao Zhang, Xiaolong Jin, Bing Li, Weiming Hu. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

LLM-to-NAS is a promising field at the intersection of Large Language Models (LLMs) and Neural Architecture Search (NAS), as recent research has explored the potential of architecture generation leveraging LLMs on multiple search spaces. However, the existing LLM-to-NAS methods face the challenges of limited search spaces, time-cost search efficiency, and uncompetitive performance across standard NAS benchmarks and multiple downstream tasks. In this work, we propose the Reflective Zero-cost NAS (RZ-NAS) method that can search NAS architectures with humanoid reflections and training-free metrics to elicit the power of LLMs. We rethink LLMs’ roles in NAS in current work and design a structured, prompt-based to comprehensively understand the search tasks and architectures from both text and code levels. By integrating LLM reflection modules, we use LLM-generated feedback to provide linguistic guidance within architecture optimization. RZ-NAS enables effective search within both micro and macro search spaces without extensive time cost, achieving SOTA performance across multiple downstream tasks.

Graph anomaly detection (GAD) is a critical task with applications in domains such as networking, finance, and bioinformatics. % However, the scarcity of labeled anomalies and the limitations of unsupervised methods hinder effective detection. % While semi-supervised and few-shot learning approaches offer improvements, they struggle with knowledge transfer and rely heavily on labeled data. % Recent advancements in prompt tuning on graphs provide a promising direction, but their application to heterophilous graphs in anomaly detection remains underexplored. % In this work, we propose AffinityTune, a novel framework for few-shot graph anomaly detection based on prompt tuning. % Our approach introduces a unified task framework grounded in affinity judgment, incorporating multi-granularity tasks (''node vs. adjacent subgraph'' and ''node vs. community'') to learn diverse anomaly priors. % For downstream tasks, we reformulate binary classification into a ''node vs. global representation'' affinity judgment, mapping anomalies to a hypersphere in feature space. % We also design learnable task heads to synergize multiple tasks and propose FlexPrompt, a flexible prompt-tuning strategy for fine-grained adaptation to downstream tasks. % Extensive experiments on real-world and synthetic datasets demonstrate that AffinityTune significantly outperforms existing methods in detection efficacy. % Our contributions include a unified task framework, multi-granularity tasks, and the FlexPrompt strategy, offering a parameter-efficient and adaptable solution for few-shot graph anomaly detection.

Brain-inspired spiking neural networks (SNNs) provide energy-efficient computation through event-driven processing. However, the shared weights across multiple timesteps lead to serious temporal feature redundancy, limiting both efficiency and performance. This issue is further aggravated when processing static images due to the duplicated input. To mitigate this problem, we propose a parameter-free and plug-and-play module named Mutual Information-based Temporal Redundancy Quantification and Reduction (MI-TRQR), constructing energy-efficient SNNs. Specifically, Mutual Information (MI) is properly introduced to quantify redundancy between discrete spike features at different timesteps on two spatial scales: pixel (local) and the entire spatial features (global). Based on the multi-scale redundancy quantification, we apply a probabilistic masking strategy to remove redundant spikes. The final representation is subsequently recalibrated to account for the spike removal. Extensive experimental results demonstrate that our MI-TRQR achieves sparser spiking firing, higher energy efficiency, and better performance concurrently with different SNN architectures in tasks of neuromorphic data classification, static data classification, and time-series forecasting. Notably, MI-TRQR increases accuracy by \textbf{1.7\%} on CIFAR10-DVS with 4 timesteps while reducing energy cost by \textbf{37.5\%}. Our codes are available at https://github.com/dfxue/MI-TRQR.

Recognizing Chinese entities in low-resource settings is a challenging but promising task, which extracts structured pre-defined entities and corresponding types from unstructured text. Compared with the prosperous Named Entity Recognition (NER) methods for Indo-European languages, such as English, the research on Chinese NER is still in its infancy. The main obstacles to the development of Chinese NER methods include the ambiguity of Chinese entity boundary recognition and limited data resources. To address these issues, in this paper, a word-segmentation-based model is present for few-shot Chinese NER. First, we enumerate all possible candidate entity spans on the character level for accurate entity boundary identification with the proposed word segmentation and combination strategy. Then, one kind of question-answer-based prompt template loaded with the candidate entity spans is proposed to cast entity extraction into the masked token prediction task, for dealing with the low-data problem by taking full advantage of the generality and transferability of the pre-trained language model. The extensive experimental results show that our method outperforms the state-of-the-art baselines in low-data settings and also achieves comparable performance in full-data settings.

Learning on temporal graphs has attracted tremendous research interest due to its wide range of applications. Some works intuitively merge graph neural networks (GNNs) and recurrent neural networks (RNNs) to capture structural and temporal information, and recent works propose to aggregate information from neighbor nodes in local subgraphs based on message passing or random walks. These methods produce node embeddings from a global or local perspective and ignore the complementarity between them, thus facing limitations in capturing complex and entangled dynamic patterns when applied to diverse datasets or evaluated by more challenging evaluation protocols. To address the issues, we propose the global and local embedding network (GLEN) for effective and efficient temporal graph representation learning. Specifically, GLEN dynamically generates embeddings for graph nodes by considering both global and local perspectives using specially designed modules. Then, global and local embeddings are combined by a devised cross-perspective fusion module to extract high-order semantic relations of node embeddings. We evaluate GLEN on multiple real-world datasets and apply more stringent evaluation procedures. Extensive experimental results demonstrate that GLEN outperforms other baselines in both link prediction and dynamic node classification tasks. Moreover, with concise and effective modules, GLEN can achieve a better balance between inference precision and training efficiency.

Heterogeneous graph neural networks (GNNs) have been successful in handling heterogeneous graphs. In existing heterogeneous GNNs, meta-path plays an essential role. However, recent work pointed out that a simple homogeneous graph model without a meta-path can also achieve comparable results, which calls into question the necessity of a meta-path. In this article, we first present the intrinsic difference between meta-path-based and meta-path-free models, i.e., how to select neighbors for node aggregation. Then, we propose a novel framework to utilize the rich type of semantic information in heterogeneous graphs comprehensively, namely, hybrid aggregation for heterogeneous GNNs (HAGNNs). The core of HAGNN is to leverage the meta-path neighbors and the directly connected neighbors simultaneously for node aggregations. HAGNN divides the overall aggregation process into two phases: meta-path-based intratype aggregation and meta-path-free intertype aggregation. During the intratype aggregation phase, we propose a new data structure called a fused meta-path graph and perform structural semantic aware aggregation on it. Finally, we combine the embeddings generated by each phase. Compared with existing heterogeneous GNN models, HAGNN can take full advantage of the heterogeneity in heterogeneous graphs. Extensive experimental results on node classification, node clustering, and link prediction tasks show that HAGNN outperforms the existing modes, demonstrating the effectiveness and efficiency of HAGNN.

Diverse video captioning aims to generate a set of sentences to describe the given video in various aspects. Mainstream methods are trained with independent pairs of a video and a caption from its ground-truth set without exploiting the intra-set relationship, resulting in low diversity of generated captions. Different from them, we formulate diverse captioning into a semantic-concept-guided set prediction (SCG-SP) problem by fitting the predicted caption set to the ground-truth set, where the set-level relationship is fully captured. Specifically, our set prediction consists of two synergistic tasks, i.e., caption generation and an auxiliary task of concept combination prediction providing extra semantic supervision. Each caption in the set is attached to a concept combination indicating the primary semantic content of the caption and facilitating element alignment in set prediction. Furthermore, we apply a diversity regularization term on concepts to encourage the model to generate semantically diverse captions with various concept combinations. These two tasks share multiple semantics-specific encodings as input, which are obtained by iterative interaction between visual features and conceptual queries. The correspondence between the generated captions and specific concept combinations further guarantees the interpretability of our model. Extensive experiments on benchmark datasets show that the proposed SCG-SP achieves state-of-the-art (SOTA) performance under both relevance and diversity metrics.

In vision-language pre-training (VLP), masked image modeling (MIM) has recently been introduced for fine-grained cross-modal alignment. However, in most existing methods, the reconstruction targets for MIM lack high-level semantics, and text is not sufficiently involved in masked modeling. These two drawbacks limit the effect of MIM in facilitating cross-modal semantic alignment. In this work, we propose a semantics-enhanced cross-modal MIM framework (SemMIM) for vision-language representation learning. Specifically, to provide more semantically meaningful supervision for MIM, we propose a local semantics enhancing approach, which harvest high-level semantics from global image features via self-supervised agreement learning and transfer them to local patch encodings by sharing the encoding space. Moreover, to achieve deep involvement of text during the entire MIM process, we propose a text-guided masking strategy and devise an efficient way of injecting textual information in both masked modeling and reconstruction target acquisition. Experimental results validate that our method improves the effectiveness of the MIM task in facilitating cross-modal semantic alignment. Compared to previous VLP models with similar model size and data scale, our SemMIM model achieves state-of-the-art or competitive performance on multiple downstream vision-language tasks.

In video-text retrieval, most existing methods adopt the dual-encoder architecture for fast retrieval, which employs two individual encoders to extract global latent representations for videos and texts. However, they face challenges in capturing fine-grained semantic concepts. In this work, we propose the UNIFY framework, which learns lexicon representations to capture fine-grained semantics and combines the strengths of latent and lexicon representations for video-text retrieval. Specifically, we map videos and texts into a pre-defined lexicon space, where each dimension corresponds to a semantic concept. A two-stage semantics grounding approach is proposed to activate semantically relevant dimensions and suppress irrelevant dimensions. The learned lexicon representations can thus reflect fine-grained semantics of videos and texts. Furthermore, to leverage the complementarity between latent and lexicon representations, we propose a unified learning scheme to facilitate mutual learning via structure sharing and self-distillation. Experimental results show our UNIFY framework largely outperforms previous video-text retrieval methods, with 4.8% and 8.2% Recall@1 improvement on MSR-VTT and DiDeMo respectively.

Dominant dual-encoder models enable efficient image-text retrieval but suffer from limited accuracy, while the cross-encoder models offer higher accuracy at the expense of efficiency. Distilling cross-modality matching knowledge from cross-encoder to dual-encoder provides a natural approach to harness their strengths. Thus, we investigate the following valuable question: how to make cross-encoder a good teacher for dual-encoder? Our findings are threefold: (1) Cross-modal similarity score distribution of cross-encoder is more concentrated, while the result of dual-encoder is nearly normal, making vanilla logit distillation less effective. However, ranking distillation remains practical, as it is not affected by the score distribution. (2) Only the relative order between hard negatives conveys valid knowledge, while the order information between easy negatives has little significance. (3) Maintaining the coordination between distillation loss and dual-encoder training loss is beneficial for knowledge transfer. Based on these findings, we propose a novel Contrastive Partial Ranking Distillation (CPRD) method, which implements the objective of mimicking relative order between hard negative samples with contrastive learning. This approach coordinates with the training of the dual-encoder, effectively transferring valid knowledge from the cross-encoder to the dual-encoder. Extensive experiments on image-text retrieval and ranking tasks show that our method surpasses other distillation methods and significantly improves the accuracy of dual-encoder.

