Xi Wang, Songlei Jian, Shasha Li, Xiaopeng Li, Zhaoye Li, Bin Ji, Baosheng Wang, Jie Yu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

As the pivot to understand the property of the whole graph, substructure extraction plays an essential role in graph representation learning. Following a “convolution + pooling” paradigm, recent progress mainly relies on cluster pooling to achieve substructure compression. However, it suffers from high complexity caused by dense implementation and poor effectiveness resulting from rigid node assignment. Besides, the separating design of the universally adopted paradigm leads to substantial redundancy of model architecture and underutilization of convolution capability. To overcome the above dilemmas, in this paper, we propose GAS, a sparsely implemented graph representation learning model that can adaptively discover critical substructures in a flexible manner through a simplified paradigm. Specifically, GAS reformulates substructure construction from a dense node assignment problem to a sparse node absorbing problem, reducing complexity by eliminating dense assignment and complex matrix multiplication. Besides, we propose an adaptive absorbing strategy to overcome the suboptimality caused by rigidity. It allows multi-substructure compatibility for high-influential nodes while single-substructure qualification for low-influential nodes, making a leap in the flexible division of substructures with better local clarity. In addition, theoretical analysis provides evidence that a simplified paradigm with merely convolutional operation is adequate for reinforced efficiency and intensive substructure extracting quality. To this end, we design AConv layer, which empowers convolution operation with substructure construction capability, leading to a simplified graph representation learning paradigm without extra pooling layers. Extensive experiments on graph classification and molecular property prediction tasks demonstrate a significant improvement of GAS with an average performance gain of 8.77% and 4.65% compared with SOTA competitors. Investigation of parameter size and training time quantifies the superiority of GAS with high efficiency and tractability for large-scale graphs. Ablation study confirms the necessity of the adaptive node absorbing strategy. A visualization further demonstrates the advantage of the method.

Few-shot Anomaly Detection (AD) for images aims to detect anomalies with few-shot normal samples from the target dataset. It is a crucial task when only few samples can be obtained, and it is challenging since it needs to be generalized to different domains. Existing methods try to enhance the generalizability of AD by incorporating large vision-language models (LVLMs).However, how to transform category semantic information in LVLMs into anomaly information to improve the generalizability of AD remains a challenge facing existing methods.To address the challenge, we propose a few-shot AD method called MetaCAN, a novel category-to-anomaly network trained with AD meta-learning scheme based on an LVLM. Specifically, MetaCAN constructs the auxiliary training data and multiple tasks based on different categories to perform AD meta-learning, which ensures that the optimization toward the achievement of optimal anomaly detection across all categories. Moreover, MetaCAN introduces an image-image anomaly discriminator and an image-text anomaly detector to fully exploit the powerful multimodal semantic representations during auxiliary training. Once trained on auxiliary datasets, MetaCAN can be applied directly to other target datasets without retraining. Extensive experiments on six real-world datasets demonstrate that MetaCAN achieves state-of-the-art performance on cross-domain and cross-category anomaly detection tasks compared with existing methods.

Graph classification is a fundamental task in graph machine learning, aiming to categorize entire graphs based on their structural information and node attributes. The graph classification task wildly existed in collaboration technology, like knowledge graph, social network analysis. However, the structure of graph samples often varies greatly within the graph dataset. This inherent diversity poses a challenge for existing Graph Neural Networks (GNNs) to make graph classification, as they often uniformly utilize the same model for all graph samples. In this paper, we propose a novel framework, GraphSGA, which enhances Graph Neural Networks for graph classification via tailoring models for each graph based on their strucure characteristics. Our framework consists of a graph topology encoder and a structure-guided adaptor. The encoder converts the graph topology into a vector, offering a rich structural representation that guides the adaptor to accordingly refine graph embeddings. This enables a structure-aware adaption to model when classify graphs with distinct structures, improving the overall classification accuracy. Comprehensive experimental results show that our method outperforms baseline methods with a significant improvement and is superior to state-of-the-art methods. Also, the proposed approach possesses good compatibility with various GNN backbones like GCN, GAT, GIN.