Haowei Liu, Xi Zhang, Haiyang Xu, Yaya Shi, Chaoya Jiang, Ming Yan, Ji Zhang, Fei Huang, Chunfeng Yuan, Bing Li, Weiming Hu. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Unsupervised graph anomaly detection has been widely used in real-world applications. Existing methods primarily focus on local inconsistency mining (LIM), based on the intuition that establishing high similarities between abnormal nodes and their neighbors is difficult. However, the message passing employed by graph neural networks (GNNs) results in local anomaly signal loss, as GNNs tend to make connected nodes similar, which conflicts with the LIM intuition. In this paper, we propose GADAM, a novel framework that not only resolves the conflict between LIM and message passing but also leverages message passing to augment anomaly detection through a transformative approach to anomaly mining beyond LIM. Specifically, we first propose an efficient MLP-based LIM approach to obtain local anomaly scores in a conflict-free way. Next, we introduce a novel approach to capture anomaly signals from a global perspective. This involves a hybrid attention based adaptive message passing, enabling nodes to selectively absorb abnormal or normal signals from their surroundings. Extensive experiments conducted on nine benchmark datasets, including two large-scale OGB datasets, demonstrate that GADAM surpassinges existing state-of-the-art methods in terms of both effectiveness and efficiency.

Multi-task learning (MTL), which aims to make full use of knowledge contained in multiple tasks to enhance overall performance and efficiency, has been broadly applied in recommendations. The main challenge for MTL models is negative transfer. Existing MTL models, mainly built on the Mixture-of-Experts (MoE) structure, seek enhancements in performance through feature selection and specific expert sharing mode design. However, one expert sharing mode may not be universally applicable due to the complex correlations and diverse demands among various tasks. Additionally, homogeneous expert architectures in such models further limit their performance. To address these issues, in this paper, we propose an innovative automatic MTL framework, AutoMTL, leveraging neural architecture search (NAS) to design optimal expert architectures and sharing modes. The Dual-level Expert Sharing mode and Architecture Navigator (DESAN) search space of AutoMTL can not only efficiently explore expert sharing modes and feature selection schemes but also focus on the architectures of expert subnetworks. Along with this, we introduce an efficient Progressively Discretizing Differentiable Architecture Search (PD-DARTS) algorithm for search space exploration. Extensive experiments demonstrate that AutoMTL can consistently outperform state-of-the-art, human-crafted MTL models. Moreover, the insights obtained from the discovered architectures provide valuable guidance for building new multi-task recommendation models.

Learning on temporal heterogeneous graphs (THGs) has attracted substantial attention in applications of information retrieval. Such graphs are ubiquitous in real-world domains like recommender systems and social networks. However, the spatial heterogeneity, rich semantic information, and intricate evolution patterns of THGs make it still difficult to generate high-quality embeddings for graph nodes. In this paper, we focus on two valuable and understudied issues related to THG learning: (a) How to capture the specific evolutionary characteristics of diverse temporal heterogeneous graphs? (b) Due to the heterogeneous nature of the graph, how to capture the unique temporal patterns of different node types? We explore these questions and present our solution by proposing a new method named CasMLN (Cascaded Multi-level Learning Network) for THG learning. Through the multi-level learning structure and aggregation methods specifically designed for different levels, we obtain information of multiple levels and fuse them to improve embedding generation. Additionally, we pioneer the use of large language models (LLMs) in the THG field. By leveraging the universality and powerful capabilities of LLMs, our method introduces LLM-based external knowledge to effectively capture the implicit nature of graphs and node types, which helps to enhance type- and graph-level representations. We evaluate our method on several real-world THG datasets for different downstream tasks. Extensive experimental results show that CasMLN outperforms the state-of-the-art baselines in both accuracy and efficiency.

Previous work for video captioning aims to objectively describe the video content but the captions lack human interest and attractiveness, limiting its practical application scenarios. The intention of video title generation (video titling) is to produce attractive titles, but there is a lack of benchmarks. This work offers CREATE, the first large-scale Chinese shoRt vidEo retrievAl and Title gEneration dataset, to assist research and applications in video titling, video captioning, and video retrieval in Chinese. CREATE comprises a high-quality labeled 210 K dataset and two web-scale 3 M and 10 M pre-training datasets, covering 51 categories, 50K+ tags, 537K+ manually annotated titles and captions, and 10M+ short videos with original video information. This work presents ACTEr, a unique Attractiveness-Consensus-based Title Evaluation, to objectively evaluate the quality of video title generation. This metric measures the semantic correlation between the candidate (model-generated title) and references (manual-labeled titles) and introduces attractive consensus weights to assess the attractiveness and relevance of the video title. Accordingly, this work proposes a novel multi-modal ALignment WIth Generation model, ALWIG, as one strong baseline to aid future model development. With the help of a tag-driven video-text alignment module and a GPT-based generation module, this model achieves video titling, captioning, and retrieval simultaneously. We believe that the release of the CREATE dataset, ACTEr metric, and ALWIG model will encourage in-depth research on the analysis and creation of Chinese short videos.

Video captioning evaluation aims at assessing the semantic consistency between video and candidate text, which should include measurement from two aspects: faithfulness (whether the information conveyed by candidate is correct w.r.t. video) and comprehensiveness (whether the main video content is covered by candidate). However, previous approaches have difficulty in evaluating faithfulness and comprehensiveness due to heavy reliance on references or heterogeneous of visual and textual data. In this paper, we propose a vision-involved evaluation metric based on a novel DuAl-Reconstruction Transformer, named DARTScore. DARTScore formulates the caption evaluation task as a dual-reconstruction problem to evaluate both faithfulness and comprehensiveness explicitly. Since the word in a candidate is usually related to several frames, DARTScore adaptively collects relevant frames to reconstruct the word and computes the reconstruction accuracy as faithfulness to inherently reflect whether the word information is contained in the video. In the inversive way, DARTScore reconstructs each frame with relevant words to evaluate comprehensiveness. By integrating fine-grained bidirectional reconstruction accuracies, DARTScore drills into each word in candidate and each frame in video to fully evaluate the semantic consistency. Furthermore, we collect and annotate two Chinese datasets with a large domain gap, named CRAETE-EVAL and VATEX-ZH-EVAL, to systematically evaluate existing metrics and fill the blank of Chinese video captioning evaluation. Experimental results show that DARTScore achieves higher correlation with human judgments, has lower reference reliance, and generalizes well to data from different domains.

As the study of graph neural networks becomes more intensive and comprehensive, their robustness and security have received great research interest. The existing global attack methods treat all nodes in the graph as their attack targets. Although existing methods have achieved excellent results, there is still considerable space for improvement. The key problem is that the current approaches rigidly follow the definition of global attacks. They ignore an important issue, i.e., different nodes have different robustness and are not equally resilient to attacks. From a global attacker's view, we should arrange the attack budget wisely, rather than wasting them on highly robust nodes. To this end, we propose a totally new method named partial graph attack (PGA), which selects the vulnerable nodes as attack targets. First, to select the vulnerable items, we propose a hierarchical target selection policy, which allows attackers to only focus on easy-to-attack nodes. Then, we propose a cost-effective anchor-picking policy to pick the most promising anchors for adding or removing edges, and a more aggressive iterative greedy-based attack method to perform more efficient attacks. Extensive experimental results demonstrate that PGA can achieve significant improvements in both attack effect and attack efficiency compared to existing graph global attack methods.

The problem of long-tailed visual recognition has been receiving increasing research attention. However, the long-tailed distribution problem remains underexplored for video-based visual recognition. To address this issue, in this article we propose a compositional learning based solution for video-based human action recognition. Our method, named Attentional Composition Networks (ACN), first learns verb-like and preposition-like components, then shuffles these components to generate samples for the tail classes in the feature space to augment the data for the tail classes. Specifically, during training, we represent each action video by a graph that captures the spatial-temporal relations (edges) among detected human/object instances (nodes). Then, ACN utilizes the position information to decompose each action into a set of verb and preposition representations using the edge features in the graph. After that, the verb and preposition features from different videos are combined via an attention structure to synthesize feature representations for tail classes. This way, we can enrich the data for the tail classes and consequently improve the action recognition for these classes. To evaluate the compositional human action recognition, we further contribute a new human action recognition dataset, namely NEU-Interaction (NEU-I). Experimental results on both Something-Something V2 and the proposed NEU-I demonstrate the effectiveness of the proposed method for long-tailed, few-shot, and zero-shot problems in human action recognition. Source code and the NEU-I dataset are available at https://github.com/YajieW99/ACN .

Dominant pre-training works for image-text retrieval adopt “dual-encoder” architecture to enable high efficiency, where two encoders are used to extract image and text representations and contrastive learning is employed for global alignment. However, coarse-grained global alignment ignores detailed semantic associations between image and text. In this work, we propose a novel proxy task, named Visual-Language Error Modeling (ViLEM), to inject detailed image-text association into “dual-encoder” model by “proofreading” each word in the text against the corresponding image. Specifically, we first edit the image-paired text to automatically generate diverse plausible negative texts with pre-trained language models. ViLEM then enforces the model to discriminate the correctness of each word in the plausible negative texts and further correct the wrong words via resorting to image information. Further-more, we propose a multi-granularity interaction framework to perform ViLEM via interacting text features with both global and local image features, which associates local text semantics with both high-level visual context and multi-level local visual information. Our method surpasses state-of-the-art “dual-encoder” methods by a large margin on the image-text retrieval task and significantly improves discriminativeness to local textual semantics. Our model can also generalize well to video-text retrieval.

Video Tagging intends to infer multiple tags spanning relevant content for a given video. Typically, video tags are freely defined and uploaded by a variety of users, so they have two characteristics: abundant in quantity and disordered intra-video. It is difficult for the existing multilabel classification and generation methods to adapt directly to this task. This paper proposes a novel generative model, Order-Prompted Tag Sequence Generation (OP-TSG), according to the above characteristics. It regards video tagging as a tag sequence generation problem guided by sample-dependent order prompts. These prompts are semantically aligned with tags and enable to decouple tag generation order, making the model focus on modeling the tag dependencies. Moreover, the word-based generation strategy enables the model to generate novel tags. To verify the effectiveness and generalization of the proposed method, a Chinese video tagging benchmark CREATE-tagging, and an English image tagging benchmark Pexel-tagging are established. Extensive results show that OP-TSG is significantly superior to other methods, especially the results on rare tags improve by 3.3% and 3% over SOTA methods on CREATE-tagging and Pexel-tagging, and novel tags generated on CREATE-tagging exhibit a tag gain of 7.04%.

Many real-world data can be modeled as heterogeneous graphs that contain multiple types of nodes and edges. Meanwhile, due to excellent performance, heterogeneous graph neural networks (GNNs) have received more and more attention. However, the existing work mainly focuses on the design of novel GNN models, while ignoring another important issue that also has a large impact on the model performance, namely the missing attributes of some node types. The handcrafted attribute completion requires huge expert experience and domain knowledge. Also, considering the differences in semantic characteristics between nodes, the attribute completion should be fine-grained, i.e., the attribute completion operation should be node-specific. Moreover, to improve the performance of the downstream graph learning task, attribute completion and the training of the heterogeneous GNN should be jointly optimized rather than viewed as two separate processes. To address the above challenges, we propose a differentiable attribute completion framework called AutoAC for automated completion operation search in heterogeneous GNNs. We first propose an expressive completion operation search space, including topology-dependent and topology-independent completion operations. Then, we propose a continuous relaxation schema and further propose a differentiable completion algorithm where the completion operation search is formulated as a bi-level joint optimization problem. To improve the search efficiency, we leverage two optimization techniques: discrete constraints and auxiliary unsupervised graph node clustering. Extensive experimental results on real-world datasets reveal that AutoAC outperforms the SOTA handcrafted heterogeneous GNNs and the existing attribute completion method.