Xi Wang, Songlei Jian, Shasha Li, Xiaopeng Li, Bin Ji, Ma Jun, Xiaodong Liu, Jing Wang, Jianfeng Zhang, Jie Yu, Feilong Bao, Wangbaosheng. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Deep time series anomaly detection (TSAD) essentially relies on learning data "normality". Current approaches leverage various neural network architectures, including RNNs, CNNs, Transformers, and graph neural networks, effectively modeling temporal and inter-variable dependencies within time series data. However, these approaches often overlook the equally crucial asynchronous inter-variable dependencies, potentially missing anomalous patterns that arise from intricate, temporally offset interactions among variables. Therefore, we propose a novel time series anomaly detection method by perceiving comprehensive intrinsic dependencies of time series data. Specifically, we propose a Comprehensive Dependency-aware Attention (CDA) that computes attention between each element and its criss-cross counterparts (i.e., intra-variable values across the whole temporal dimension and synchronous values of different variables) and captures asynchronous inter-variable dependency by repeating this step. Additionally, to facilitate more effective downstream dependency learning, we propose a Frequency Domain Rectifier (FDR) to capture the signal’s underlying patterns and reduce noise in the frequency domain. We replace the self-attention in the classic Transformer with CDA, and introduce the FDR module to construct a new method for TSAD, i.e., Frequency-enhanced Comprehensive Dependency Attention-based Time Series Anomaly Detection (FCDATA). Extensive experiments show that: (i) our detection method significantly outperforms state-of-the-art competitors by 15% in the F1-score; and (ii) by incorporating the proposed network architecture, existing anomaly detectors achieve over 20% performance gain compared to current temporal network structures.

Depth completion is a popular research direction in the field of depth estimation. The fusion of color and depth features is the critical challenge in this task, mainly due to the asymmetry between the rich scene details in color images and the sparse pixels in depth maps. To tackle this issue, we design an efficient Gated Cross-Attention Network that propagates confidence via a gating mechanism, simultaneously extracting and refining key information in both color and depth branches to achieve local spatial feature fusion. Additionally, we incorporate a Transformer-based attention network in low-dimensional space to effectively fuse global features and increase the network’s receptive field. At the same time, we use the Ray Tune mechanism with the AsyncHyperBandScheduler and the HyperOptSearch algorithm to automatically search for the optimal number of module iterations, which also allows us to achieve performance comparable to state-of-the-art methods. We conduct experiments on both indoor and outdoor scene datasets. Our fast network ranked first among real-time methods (below 30ms and 100ms), and our accurate network ranked first among all methods on the KITTI official website at the time of submission.

As the complexity of models and the scale of parameters grows rapidly, the limited memory capacity of intelligent acceleration devices such as GPUs has become a constraint on the development of large-scale deep neural networks (DNNs). Deep learning frameworks employ caching allocators to manage memory pools through a "split-merge" mechanism, enhancing GPU memory utilization while inevitably introducing memory fragmentation. While common memory optimization techniques effectively alleviate memory pressure during the training phase of DNNs, they also exacerbate the fragmentation issue. Existing methods rely on pre-acquired runtime information to guide memory allocation, which fails to adapt to dynamic and complex large-scale DNNs training scenarios. This paper proposes MOVie, a GPU memory optimization method for large-scale DNNs training based on virtual memory management. MOVie utilizes an over-subscribed virtual memory merging mechanism and a dynamic programming algorithm to combine non-contiguous memory blocks. Experimental results demonstrate that for various large-scale DNNs, MOVie can improve GPU memory utilization by up to 21.76% and achieves an average performance improvement of 1.5 times compared to the state-of-the-art methods.

Deep learning inference tasks develop towards the edge due to their latency requirements and privacy issues. However, edge devices are limited by their power consumption and size, and generally have limited resources. The convolutional neural networks (CNN) is commonly used in image processing tasks which contains a large number of convolutional layers, accounting for more than 95% of the calculation time in most general used CNN model. This paper proposes a general convolutional layer optimization method called MACA, we implement and optimize a variety of convolution operators and design a memory-aware convolution operator automatic selection strategy to select appropriate operator, without modifying user code. Finally, we integrate MACA into PyTorch and conduct extensive experiments. The results show that when memory resources are sufficient, MACA can effectively increase the inference by 36.10% on average, and can reduce memory usage by an average of 29.14% to complete inference when resources are tight. This paper provides an effective solution for deploying deep learning models on resource-constrained edge devices.

Link prediction in directed graphs offers a solution for uncovering detailed and accurate relationships among distinct entities. Unlike conventional link prediction in undirected graphs, the task becomes more intricate in directed graphs as it involves predicting both associations and orientations. Existing methods simply apply classic graph embedding techniques to learn node representations, followed by mapping representations of corresponding node pairs into probabilities indicating potential links. However, the inadequate capture of orientation information and sole reliance on node representations for prediction hinder the effective differentiation of orientation, thereby impeding the link prediction accuracy. In response, we introduce DGLP, an orientation-aware link prediction method tailored for directed graphs. DGLP utilizes the incidence matrix to learn both node and edge representations, effectively capturing structural and orientation information. By leveraging edge representations, DGLP achieves accurate link prediction in directed graphs without relying solely on implicit node representations. Experiments across six datasets demonstrate the effectiveness of DGLP, achieving a 1.2x improvement in prediction results.