As one of the typical tasks of spatial temporal forecasting, traffic prediction has attracted extensive research attention in recent studies. Recent works usually combine time series modeling methods and graph neural networks to capture the temporal dynamic trends and spatial static dependencies together. However, we find that the temporal static characteristic and spatial dynamic correlations do also exist in the traffic flow, which are neglected or partially considered in previous methods. To capture such characteristics, we propose a novel method for traffic prediction, which models the Spatial Temporal Staticity and Dynamicity (STSD) together. Specifically, besides the traditional spatial and temporal encoder, we propose a pre-training module based on self-supervised contrastive learning to learn the location representation and thus capture the temporal static characteristic. Then, we propose a flow-aware dynamic graph learning module to capture the spatial dynamic correlations. Besides, to correct the original graph built from node distance, a feature reconstruction-based static graph learning module is introduced. The original, static, and dynamic graph are combined together in STSD to model the spatial staticity and dynamicity. Extensive experimental results on real-world datasets reveal that STSD can outperform existing traffic forecasting methods.

Previous dual-encoder pre-training methods for video-text retrieval employ contrastive learning for cross-modal alignment in a latent space. However, such learned latent spaces often result in modality gap problem [26]. In this paper, we introduce a novel SemVTR framework designed to learn semantics-grounded video-text representations in a vocabulary space, in which each dimension corresponds to a semantic concept represented by a word. The representation is obtained by grounding video and text into semantically-related dimensions with high activation values. As video-text pairs share grounded dimensions, their vocabulary representations are expected to cluster together and thus alleviate modality gap problem. So, the crux of our method lies in grounding video and text into vocabulary space. Specifically, we propose a Multi-Granularity Video Semantics Grounding approach and a Textual Semantics Preserving training strategy. The visualization illustrates that SemVTR obtains semantics-gronded vocabulary representation and also alleviates the modality gap problem. SemVTR significantly outperforms existing methods on four video-text retrieval benchmarks.

Differentiable NAS (DARTS) is a simple and efficient neural architecture search method that has been extensively adopted in various machine learning tasks. % Nevertheless, DARTS still encounters several robustness issues, mainly the domination of skip connections. % The resulting architectures are full of parametric-free operations, leading to performance collapse. % Existing methods suggest that the skip connection has additional advantages in optimization compared to other parametric operations and propose to alleviate the domination of skip connections by eliminating these additional advantages. % In this paper, we analyze this issue from a simple and straightforward perspective and propose that the domination of skip connections results from parametric operations overfitting the training data while architecture parameters are trained on the validation data, leading to undesired behaviors. % Based on this observation, we propose the operation-level early stopping (OLES) method to overcome this issue and robustify DARTS without introducing any computation overhead. % Extensive experimental results can verify our hypothesis and the effectiveness of OLES.

3D visual grounding, the task of identifying visual objects in 3D scenes based on natural language inputs, plays a critical role in enabling machines to understand and engage with the real-world environment. However, this task is challenging due to the necessity to capture 3D contextual information to distinguish target objects from complex 3D scenes. The absence of annotations for contextual objects and relations further exacerbates the difficulties. In this paper, we propose a novel model, CORE-3DVG, to address these challenges by explicitly learning about contextual objects and relations. Our method accomplishes 3D visual grounding via three sequential modular networks, including a text-guided object detection network, a relation matching network, and a target identification network. During training, we introduce a pseudo-label self-generation strategy and a weakly-supervised method to facilitate the learning of contextual objects and relations, respectively. The proposed techniques allow the networks to focus more effectively on referred objects within 3D scenes by understanding their context better. We validate our model on the challenging Nr3D, Sr3D, and ScanRefer datasets and demonstrate state-of-the-art performance. Our code will be public at https://github.com/yangli18/CORE-3DVG.

Graph collaborative filtering has achieved great success in capturing users' preferences over items. Despite effectiveness, graph neural network (GNN)-based methods suffer from data sparsity in real scenarios. Recently, contrastive learning (CL) has been used to address the problem of data sparsity. However, most CL-based methods only leverage the original user-item interaction graph to construct the CL task, lacking the explicit exploitation of the higher-order information (i.e., user-user and item-item relationships). Even for the CL-based method that uses the higher-order information, the reception field of the higher-order information is fixed and regardless of the difference between nodes. In this paper, we propose a novel adaptive multi-view fusion contrastive learning framework, named AdaMCL, for graph collaborative filtering. To exploit the higher-order information more accurately, we propose an adaptive fusion strategy to fuse the embeddings learned from the user-item and user-user graphs. Moreover, we propose a multi-view fusion contrastive learning paradigm to construct effective CL tasks. Besides, to alleviate the noisy information caused by aggregating higher-order neighbors, we propose a layer-level CL task. Extensive experimental results reveal that AdaMCL is effective and outperforms existing collaborative filtering models significantly.

Blind image inpainting involves two critical aspects, i.e., "where to inpaint" and "how to inpaint". Knowing "where to inpaint" can eliminate the interference arising from corrupted pixel values; a good "how to inpaint" strategy yields high-quality inpainted results robust to various corruptions. In existing methods, these two aspects usually lack explicit and separate consideration. This paper fully explores these two aspects and proposes a self-prior guided inpainting network (SIN). The self-priors are obtained by detecting semantic-discontinuous regions and by predicting global semantic structures of the input image. On the one hand, the self-priors are incorporated into the SIN, which enables the SIN to perceive valid context information from uncorrupted regions and to synthesize semantic-aware textures for corrupted regions. On the other hand, the self-priors are reformulated to provide a pixel-wise adversarial feedback and a high-level semantic structure feedback, which can promote the semantic continuity of inpainted images. Experimental results demonstrate that our method achieves state-of-the-art performance in metric scores and in visual quality. It has an advantage over many existing methods that assume "where to inpaint" is known in advance. Extensive experiments on a series of related image restoration tasks validate the effectiveness of our method in obtaining high-quality inpainting.

In skeleton-based action recognition, it has been a dominant paradigm to extract motion features with temporal convolution and model spatial correlations with graph convolution. However, it’s difficult for temporal convolution to capture long-range dependencies effectively. Meanwhile, commonly used multi-branch graph convolution leads to high complexity. In this paper, we propose TranSkeleton, a powerful Transformer framework which neatly unifies the spatial and temporal modeling of skeleton sequences. For temporal modeling, we propose a novel partition-aggregation temporal Transformer. It works with hierarchical temporal partition and aggregation, and can capture both long-range dependencies and subtle temporal structures effectively. A difference-aware aggregation approach is designed to reduce information loss during temporal aggregation. For spatial modeling, we propose a topology-aware spatial Transformer which utilizes the prior information of human body topology to facilitate spatial correlation modeling. Extensive experiments on two challenging benchmark datasets demonstrate that TranSkeleton notably outperforms the state of the arts.

Video captioning requires that the model has the abilities of video understanding, video-text alignment, and text generation. Due to the semantic gap between vision and language, conducting video-text alignment is a crucial step to reduce the semantic gap, which maps the representations from the visual to the language domain. However, the existing methods often overlook this step, so the decoder has to directly take the visual representations as input, which increases the decoder’s workload and limits its ability to generate semantically correct captions. In this paper, we propose a video-text alignment module with a retrieval unit and an alignment unit to learn video-text aligned representations for video captioning. Specifically, we firstly propose a retrieval unit to retrieve sentences as additional input which is used as the semantic anchor between visual scene and language description. Then, we employ an alignment unit with the input of the video and retrieved sentences to conduct the video-text alignment. The representations of two modal inputs are aligned in a shared semantic space. The obtained video-text aligned representations are used to generate semantically correct captions. Moreover, retrieved sentences provide rich semantic concepts which are helpful for generating distinctive captions. Experiments on two public benchmarks, i.e., VATEX and MSR-VTT, demonstrate that our method outperforms state-of-the-art performances by a large margin. The qualitative analysis shows that our method generates correct and distinctive captions.

Recently emerged heterophilous graph neural networks have significantly reduced the reliance on the assumption of graph homophily where linked nodes have similar features and labels. These methods focus on a supervised setting that relies on labeling information heavily and presents the limitations on general graph downstream tasks. In this work, we propose a self-supervised representation learning paradigm on graphs with heterophily (namely HGRL) for improving the generalizability of node representations, where node representations are optimized without any label guidance. Inspired by the designs of existing heterophilous graph neural networks, HGRL learns the node representations by preserving the node original features and capturing informative distant neighbors. Such two properties are obtained through carefully designed pretext tasks that are optimized based on estimated high-order mutual information. Theoretical analysis interprets the connections between HGRL and existing advanced graph neural network designs. Extensive experiments on different downstream tasks demonstrate the effectiveness of the proposed framework.

Current metrics for video captioning are mostly based on the text-level comparison between reference and candidate captions. However, they have some insuperable drawbacks, e.g., they cannot handle videos without references, and they may result in biased evaluation due to the one-to-many nature of video-to-text and the neglect of visual relevance. From the human evaluator's viewpoint, a high-quality caption should be consistent with the provided video, but not necessarily be similar to the reference in literal or semantics. Inspired by human evaluation, we propose EMScore (Embedding Matching-based score), a novel reference-free metric for video captioning, which directly measures similarity between video and candidate captions. Benefiting from the recent development of large-scale pre-training models, we exploit a well pre-trained vision-language model to extract visual and linguistic embeddings for computing EMScore. Specifically, EMScore combines matching scores of both coarse-grained (video and caption) and fine-grained (frames and words) levels, which takes the overall understanding and detailed characteristics of the video into account. Furthermore, considering the potential information gain, EMScore can be flexibly extended to the conditions where human-labeled references are available. Last but not least, we collect VATEX-EVAL and ActivityNet-FOIl datasets to systematically evaluate the existing metrics. VATEX-EVAL experiments demonstrate that EMScore has higher human correlation and lower reference dependency. ActivityNet-FOIL experiment verifies that EMScore can effectively identify “hallucinating” captions. Code and datasets are available at https://github.com/shiyaya/emscore.

Visual grounding is a task to locate the target indicated by a natural language expression. Existing methods extend the generic object detection framework to this problem. They base the visual grounding on the features from pre-generated proposals or anchors, and fuse these features with the text embeddings to locate the target mentioned by the text. However, modeling the visual features from these predefined locations may fail to fully exploit the visual context and attribute information in the text query, which limits their performance. In this paper, we propose a transformer-based framework for accurate visual grounding by establishing text-conditioned discriminative features and performing multi-stage cross-modal reasoning. Specifically, we develop a visual-linguistic verification module to focus the visual features on regions relevant to the textual descriptions while suppressing the unrelated areas. A language-guided feature encoder is also devised to aggregate the visual contexts of the target object to improve the object's distinctiveness. To retrieve the target from the encoded visual features, we further propose a multi-stage cross-modal decoder to iteratively speculate on the correlations between the image and text for accurate target localization. Extensive experiments on five widely used datasets validate the efficacy of our proposed components and demonstrate state-of-the-art performance.

Recently, graph neural network (GNN) has achieved great success in many graph learning tasks such as node classifi-cation and graph classification. However, there is no single GNN architecture that can fit different graph datasets. Designing an effective GNN for a specific graph dataset requires considerable expert experience and huge computational costs. Inspired by the success of neural architecture search (NAS), searching the GNN architectures automatically has attracted more and more attention. Motivated by the fact that the search space plays a critical role in the NAS, we propose a novel and effective graph neural architecture search method called PSP from the perspective of search space design in this paper. We first propose an expressive search space composed of multiple cells. Instead of searching the entire architecture, we focus on searching the architecture of the cell. Then, we propose a progressive space pruning-based algorithm to search the architectures efficiently. Moreover, the data-specific search spaces and architectures ob-tained by PSP can be transferred to new graph datasets based on meta-learning. Extensive experimental results on different types of graph datasets reveal that PSP outperforms the state-of-the-art handcrafted architectures and the existing NAS methods in terms of effectiveness and efficiency.

Unsupervised image segmentation is a challenge task, since a high-quality segmented image should perceive not only local object structures but also certain semantics without any annotations. In this paper, we propose a novel encoder-decoder pixel clustering framework with dual constraints to incorporate local structure and global semantic information for guiding pixel feature learning in a self-supervised manner. On one hand, a Local Structure Constraint (LStC) is constructed based on fine-grained superpixels, which improves the boundary perception of pixel features by keeping intra-superpixel feature consistency and largening inter-superpixel feature distance. On the other hand, a new Global Semantic Constraint (GSeC) is proposed via adapting the mutual information maximization technique to the single-image setting, and it strengthens the global semantic perception of pixel features and thus improves the segmenting integrity of objects. Finally, based on the learned pixel features, a smoothing component is employed to achieve semantically meaningful pixel clustering. The experimental evaluation on BSDS500 and PASCAL Context datasets show the superiority of our method on region and boundary qualities.

Recent studies have shown that deep neural networks-based recommender systems are vulnerable to adversarial attacks, where attackers can inject carefully crafted fake user profiles (i.e., a set of items that fake users have interacted with) into a target recommender system to achieve malicious purposes, such as promote or demote a set of target items. Due to the security and privacy concerns, it is more practical to perform adversarial attacks under the black-box setting, where the architecture/parameters and training data of target systems cannot be easily accessed by attackers. However, generating high-quality fake user profiles under black-box setting is rather challenging with limited resources to target systems. To address this challenge, in this work, we introduce a novel strategy by leveraging items' attribute information (i.e., items' knowledge graph), which can be publicly accessible and provide rich auxiliary knowledge to enhance the generation of fake user profiles. More specifically, we propose a knowledge graph-enhanced black-box attacking framework (KGAttack) to effectively learn attacking policies through deep reinforcement learning techniques, in which knowledge graph is seamlessly integrated into hierarchical policy networks to generate fake user profiles for performing adversarial black-box attacks. Comprehensive experiments on various real-world datasets demonstrate the effectiveness of the proposed attacking framework under the black-box setting.

Session-based recommendation aims to predict next click action (e.g., item) of anonymous users based on a fixed number of previous actions. Recently, Graph Neural Networks (GNNs) have shown superior performance in various applications. Inspired by the success of GNNs, tremendous endeavors have been devoted to introduce GNNs into session-based recommendation and have achieved significant results. Nevertheless, due to the highly diverse types of potential information in sessions, existing GNNs-based methods perform differently on different session datasets, leading to the need for efficient design of neural networks adapted to various session recommendation scenarios. To address this problem, we propose Automated neural architecture search for Graph-based Session Recommendation, namely AutoGSR, a framework that provides a practical and general solution to automatically find the optimal GNNs-based session recommendation model. In AutoGSR, we propose two novel GNN operations to build an expressive and compact search space. Building upon the search space, we employ a differentiable search algorithm to search for the optimal graph neural architecture. Furthermore, to consider all types of session information together, we propose to learn the item meta knowledge, which acts as a priori knowledge for guiding the optimization of final session representations. Comprehensive experiments on three real-world datasets demonstrate that AutoGSR is able to find effective neural architectures and achieve state-of-the-art results. To the best of our knowledge, we are the first to study the neural architecture search for the session-based recommendation.

Click-Through Rate (CTR) prediction has been widely used in many machine learning tasks such as online advertising and personalization recommendation. Unfortunately, given a domain-specific dataset, searching effective feature interaction operations and combinations from a huge candidate space requires significant expert experience and computational costs. Recently, Neural Architecture Search (NAS) has achieved great success in discovering high-quality network architectures automatically. However, due to the diversity of feature interaction operations and combinations, the existing NAS-based work that treats the architecture search as a black-box optimization problem over a discrete search space suffers from low efficiency. Therefore, it is essential to explore a more efficient architecture search method. To achieve this goal, we propose NAS-CTR, a differentiable neural architecture search approach for CTR prediction. First, we design a novel and expressive architecture search space and a continuous relaxation scheme to make the search space differentiable. Second, we formulate the architecture search for CTR prediction as a joint optimization problem with discrete constraints on architectures and leverage proximal iteration to solve the constrained optimization problem. Additionally, a straightforward yet effective method is proposed to eliminate the aggregation of skip connections. Extensive experimental results reveal that NAS-CTR can outperform the SOTA human-crafted architectures and other NAS-based methods in both test accuracy and search efficiency.

Named Entity Recognition (NER) aims to identify the pre-defined entities from the unstructured text. Compared with English NER, Chinese NER faces more challenges: the ambiguity problem in entity boundary recognition due to unavailable explicit delimiters between Chinese characters, and the out-of-vocabulary (OOV) problem caused by rare Chinese characters. However, two important features specific to the Chinese language are ignored by previous studies: glyphs and phonetics, which contain rich semantic information of Chinese. To overcome these issues by exploiting the linguistic potential of Chinese as a logographic language, we present MPM-CNER (short for Multi-modal Pretraining Model for Chinese NER), a model for learning multi-modal representations of Chinese semantics, glyphs, and phonetics, via four pretraining tasks: Radical Consistency Identification (RCI), Glyph Image Classification (GIC), Phonetic Consistency Identification (PCI), and Phonetic Classification Modeling (PCM). Meanwhile, a novel cross-modality attention mechanism is proposed to fuse these multimodal features for further improvement. The experimental results show that our method outperforms the state-of-the-art baseline methods on four benchmark datasets, and the ablation study also verifies the effectiveness of the pre-trained multi-modal representations.

Abstract Decision trees have favorable properties, including interpretability, high computational efficiency, and the ability to learn from little training data. Learning a decision tree is known to be NP-complete. The researchers have proposed many greedy algorithms such as CART to learn approximate solutions. Inspired by the current popular neural networks, soft trees that support end-to-end training with back-propagation have attracted more and more attention. However, existing soft trees either lose the interpretability due to the continuous relaxation or employ the two-stage method of end-to-end building and then pruning. In this paper, we propose One-Stage Tree to build and prune the decision tree jointly through a bilevel optimization problem. Moreover, we leverage the reparameterization trick and proximal iterations to keep the tree discrete during end-to-end training. As a result, One-Stage Tree reduces the performance gap between training and testing and maintains the advantage of interpretability. Extensive experiments demonstrate that the proposed One-Stage Tree outperforms CART and the existing soft trees on classification and regression tasks.

For CNN-based visual action recognition, the accuracy may be increased if local key action regions are focused on. The task of self-attention is to focus on key features and ignore irrelevant information. So, self-attention is useful for action recognition. However, current self-attention methods usually ignore correlations among local feature vectors at spatial positions in CNN feature maps. In this paper, we propose an effective interaction-aware self-attention model which can extract information about the interactions between feature vectors to learn attention maps. Since the different layers in a network capture feature maps at different scales, we introduce a spatial pyramid with the feature maps at different layers for attention modeling. The multi-scale information is utilized to obtain more accurate attention scores. These attention scores are used to weight the local feature vectors of the feature maps and then calculate attentional feature maps. Since the number of feature maps input to the spatial pyramid attention layer is unrestricted, we easily extend this attention layer to a spatio-temporal version. Our model can be embedded in any general CNN to form a video-level end-to-end attention network for action recognition. Several methods are investigated to combine the RGB and flow streams to obtain accurate predictions of human actions. Experimental results show that our method achieves state-of-the-art results on the datasets UCF101, HMDB51, Kinetics-400, and untrimmed Charades.

Recent one-stage object detectors follow a per-pixel prediction approach that predicts both the object category scores and boundary positions from every single grid location. However, the most suitable positions for inferring different targets, i.e., the object category and boundaries, are generally different. Predicting all these targets from the same grid location thus may lead to sub-optimal results. In this paper, we analyze the suitable inference positions for object category and boundaries, and propose a prediction-target-decoupled detector named PDNet to establish a more flexible detection paradigm. Our PDNet with the prediction decoupling mechanism encodes different targets separately in different locations. A learnable prediction collection module is devised with two sets of dynamic points, i.e., dynamic boundary points and semantic points, to collect and aggregate the predictions from the favorable regions for localization and classification. We adopt a two-step strategy to learn these dynamic point positions, where the prior positions are estimated for different targets first, and the network further predicts residual offsets to the positions with better perceptions of the object properties. Extensive experiments on the MS COCO benchmark demonstrate the effectiveness and efficiency of our method. With a single ResNeXt-64x4d-101-DCN as the backbone, our detector achieves 50.1 AP with single-scale testing, which outperforms the state-of-the-art methods by an appreciable margin under the same experimental settings. Moreover, our detector is highly efficient as a one-stage framework. Our code will be public.

Filter pruning is a commonly used method for compressing Convolutional Neural Networks (ConvNets), due to its friendly hardware supporting and flexibility. However, existing methods mostly need a cumbersome procedure, which brings many extra hyper-parameters and training epochs. This is because only using sparsity and pruning stages cannot obtain a satisfying performance. Besides, many works do not consider the difference of pruning ratio across different layers. To overcome these limitations, we propose a novel dynamic and progressive filter pruning (DPFPS) scheme that directly learns a structured sparsity network from Scratch. In particular, DPFPS imposes a new structured sparsity-inducing regularization specifically upon the expected pruning parameters in a dynamic sparsity manner. The dynamic sparsity scheme determines sparsity allocation ratios of different layers and a Taylor series based channel sensitivity criteria is presented to identify the expected pruning parameters. Moreover, we increase the structured sparsity-inducing penalty in a progressive manner. This helps the model to be sparse gradually instead of forcing the model to be sparse at the beginning. Our method solves the pruning ratio based optimization problem by an iterative soft-thresholding algorithm (ISTA) with dynamic sparsity. At the end of the training, we only need to remove the redundant parameters without other stages, such as fine-tuning. Extensive experimental results show that the proposed method is competitive with 11 state-of-the-art methods on both small-scale and large-scale datasets (i.e., CIFAR and ImageNet). Specifically, on ImageNet, we achieve a 44.97% pruning ratio of FLOPs by compressing ResNet-101, even with an increase of 0.12% Top-5 accuracy. Our pruned models and codes are released at https://github.com/taoxvzi/DPFPS.

In this paper, we convert traditional video captioning task into a new paradigm, i.e., Open-book Video Captioning, which generates natural language under the prompts of video-content-relevant sentences, not limited to the video itself. To address the open-book video captioning problem, we propose a novel Retrieve-Copy-Generate network, where a pluggable video-to-text retriever is constructed to retrieve sentences as hints from the training corpus effectively, and a copy-mechanism generator is introduced to extract expressions from multi-retrieved sentences dynamically. The two modules can be trained end-to-end or separately, which is flexible and extensible. Our framework co-ordinates the conventional retrieval-based methods with orthodox encoder-decoder methods, which can not only draw on the diverse expressions in the retrieved sentences but also generate natural and accurate content of the video. Extensive experiments on several benchmark datasets show that our proposed approach surpasses the state-of-the-art performance, indicating the effectiveness and promising of the proposed paradigm in the task of video captioning.

Graph convolutional networks (GCNs) have been widely used and achieved remarkable results in skeleton-based action recognition. In GCNs, graph topology dominates feature aggregation and therefore is the key to extracting representative features. In this work, we propose a novel Channel-wise Topology Refinement Graph Convolution (CTR-GC) to dynamically learn different topologies and effectively aggregate joint features in different channels for skeleton-based action recognition. The proposed CTR-GC models channel-wise topologies through learning a shared topology as a generic prior for all channels and refining it with channel-specific correlations for each channel. Our refinement method introduces few extra parameters and significantly reduces the difficulty of modeling channel-wise topologies. Furthermore, via reformulating graph convolutions into a unified form, we find that CTR-GC relaxes strict constraints of graph convolutions, leading to stronger representation capability. Combining CTR-GC with temporal modeling modules, we develop a powerful graph convolutional network named CTR-GCN which notably outperforms state-of-the-art methods on the NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets.1

The industry and academia have proposed many distributed graph processing systems. However, the existing systems are not friendly enough for users like data analysts and algorithm engineers. On the one hand, the programming models and interfaces differ a lot in the existing systems, leading to high learning costs and program migration costs. On the other hand, these graph processing systems are tightly bound to the underlying distributed computing platforms, requiring users to be familiar with distributed computing. To improve the usability of distributed graph processing, we propose a unified distributed graph programming framework UniGPS. Firstly, we propose a unified cross-platform graph programming model VCProg for UniGPS. VCProg hides details of distributed computing from users. It is compatible with the popular graph programming models Pregel, GAS, and Push-Pull. VCProg-based programs can be executed by compatible distributed graph processing systems without modification, reducing the learning overheads of users. Secondly, UniGPS supports Python as the programming language. We propose an interprocess-communication-based execution environment isolation mechanism to enable Java/C++-based graph processing systems to call user-defined methods written in Python. The experimental results show that UniGPS enables users to process big graphs beyond the memory capacity of a single machine without sacrificing usability. UniGPS shows near-linear data scalability and machine scalability.

Abstract Nowadays, the financial securities investment and transaction are processed digitally. The securities companies have accumulated a large amount of financial data. Given that, quantitative analysis has become a common method for securities investors. However, as the scale of data continues to grow, data storing and processing have increasingly become a big challenge to financial quantitative analysts. First, most existing stand‐alone quantitative analysis systems are hard to accelerate data processing in a distributed way. Second, existing distributed computing frameworks such as Apache Spark demands financial quantitative analysts with expertise knowledge on distributed computing. Third, transition from traditional stand‐alone financial quantitative analysis (FQA) system to such distributed computing system often introduces learning efforts and development overheads. To solve these problems, we propose Alchemy, a distributed processing platform tailored for FQA with high‐level programming interface. Alchemy offers distributed computing service and provides a pipeline‐based programming model which shares many similarities with the traditional stand‐alone systems, reducing learning efforts for financial quantitative analysts. The performance of Alchemy is evaluated in both experimental environment and a real‐world production environment in Huatai Securites which is a leading financial company in China. The results show that Alchemy is able to achieve up to 300 times speedup compared against existing financial quantitative analysis systems.

For weakly supervised object localization (WSOL), how to avoid the network focusing only on some small discriminative parts is a main challenge needed to solve. The widely-used Class Activation Mapping (CAM) based paradigm usually employs Adversarial Learning (AL) strategy to search more object parts by constantly hiding discovered object features, but the adversarial process is difficult to control. In this paper, we propose a novel CAM-based framework with Multi-scale Low-Discriminative Feature Reactivation (mLDFR) for WSOL. The mLDFR framework reactivates the low-discriminative object parts via bottom-up continuous feature maps recalibration and multi-scale object category mapping. Compared with the AL-based methods, our method fully improves the localization power of the network without damaging the classification power and can perform multi-instance localization, which are hard to achieve under the AL-based framework. Moreover, the mLDFR framework is flexible, and can be built on the top of various classical CNN backbones. Experimental results demonstrate the superiority of our method. With VGG16 as backbone, we achieve 46.96% Cls-Loc top1 err and 66.12% CorLoc on ILSVRC2014, 38.07% Cls-Loc top1 err and 75.04% CorLoc on CUB200-2011, surpassing the state-of-the-arts by a large margin.

Model compression methods have become popular in recent years, which aim to alleviate the heavy load of deep neural networks (DNNs) in real-world applications. However, most of the existing compression methods have two limitations: 1) they usually adopt a cumbersome process, including pretraining, training with a sparsity constraint, pruning/decomposition, and fine-tuning. Moreover, the last three stages are usually iterated multiple times. 2) The models are pretrained under explicit sparsity or low-rank assumptions, which are difficult to guarantee wide appropriateness. In this article, we propose an efficient decomposition and pruning (EDP) scheme via constructing a compressed-aware block that can automatically minimize the rank of the weight matrix and identify the redundant channels. Specifically, we embed the compressed-aware block by decomposing one network layer into two layers: a new weight matrix layer and a coefficient matrix layer. By imposing regularizers on the coefficient matrix, the new weight matrix learns to become a low-rank basis weight, and its corresponding channels become sparse. In this way, the proposed compressed-aware block simultaneously achieves low-rank decomposition and channel pruning by only one single data-driven training stage. Moreover, the network of architecture is further compressed and optimized by a novel Pruning & Merging (PM) module which prunes redundant channels and merges redundant decomposed layers. Experimental results (17 competitors) on different data sets and networks demonstrate that the proposed EDP achieves a high compression ratio with acceptable accuracy degradation and outperforms state-of-the-arts on compression rate, accuracy, inference time, and run-time memory.

Finding or monitoring subgraph instances that are isomorphic to a given pattern graph in a data graph is a fundamental query operation in many graph analytic applications, such as network motif mining and fraud detection. Existing distributed methods are inefficient in communication. They have to shuffle partial matching results during the distributed multiway join. The partial matching results may be much larger than the data graph itself. To overcome the drawback, we develop the Batch-BENU framework for distributed subgraph enumeration on static data graphs. Batch-BENU executes a group of local search tasks in parallel. Each task enumerates subgraphs around a vertex in the data graph, guided by a backtracking-based execution plan. To handle large-scale data graphs that may exceed the memory capacity of a single machine, Batch-BENU stores the data graph in a distributed database. Each task queries adjacency sets of the data graph on demand, shuffling the data graph instead of partial matching results. To support incremental subgraph enumeration on dynamic data graphs, we propose the Streaming-BENU framework. Streaming-BENU turns the problem of enumerating incremental matching results into enumerating all matching results of incremental pattern graphs at each time step. We implement Batch-BENU and Streaming-BENU with the local database cache and the load balance optimization to improve their efficiency. Extensive experiments show that Batch-BENU and Streaming-BENU can scale to big graphs and complex pattern graphs. They outperform the state-of-the-art distributed methods by up to one and two orders of magnitude, respectively.

Taking full advantage of the information from both vision and language is critical for the video captioning task. Existing models lack adequate visual representation due to the neglect of interaction between object, and sufficient training for content-related words due to long-tailed problems. In this paper, we propose a complete video captioning system including both a novel model and an effective training strategy. Specifically, we propose an object relational graph (ORG) based encoder, which captures more detailed interaction features to enrich visual representation. Meanwhile, we design a teacher-recommended learning (TRL) method to make full use of the successful external language model (ELM) to integrate the abundant linguistic knowledge into the caption model. The ELM generates more semantically similar word proposals which extend the groundtruth words used for training to deal with the long-tailed problem. Experimental evaluations on three benchmarks: MSVD, MSR-VTT and VATEX show the proposed ORG-TRL system achieves state-of-the-art performance. Extensive ablation studies and visualizations illustrate the effectiveness of our system.

Convolutional neural networks are built upon simple but useful convolution modules. The traditional convolution has a limitation on feature extraction and object localization due to its fixed scale and geometric structure. Besides, the loss of spatial information also restricts the networks' performance and depth. To overcome these limitations, this paper proposes a novel anisotropic convolution by adding a scale factor and a shape factor into the traditional convolution. The anisotropic convolution augments the receptive fields flexibly and dynamically depending on the valid sizes of objects. In addition, the anisotropic convolution is a generalized convolution. The traditional convolution, dilated convolution and deformable convolution can be viewed as its special cases. Furthermore, in order to improve the training efficiency and avoid falling into a local optimum, this paper introduces a simplified implementation of the anisotropic convolution. The anisotropic convolution can be applied to arbitrary convolutional networks and the enhanced networks are called ACNs (anisotropic convolutional networks). Experimental results show that ACNs achieve better performance than many state-of-the-art methods and the baseline networks in tasks of image classification and object localization, especially in classification task of tiny images.

Convolutional Neural Networks have achieved excellent successes for object recognition in still images. However, the improvement of Convolutional Neural Networks over the traditional methods for recognizing actions in videos is not so significant, because the raw videos usually have much more redundant or irrelevant information than still images. In this paper, we propose a Spatial-Temporal Attentive Convolutional Neural Network (STA-CNN) which selects the discriminative temporal segments and focuses on the informative spatial regions automatically. The STA-CNN model incorporates a Temporal Attention Mechanism and a Spatial Attention Mechanism into a unified convolutional network to recognize actions in videos. The novel Temporal Attention Mechanism automatically mines the discriminative temporal segments from long and noisy videos. The Spatial Attention Mechanism firstly exploits the instantaneous motion information in optical flow features to locate the motion salient regions and it is then trained by an auxiliary classification loss with a Global Average Pooling layer to focus on the discriminative non-motion regions in the video frame. The STA-CNN model achieves the state-of-the-art performance on two of the most challenging datasets, UCF-101 (95.8%) and HMDB-51 (71.5%).

The key challenge of knowledge distillation is to extract general, moderate and sufficient knowledge from a teacher network to guide a student network. In this paper, a novel Instance Relationship Graph (IRG) is proposed for knowledge distillation. It models three kinds of knowledge, including instance features, instance relationships and feature space transformation, while the latter two kinds of knowledge are neglected by previous methods. Firstly, the IRG is constructed to model the distilled knowledge of one network layer, by considering instance features and instance relationships as vertexes and edges respectively. Secondly, an IRG transformation is proposed to models the feature space transformation across layers. It is more moderate than directly mimicking the features at intermediate layers. Finally, hint loss functions are designed to force a student's IRGs to mimic the structures of a teacher's IRGs. The proposed method effectively captures the knowledge along the whole network via IRGs, and thus shows stable convergence and strong robustness to different network architectures. In addition, the proposed method shows superior performance over existing methods on datasets of various scales.

Given a small pattern graph and a large data graph, the task of subgraph enumeration is to find all the subgraphs of the data graph that are isomorphic to the pattern graph. The state-of-the-art distributed algorithms like SEED and CBF turn subgraph enumeration into a distributed multi-way join problem. They are inefficient in communication as they have to shuffle partial matching results that are much larger than the data graph itself during the join. They also spend non-trivial costs on constructing indexes for data graphs. Different from those join-based algorithms, we develop a new backtracking-based framework BENU for distributed subgraph enumeration. BENU divides a subgraph enumeration task into a group of local search tasks that can be executed in parallel. Each local search task follows a backtracking-based execution plan to enumerate subgraphs. The data graph is stored in a distributed database and is queried as needed. BENU only queries the necessary edges of the data graph and avoids shuffling partial matching results. We also develop an efficient implementation for BENU. We set up an in-memory database cache on each machine. Taking advantage of the inter-task and intra-task locality, the cache significantly reduces the communication cost with controllable memory usage. We conduct extensive experiments to evaluate the performance of BENU. The results show that BENU is scalable and outperforms the state-of-the-art methods by up to an order of magnitude.

The multi-way join query plays a fundamental role in many big data analytic scenarios. Recently, the hybrid join query is becoming increasingly important. However, the existing one-round and multi-round algorithms have limitations in the process of the hybrid query. In this paper, we present a novel hybrid structure-aware multi-way join algorithm called HyMJ, which combines the one-round and multi-round algorithms to compute the hybrid query efficiently. First, we propose the query structure graph (QSG) to represent the internal query structure of a given join query and the query structure decomposition tree (QSDT) to represent the structure-aware query plan. Each internal node of the QSDT denotes a subquery with a cyclic or acyclic query structure. Then, we design a graph contraction based algorithm to construct QSDT from QSG. Furthermore, to select the optimal join strategy for each subquery in the QSDT, we introduce a heuristic strategy selection model. Experimental results on Apache Spark reveal that HyMJ outperforms both the one-round and multi-round algorithms for hybrid multi-way join queries on real-world datasets.

Inspired by the fact that different modalities in videos carry complementary information, we propose a Multimodal Semantic Attention Network(MSAN), which is a new encoder-decoder framework incorporating multimodal semantic attributes for video captioning. In the encoding phase, we detect and generate multimodal semantic attributes by formulating it as a multi-label classification problem. Moreover, we add auxiliary classification loss to our model that can obtain more effective visual features and high-level multimodal semantic attribute distributions for sufficient video encoding. In the decoding phase, we extend each weight matrix of the conventional LSTM to an ensemble of attribute-dependent weight matrices, and employ attention mechanism to pay attention to different attributes at each time of the captioning process. We evaluate algorithm on two popular public benchmarks: MSVD and MSR-VTT, achieving competitive results with current state-of-the-art across six evaluation metrics.

In the big data era, data-intensive cluster computing systems like Hadoop, have gained much popularity, and YARN, the second generation of Hadoop becomes the general resource manager in the Hadoop ecosystem. In the distributed computing scenarios, data locality (scheduling tasks on where the data resides) is essential to the performance since higher data locality brings lower network transmission cost and higher throughput. However, we find that the native YARN scheduling mechanism has little data locality and the delay scheduling strategy leads to the long-tail effect while achieving data locality for in-memory computing scenarios. Therefore, in this paper we propose the push-based YARN scheduling mechanism for the in-memory computing environment. First, we classify the Resource Requests into various categories. Then, we prune the non-local Resource Requests to achieve fast datalocality in-memory computation. Finally, we push the left longtail Resource Requests to the data-locality nodes to avoid the long-tail effect. The experimental results demonstrate that the proposed scheduling mechanism achieves nearly 100% datalocality percentage comparing to the native YARN scheduling mechanism that only achieves 10% 20% data-locality percentage. Under the identical data-locality percentage, the proposed push based scheduling mechanism promotes nearly 20% throughput and reduces nearly 10% application running time comparing to the existing delay scheduling mechanism used in YARN.

Summary Multitiered storage systems, which are made up of heterogeneous devices, are widely used in distributed environments to accelerate the I/O performance of upper big data applications. It raises new challenges in efficient data migration through smart caching mechanisms among heterogeneous storage levels, such as MEM‐SSD‐HDD. To optimize the cache policy scheduling mechanism on the distributed tiered storage architecture, we proposed a general framework with five layers, including a tiered storage system layer, a cache migration policy layer, a cache policy adaptive scheduling layer, a data access pattern layer, and a big data application layer. The framework prototype has been designed and implemented on the widely used distributed hybrid storage system named Alluxio. To meet the demands of the big data application layer, on the one hand, we designed a couple of cache eviction policies and promotion policies covering various access patterns on the cache migration policy layer (several proposed eviction policies have been adopted by the Alluxio open‐source community). On the other hand, two adaptive cache policy scheduling algorithms for selecting appropriate policies in various scenarios are designed and implemented on the cache policy adaptive scheduling layer. The scheduling algorithms are designed based on the hit ratio statistics and data access pattern model prediction, respectively. Experimental results show that the proposed cache policies are very effective for various big data applications, such as Spark SQL. The proposed cache policy scheduling algorithms with various eviction policies can improve around 20% hit ratio than that with a single eviction policy.

Functional dependencies (FDs) play a very important role in many data management tasks such as schema normalization, data cleaning, and query optimization. Meanwhile, there are ever-increasing application demands for efficient FD discovery on large-scale datasets. Unfortunately, due to huge runtime and memory overhead, the existing single-machine FD discovery algorithms are inefficient for large-scale datasets. Recently, distributed data-parallel computing has become the de facto standard for large-scale data processing. However, it is challenging to design an efficient distributed FD discovery algorithm. In this paper, we present SmartFD, which is an efficient and scalable algorithm for distributed FD discovery. First, we propose a novel attribute sorting-based algorithm framework. Next, to discover all the FDs grouped by a given attribute, we propose an efficient distributed algorithm Attribute-centric Functional Dependency Discovery (AFDD). In AFDD, we design an Fast Sampling and Early Aggregation (FSEA) mechanism to improve the efficiency of distributed sampling and propose a memory-efficient index-based method for distributed FD validation. Moreover, AFDD employs an attribute-parallel method to accelerate the pruning-and-generation of candidate FDs. Furthermore, we propose an adaptive switching strategy between distributed sampling and distributed validation based on the unified time-based efficiency metric. Also, we employ a distributed probing based method to make the switching strategy more accurate. Experimental results on Apache Spark reveal that SmartFD outperforms the state-of-the-art single-machine algorithm HyFD and the existing distributed algorithm HFDD with 3.2×-44.9× and 2.5×-455.7× speedup respectively. Moreover, SmartFD achieves good row scalability and column scalability. Additionally, SmartFD has sub-linear node scalability.

Feature extraction is a critical step in the task of action recognition. Hand-crafted features are often restricted because of their fixed forms and deep learning features are more effective but need large-scale labeled data for training. In this paper, we propose a new hierarchical Nonlinear Orthogonal Adaptive-Subspace Self-Organizing Map(NOASSOM) to adaptively and learn effective features from data without supervision. NOASSOM is extended from Adaptive-Subspace Self-Organizing Map (ASSOM) which only deals with linear data and is trained with supervision by the labeled data. Firstly, by adding a nonlinear orthogonal map layer, NOASSOM is able to handle the nonlinear input data and it avoids defining the specific form of the nonlinear orthogonal map by a kernel trick. Secondly, we modify loss function of ASSOM such that every input sample is used to train model individually. In this way, NOASSOM effectively learns the statistic patterns from data without supervision. Thirdly, we propose a hierarchical NOASSOM to extract more representative features. Finally, we apply the proposed hierarchical NOASSOM to efficiently describe the appearance and motion information around trajectories for action recognition. Experimental results on widely used datasets show that our method has superior performance than many state-of-the-art hand-crafted features and deep learning features based methods.

Statistical machine translation (SMT) is an important research branch in natural language processing (NLP). Similar to many other NLP applications, large scale training data can potentially bring higher translation accuracy for SMT models. However, the traditional single-node SMT model training systems can hardly cope with the fast-growing amount of large scale training corpus in the big data era, which makes the urgent requirement of efficient large scale machine translation model training systems. In this paper, we propose Seal, an efficient, scalable, and end-to-end offline SMT model training toolkit based on Apache Spark which is a widely-used distributed data-parallel platform. Seal parallelizes the training process of the entire three key SMT models that are the word alignment model, the translation model, and the N -Gram language model, respectively. To further improve the performance of the model training in Seal, we also propose a number of system optimization methods. In word alignment model training, by optimizing the block size tuning, the overhead of IO operation and communication is greatly reduced. In translation model training, by well encoding the training corpus, the data size transferred over the network can be reduced significantly, thus improving the overall training efficiency. We also optimize the maximum likelihood estimation (MLE) algorithm to solve the data skew issue on the join operation which is adopted both in the translation model training and the language model training. The experiment results show that Seal outperforms the well-known SMT training system Chaski with about 5× speedup for word alignment model training. For the syntactic translation model and language model training, Seal outperforms the existing cutting-edge tools with about 9~18× speedup and 8~9× speedup on average, respectively. On the whole, Seal outperforms the existing distributed system with 4~6× speedup and the single-node system with 9~60× speedup on average respectively. Besides, Seal achieves near-linear data and node scalability.

In the big data era, machine learning optimization algorithms usually need to be designed and implemented on widely-used distributed computing platforms, such as Apache Hadoop, Spark, and Flink. However, these general distributed computing platforms themselves do not focus on parallelizing machine learning optimization algorithms. In this paper, we present a parallel optimization algorithm framework for scalable machine learning, and empirically evaluate the synchronous Elastic Averaging SGD (EASGD) and other distributed SGD-based optimization algorithms. First, we design a distributed machine learning optimization algorithm framework based on Apache Spark by adopting the parameter server. Then, we design and implement the widely-used distributed synchronous EASGD and several other popular SGD-based optimization algorithms, such as Adadelta and Adam, on top of the framework. In addition, we evaluate the performance of synchronous distributed EASGD compared with other distributed optimization algorithms based on the same framework. Finally, to explore the optimal settings of mini-batch size in large-scale distributed optimization, we further analyze the empirical linear scaling rule originally proposed in the single-node environment. Experimental results show that our parallel optimization algorithm framework achieves good flexibility and scalability. And, the distributed synchronous EASGD runs over the proposed framework gains a competitive convergence performance and is about 5.7% faster than other distributed SGD-based optimization algorithms. Experimental results also verified that the empirical linear scaling rule only holds well before the mini-batch size exceeds certain threshold on large-scale benchmarks in the distributed environment.

In this paper, a new nonparametric Bayesian model called the dual sticky hierarchical Dirichlet process hidden Markov model (HDP-HMM) is proposed for mining activities from a collection of time series data such as trajectories. All the time series data are clustered. Each cluster of time series data, corresponding to a motion pattern, is modeled by an HMM. Our model postulates a set of HMMs that share a common set of states (topics in an analogy with topic models for document processing), but have unique transition distributions. The number of HMMs and the number of topics are both automatically determined. The sticky prior avoids redundant states and makes our HDP-HMM more effective to model multimodal observations. For the application to motion trajectory modeling, topics correspond to motion activities. The learnt topics are clustered into atomic activities which are assigned predicates. We propose a Bayesian inference method to decompose a given trajectory into a sequence of atomic activities. The sources and sinks in the scene are learnt by clustering endpoints (origins and destinations) of trajectories. The semantic motion regions are learnt using the points in trajectories. On combining the learnt sources and sinks, the learnt semantic motion regions, and the learnt sequence of atomic activities, the action represented by a trajectory can be described in natural language in as automatic a way as possible. The effectiveness of our dual sticky HDP-HMM is validated on several trajectory datasets. The effectiveness of the natural language descriptions for motions is demonstrated on the vehicle trajectories extracted from a traffic scene.

Graphs are effective tools for modeling complex data. Setting out from two basic substructures, random walks and trees, we propose a new family of context-dependent random walk graph kernels and a new family of tree pattern graph matching kernels. In our context-dependent graph kernels, context information is incorporated into primary random walk groups. A multiple kernel learning algorithm with a proposed l1,2-norm regularization is applied to combine context-dependent graph kernels of different orders. This improves the similarity measurement between graphs. In our tree-pattern graph matching kernel, a quadratic optimization with a sparse constraint is proposed to select the correctly matched tree-pattern groups. This augments the discriminative power of the tree-pattern graph matching. We apply the proposed kernels to human action recognition, where each action is represented by two graphs which record the spatiotemporal relations between local feature vectors. Experimental comparisons with state-of-the-art algorithms on several benchmark datasets demonstrate the effectiveness of the proposed kernels for recognizing human actions. It is shown that our kernel based on tree-pattern groups, which have more complex structures and exploit more local topologies of graphs than random walks, yields more accurate results but requires more runtime than the context-dependent walk graph kernel.

In the big data era, there are many demands for efficient and easy to use data replay services over large scale historical data. For example, stock security trading and on-line e-business services need historical data replay services to conduct system testing or ex-post review and analysis, just like replaying videos for security monitoring. Stream processing systems are designed for processing stream data, thus can not perform complex replay jobs over the static historical data. Database management systems support easy to use complex queries, but lack stream processing abilities. In this paper, we present a data replay model that combines the stream replay and complex query ability together, to allow the applications to replay large scale historical data from various data sources. First, to meet the demands of flexible replay semantics, we designed a set of easy to use replay operators to describe various replay behaviors and semantics. Users can use these operators to build up their complex replay jobs with diversified requirements. Then, we proposed a query mechanism to provide a flexible data loading service. Next, we presented the Penguin framework to support the proposed query-based replay model along with the replay operators. Penguin enables users to develop high-throughput and easy to use replay services over various large scale data sources with tunable replay speeds. To further improve the data replay performance, we proposed four system-level optimizations, including caching loading task results, cascading merging intermediate record queues, producing the replay queue in parallel, and caching remote file streams. Experimental results over replaying millions of records demonstrate that Penguin can achieve up to 4x and 144x speedup in data preparation and up to 16x and 9x speedup in replay speed compared to Apache Phoenix and Apache Hive respectively. As a case study, Penguin has been deployed in production environments of some Securities companies to provide online historical stock data replay services to large number of stock market users.

Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) are two typical kinds of neural networks. While CNN models have achieved great success on image recognition due to their strong abilities in abstracting spatial information from multiple levels, RNN models have not achieved significant progress in video analyzing tasks (e.g. action recognition), although RNN can inherently model temporal dependencies from videos. In this work, we propose a Sequential Convolutional Neural Network, denoted as SCNN, to extract effective spatial-temporal features from videos, thus incorporating the strengths of both convolutional operation and recurrent operation. Our SCNN model extends RNN to directly process feature maps, rather than vectors flattened from feature maps, to keep spatial structures of the inputs. It replaces the full connections of RNN with convolutional connections to decrease parameter numbers, computational cost, and over-fitting risk. Moreover, we introduce asymmetric convolutional layers into SCNN to reduce parameter numbers and computational cost further. Our final SCNN deep architecture used for action recognition achieves very good performances on two challenging benchmarks, UCF-101 and HMDB-51, outperforming many state-of-the-art methods.

We present a simple and effective ensemble method, Diversity Encouraging Ensemble (DEE), for deep convolutional networks to boost their performances. By training the convolutional network in two stages, we generate multiple component networks without adding any training cost. On the one hand, we modify the structure parameters of component networks in the training process to enlarge the diversities of the networks, which is found to be beneficial to improving the ensemble performance. On the other hand, we exploit monotonous decreasing learning rate schedule to accelerate the speed of deep network converging to different local minima, and we decrease the training time of integrating multiple networks to that of training a single network from traditional multi-step learning policy. We evaluate our ensemble method on two challenging action datasets, UCF-101 and HMDB-51, and obtain performance improvements from single deep network and other ensemble methods. Our results also outperform many state-of-the-art action recognition methods.

The multi-way join query has attracted considerable attention from research community for its importance in many big data analytic applications. For the multi-round multi-way join algorithm in distributed data-parallel platforms, the huge communication cost caused by shuffling large intermediate results over the network is the main bottleneck. The one-round multi-way join algorithm processes the join query in a single communication round, which can significantly reduce the communication cost in complex queries, including cyclic queries. However, the one-round method is not always superior to the multi-round method, because the intermediate result size of the multi-round method may the much smaller than the size of data shuffled in the one-round method. Therefore, it is challenging to choose the best multi-way join algorithm in practice. To solve this problem, in this paper, we present AutoMJ, an efficient framework for multi-way join queries. In AutoMJ, we propose a novel automatic join strategy selection model based on the size estimation of intermediate join results. AutoMJ chooses the multi-way join strategy with the minimal shuffle data size. In addition, we propose an optimized HyperCube algorithm for the one-round multi-way join. We have implemented the prototype of AutoMJ on the widely-used distributed data-parallel platform Apache Spark. Experiments show that for multi-way join queries with large intermediate results, the one-round join strategy can outperform the multi-round join strategy built in Spark SQL 1.2 - 159.3× faster. In contrast, the multi-round join strategy is 2.1 - 6.2× faster than the one-round method for the queries with small intermediate results. Experiments also show that the relative error of size estimation can be within 0.1 for the Twitter dataset and 0.25 for the Wikidata dataset. Furthermore, experiments verify that the automatic join strategy selection model is effective for choosing the optimal multi-way join algorithm.

With the rapid growth of opinionated contents, e.g. product reviews, sentiment analysis has drawn much attention from the researchers. The most fundamental task of sentiment analysis is document sentiment classification which aims to predict the overall sentiment (e.g. positive or negative) towards the opinion target in a review. There are usually various opinion sentences towards different aspects with different sentiments. Among them, the overall opinion towards the whole target should be more deterministic in document sentiment prediction. However, most existing methods treat all the sentences equally, thus, they may encounter difficulty especially when the sentiments of most aspect opinion sentences differ from the overall sentiment. To address this, we propose a novel method for document sentiment classification which adequately explores the effect of overall opinion sentences. The method is extended from structural SVM, and the overall opinion sentences are taken as the hidden variables for document sentiment. Experiments on several standard product review datasets show the effectiveness of our method.

In multi-instance learning (MIL), the relations among instances in a bag convey important contextual information in many applications. Previous studies on MIL either ignore such relations or simply model them with a fixed graph structure so that the overall performance inevitably degrades in complex environments. To address this problem, this paper proposes a novel multi-view multi-instance learning algorithm (MIL) that combines multiple context structures in a bag into a unified framework. The novel aspects are: (i) we propose a sparse -graph model that can generate different graphs with different parameters to represent various context relations in a bag, (ii) we propose a multi-view joint sparse representation that integrates these graphs into a unified framework for bag classification, and (iii) we propose a multi-view dictionary learning algorithm to obtain a multi-view graph dictionary that considers cues from all views simultaneously to improve the discrimination of the MIL. Experiments and analyses in many practical applications prove the effectiveness of the M IL.

This paper proposes a stock market prediction method exploiting sentiment analysis using financial microblogs (Sina Weibo). We analyze the microblog texts to find the financial sentiments, then combine the sentiments and the historical data of the Shanghai Composite Index (SH000001) to predict the stock market movements. Our framework includes three modules: Microblog Filter (MF), Sentiment Analysis (SA), and Stock Prediction (SP). The MF module is based on LDA to get the financial microblogs. The SA module first sets up a financial lexicon, then gets the sentiments of the microblogs obtained from the MF module. The SP module proposes a user-group model which adjusts the importance of different people, and combines it with stock historical data to predict the movement of the SH000001. We use about 6.1 million microblogs to test our method, and the result demonstrates our method is effective.

Matrix computation is the core of many massive data-intensive analytical applications such mining social networks, recommendation systems and nature language processing. Due to the importance of matrix computation, it has been widely studied for many years. In the Big Data ear, as the scale of the matrix grows, traditional single-node matrix computation systems can hardly cope with such large data and computation. Existing distributed matrix computation solutions are still not efficient enough, or have poor fault tolerance and usability. In this paper, we propose Marlin, an efficient distributed matrix computation library which is built on top of Spark. Marlin contains several distributed matrix operation algorithms and provides high-level matrix computation primitives for users. In Marlin, we proposed three distributed matrix multiplication algorithms for different situations. Based on this, we designed an adaptive model to choose the best approach for different problems. Moreover, to improve the computation performance, instead of naively using Spark, we put forward some optimizations including taking advantage of the native linear algebra library, reducing shuffle communication and increasing parallelism. Experimental results show that Marlin is over an order of magnitude faster than R (a widely-used statistical computing system) and the existing distributed matrix operation algorithms based on MapReduce. Moreover, Marlin achieves comparable performance to the specialized MPI-based matrix multiplication algorithm SUMMA but uses a general dataflow engine and gains common dataflow features such as scalability and fault tolerance.

In this paper, a multi-feature max-margin hierarchical Bayesian model (M3HBM) is proposed for action recognition. Different from existing methods which separate representation and classification into two steps, M3HBM jointly learns a high-level representation by combining a hierarchical generative model (HGM) and discriminative max-margin classifiers in a unified Bayesian framework. Specifically, HGM is proposed to represent actions by distributions over latent spatial temporal patterns (STPs) which are learned from multiple feature modalities and shared among different classes. For recognition, we employ Gibbs classifiers to minimize the expected loss function based on the max-margin principle and use the classifiers as regularization terms of M3HBM to perform Bayeisan estimation for classifier parameters together with the learning of STPs. In addition, multi-task learning is applied to learn the model from multiple feature modalities for different classes. For test videos, we obtain the representations by the inference process and perform action recognition by the learned Gibbs classifiers. For the learning and inference process, we derive an efficient Gibbs sampling algorithm to solve the proposed M3HBM. Extensive experiments on several datasets demonstrate both the representation power and the classification capability of our approach for action recognition.

In the era of big data, the volume of semantic data grows rapidly. The large scale semantic data contains a lot of significant but often implicit information that needs to be derived by reasoning. The semantic data reasoning is a challenging process. On one hand, the traditional single-node reasoning systems can hardly cope with such large amount of data due to the resource limitations. On the other hand, the existing large scale reasoning systems are not very efficient and scalable due to the complexity of reasoning process. In this paper, we propose Cichlid, an efficient distributed reasoning engine for the widely-used RDFS and OWL Horst rule sets. Cichlid is built on top of Spark. It implements parallel reasoning algorithms with the Spark RDD programming model. Further, we proposed the optimized parallel RDFS reasoning algorithm from three aspects, including data partition model, the execution order of reasoning rules and removing of duplicated data. Then, for the parallel OWL reasoning process, we optimized the most time-consuming parts, including large-scale data join, the transitive closure computation and the equivalent relation computation. In addition to above optimizations at the reasoning algorithm level, we also optimized the inner Spark execution mechanism by proposing an off-heap memory storage mechanism for RDD. This system-level optimization patch has been accepted and integrated into Apache Spark 1.0. The experimental results show that Cichlid is around 10 times faster on average than the state-of-the-art distributed reasoning systems for both large scale synthetic and real-world benchmarks. The proposed reasoning algorithms and engine also achieve excellent scalability and fault tolerance.

Interactions between moving targets often provide discriminative clues for multiple target tracking (MTT), though many existing approaches ignore such interactions due to difficulty in effectively handling them. In this paper, we model interactions between neighbor targets by pair-wise motion context, and further encode such context into the global association optimization. To solve the resulting global non-convex maximization, we propose an effective and efficient power iteration framework. This solution enjoys two advantages for MTT: First, it allows us to combine the global energy accumulated from individual trajectories and the between-trajectory interaction energy into a united optimization, which can be solved by the proposed power iteration algorithm. Second, the framework is flexible to accommodate various types of pairwise context models and we in fact studied two different context models in this paper. For evaluation, we apply the proposed methods to four public datasets involving different challenging scenarios such as dense aerial borne traffic tracking, dense point set tracking, and semi-crowded pedestrian tracking. In all the experiments, our approaches demonstrate very promising results in comparison with state-of-the-art trackers.

Graphs are a powerful tool to model structured objects, but it is nontrivial to measure the similarity between two graphs. In this paper, we construct a two-graph model to represent human actions by recording the spatial and temporal relationships among local features. We also propose a novel family of context-dependent graph kernels (CGKs) to measure similarity between graphs. First, local features are used as the vertices of the two-graph model and the relationships among local features in the intra-frames and inter-frames are characterized by the edges. Then, the proposed CGKs are applied to measure the similarity between actions represented by the two-graph model. Graphs can be decomposed into numbers of primary walk groups with different walk lengths and our CGKs are based on the context-dependent primary walk group matching. Taking advantage of the context information makes the correctly matched primary walk groups dominate in the CGKs and improves the performance of similarity measurement between graphs. Finally, a generalized multiple kernel learning algorithm with a proposed l12-norm regularization is applied to combine these CGKs optimally together and simultaneously train a set of action classifiers. We conduct a series of experiments on several public action datasets. Our approach achieves a comparable performance to the state-of-the-art approaches, which demonstrates the effectiveness of the two-graph model and the CGKs in recognizing human actions.

Inspired by the recent success of hierarchical representation, we propose a new hierarchical variant of latent Dirichlet allocation (h-LDA) for action recognition. The model consists of an appearance group and a motion group, and we introduce a new hierarchical structure including two-layer topics in each group to learn the spatial temporal patterns (STPs) of human actions. The basic idea is that the two-layer topics are used to model the global STPs and the local STPs of the actions respectively. Two groups of discrete words are generated from two complementary kinds of features for each group. Each topic learned in these two groups is used to describe a particular aspect of the actions. Specifically, the mid-level topics are learned to describe the local STPs by including the geometric structure information in the lower-level words. The top-level topics are learned from the mid-level topics and are the mixture distribution of the local STPs, which makes the top-level topics appropriate to represent the global STPs. In addition, we give the learning and inference process by Gibbs sampling with reasonable assumptions. Finally, each sample is discriminatively represented as the probabilistic distribution over the global STPs learned by the proposed h-LDA. Experimental results on two datasets demonstrate the effectiveness of our approach for action recognition.

In this paper, we address the problem of human action recognition through combining global temporal dynamics and local visual spatio-temporal appearance features. For this purpose, in the global temporal dimension, we propose to model the motion dynamics with robust linear dynamical systems (LDSs) and use the model parameters as motion descriptors. Since LDSs live in a non-Euclidean space and the descriptors are in non-vector form, we propose a shift invariant subspace angles based distance to measure the similarity between LDSs. In the local visual dimension, we construct curved spatio-temporal cuboids along the trajectories of densely sampled feature points and describe them using histograms of oriented gradients (HOG). The distance between motion sequences is computed with the Chi-Squared histogram distance in the bag-of-words framework. Finally we perform classification using the maximum margin distance learning method by combining the global dynamic distances and the local visual distances. We evaluate our approach for action recognition on five short clips data sets, namely Weizmann, KTH, UCF sports, Hollywood2 and UCF50, as well as three long continuous data sets, namely VIRAT, ADL and CRIM13. We show competitive results as compared with current state-of-the-art methods.

In this paper, we propose using high-level action units to represent human actions in videos and, based on such units, a novel sparse model is developed for human action recognition. There are three interconnected components in our approach. First, we propose a new context-aware spatial-temporal descriptor, named locally weighted word context, to improve the discriminability of the traditionally used local spatial-temporal descriptors. Second, from the statistics of the context-aware descriptors, we learn action units using the graph regularized nonnegative matrix factorization, which leads to a part-based representation and encodes the geometrical information. These units effectively bridge the semantic gap in action recognition. Third, we propose a sparse model based on a joint l2,1-norm to preserve the representative items and suppress noise in the action units. Intuitively, when learning the dictionary for action representation, the sparse model captures the fact that actions from the same class share similar units. The proposed approach is evaluated on several publicly available data sets. The experimental results and analysis clearly demonstrate the effectiveness of the proposed approach.

In this paper, we present a new geometric-temporal representation for visual action recognition based on local spatio-temporal features. First, we propose a modified covariance descriptor under the log-Euclidean Riemannian metric to represent the spatio-temporal cuboids detected in the video sequences. Compared with previously proposed covariance descriptors, our descriptor can be measured and clustered in Euclidian space. Second, to capture the geometric-temporal contextual information, we construct a directional pyramid co-occurrence matrix (DPCM) to describe the spatio-temporal distribution of the vector-quantized local feature descriptors extracted from a video. DPCM characterizes the co-occurrence statistics of local features as well as the spatio-temporal positional relationships among the concurrent features. These statistics provide strong descriptive power for action recognition. To use DPCM for action recognition, we propose a directional pyramid co-occurrence matching kernel to measure the similarity of videos. The proposed method achieves the state-of-the-art performance and improves on the recognition performance of the bag-of-visual-words (BOVWs) models by a large margin on six public data sets. For example, on the KTH data set, it achieves 98.78% accuracy while the BOVW approach only achieves 88.06%. On both Weizmann and UCF CIL data sets, the highest possible accuracy of 100% is achieved.

In this paper, we formulate human action recognition as a novel Multi-Task Sparse Learning(MTSL) framework which aims to construct a test sample with multiple features from as few bases as possible. Learning the sparse representation under each feature modality is considered as a single task in MTSL. Since the tasks are generated from multiple features associated with the same visual input, they are not independent but inter-related. We introduce a Beta process(BP) prior to the hierarchical MTSL model, which efficiently learns a compact dictionary and infers the sparse structure shared across all the tasks. The MTSL model enforces the robustness in coefficient estimation compared with performing each task independently. Besides, the sparseness is achieved via the Beta process formulation rather than the computationally expensive L1 norm penalty. In terms of non-informative gamma hyper-priors, the sparsity level is totally decided by the data. Finally, the learning problem is solved by Gibbs sampling inference which estimates the full posterior on the model parameters. Experimental results on the KTH and UCF sports datasets demonstrate the effectiveness of the proposed MTSL approach for action recognition.

Spatio-temporal interest points serve as an elementary building block in many modern action recognition algorithms, and most of them exploit the local spatio-temporal volume features using a Bag of Visual Words (BOVW) representation. Such representation, however, ignores potentially valuable information about the global spatio-temporal distribution of interest points. In this paper, we propose a new global feature to capture the detailed geometrical distribution of interest points. It is calculated by using the R transform which is defined as an extended 3D discrete Radon transform, followed by applying a two-directional two-dimensional principal component analysis. Such R feature captures the geometrical information of the interest points and keeps invariant to geometry transformation and robust to noise. In addition, we propose a new fusion strategy to combine the R feature with the BOVW representation for further improving recognition accuracy. We utilize a context-aware fusion method to capture both the pairwise similarities and higher-order contextual interactions of the videos. Experimental results on several publicly available datasets demonstrate the effectiveness of the proposed approach for action recognition.

This paper presents a new method to detect human actions in video by combining sparse appearance features and dense motion features in the unified random forest framework. We compute sparse appearance features to capture the main appearance changes and dense motion features to capture the tiny motion changes in the video. We take advantage of the randomization of channel selection in random trees to combine these two complementary types of features. In addition, linear classification is applied to grow each tree with high efficiency. Each leaf in these trees stores the class distribution and location information of the training samples and action detection for the test video is accomplished by Hough voting of the leaves in each tree. Experimental results demonstrate that our method achieves the state-of-the-art performance on two datasets.

In this paper, a new nonparametric Bayesian model called Sticky Multimodal Dual Hierarchical Dirichlet Process Hidden Markov Model (SMD-HDP-HMM) is proposed for mining activities from a collection of time series. An activity is modeled as an HMM where each state corresponds to an atomic activity. By extensively using Dirichlet Process (DP), multiple HMMs sharing a common set of states are learned and the numbers of HMMs and states are both automatically determined. Each time series is modeled to be generated by one of the HMMs such that all time series are clustered into activities. Simultaneously state sequences for time series are learned and each of them is decomposed into a sequence of atomic activities. Experimental results on KTH activity dataset demonstrate the advantage of our method.

A large scale of duplicate and near-duplicate web pages on the Internet create a lot of problems for search engines. Currently each single duplicate and near-duplicate web document detection algorithms cannot achieve both good performance and accuracy. Also most of them are designed to process English documents and not able to use for Chinese documents. This paper presents an integrated algorithm, KMatch, for near-duplicate document detection of large scale Chinese Web pages. First of all, KMatch employs Chinese segmentation algorithm to prepare Chinese words into meaningful features to compress documents. Then keywords matching technique is used to improve the accuracy of document detection. For further accuracy improvement, KMatch also combines IMatch algorithms to filter out the noise contents of a web document and retain the body text. To improve detection performance, we integrate the Shingling algorithm to compress huge datasets into smaller ones. Finally, to further improve the detection performance on large scale Chinese web pages, we design and implement KMatch algorithm in parallel with MapReduce. The experimental results show that our approach achieves both high precision and recall, and the parallelized algorithm with MapReduce achieves good performance and scalability when dealing with large scale of datasets.

Measures of graph similarity have a broad range of applications but involve compute-intensive process. Similarity flooding algorithm is an efficient algorithm for comparing the similarity of graphs of small size and small datasets. However, nowadays more and more large-scale graph applications emerge and existing stand-alone similarity flooding algorithm cannot efficiently conduct the similarity comparison process for large scale graph datasets in acceptable time. This paper presents a parallelized similarity flooding algorithm with MapReduce for large-scale graph datasets. The experimental results demonstrate that the parallelized algorithm achieves significant performance improvement compared to the stand-alone similarity flooding algorithm. Experimental results also reveal that the parallelized algorithm can obtain excellent speedup when the size of cluster increases.

How to fuse static and dynamic information is a key issue in event analysis. In this paper, a top-down motion guided fusing method is proposed for recognizing events in an unconstrained news video. In the method, the static information is represented as a Bag-of-SIFT-features and motion information is employed to generate event specific attention map to direct the sampling of the interest points. We build class-specific motion histograms for each event so as to give more weight on the interest points that are discriminative to the corresponding event. Experimental results on TRECVID 2005 video corpus demonstrate that the proposed method can improve the mean average accuracy of recognition.

Behavior recognition is an attractive direction in the computer vision domain. In this paper, we propose a novel behavior recognition method based on prototype learning using metric learning. Prototype learning algorithm can improve the classification performance of nearest-neighbor classifier, reduce the storage and computation requirements. And the metric learning algorithm is used to advance the performance of the prototype learning. In this paper, We use a kind of compound feature including local feature and motion feature to recognize human behaviors. The experimental results show the effectiveness of our method.