Graph pooling technique in GNNs for learning expressive graph-level representation is critical yet still chal-lenging. Existing pooling methods either struggle to capture local substructures or fail to utilize high-order dependency, thus diminishing the expression capability. To solve this problem, we propose HAP, a hierarchical graph-level representation learning framework adaptively sensitive to graph structures. Specifically, HAP utilizes a novel cross-level attention mechanism MOA to naturally focus more on the close neighborhood while effectively capturing higher-order dependency. It also learns a global graph content GCont that extracts the graph pattern properties to stabilize the pre- and post-coarsening graph content, thus providing global guidance in graph coarsening. Experiments show that HAP significantly outperforms the state-of-the-art graph pooling methods.

To improve data processing and learning capabilities, DNN models are developing towards increasingly complex structures and a huge number of parameters. However, the limited memory capacity of specialized acceleration chips such as GPUs has become a critical constraint in exploring advanced DNN models. Memory swapping and recomputation are two effective GPU memory optimization techniques that enable the training of larger models in memory-constrained hardware environments. Existing methods mostly require prior knowledge of the complete computational graph, which is not suitable for the dynamic computational graph mode adopted by mainstream deep learning frameworks. This paper proposes DOME, a tensor-level GPU memory optimization method for dynamic computational graphs. DOME combines tensor swapping and recomputation, inspired by cache replacement algorithms, to release the least costly tensors when GPU memory is insufficient, thereby supporting larger batch sizes during model training. DOME employs a heuristic evaluation function with higher training efficiency, estimating the cost of tensor swapping or recomputation using information obtained online at runtime. Experiments demonstrate that, for different DNN models, DOME can increase the maximum training batch size by 1.56 to 2.33 times. Additionally, DOME can improve training throughput by 9.8% to 35.0% compared to SOTA methods.

Deep learning (DL) is characterized by its demanding computational and memory requirements, which creates a significant challenge when deploying on edge devices. These devices often have limited computational capabilities and constrained resources. Most existing methods primarily focus on model-level techniques, such as model pruning or parameter quantization, to reduce model size and computation for accelerating inference. Considering the prevalent programming paradigms in DL, we propose a host memory optimization method, namely HMO, which can be integrated into DL programming framework, e.g., PyTorch, to improve the inference efficiency of DL models without modifying any model code. We particularly focus on memory optimization for intermediate variables in inference, aiming to enhance inference speed while maintaining a lower memory footprint. HMO involves a single profiling of inference to gather memory statistics about intermediate variables. These statistics are then used to guide subsequent inference. Additionally, we incorporate huge pages in operating systems to improve the memory access performance of HMO. Our experimental results show that HMO can achieve an average inference latency optimization ratio of 20.13 % compared with native PyTorch on six typical DL image representation models while effectively managing memory usage. Importantly, this is achieved without compromising model accuracy.

Time series anomaly detection is instrumental in maintaining system availability in various domains. Current work in this research line mainly focuses on learning data normality deeply and comprehensively by devising advanced neural network structures and new reconstruction/prediction learning objectives. However, their one-class learning process can be misled by latent anomalies in training data (i.e., anomaly contamination) under the unsupervised paradigm. Their learning process also lacks knowledge about the anomalies. Consequently, they often learn a biased, inaccurate normality boundary. To tackle these problems, this paper proposes calibrated one-class classification for anomaly detection, realizing contamination-tolerant, anomaly-informed learning of data normality via uncertainty modeling-based calibration and native anomaly-based calibration. Specifically, our approach adaptively penalizes uncertain predictions to restrain irregular samples in anomaly contamination during optimization, while simultaneously encouraging confident predictions on regular samples to ensure effective normality learning. This largely alleviates the negative impact of anomaly contamination. Our approach also creates native anomaly examples via perturbation to simulate time series abnormal behaviors. Through discriminating these dummy anomalies, our one-class learning is further calibrated to form a more precise normality boundary. Extensive experiments on ten real-world datasets show that our model achieves substantial improvement over sixteen state-of-the-art contenders.

Multimodal rumor detection aims at detecting rumors using information from textual and visual modalities. The most critical difficulty in multimodal rumor detection lies in capturing both the intra-modal and inter-modal relationships from multimodal data. However, existing methods mainly focus on the multimodal fusion process while paying little attention to the intra-modal relationships. To address these limitations, we propose a multimodal rumor detection method with deep metric learning (MRML) to effectively extract multimodal relationships of news for detecting rumors. Specifically, we design the metric-based triplet learning to extract the intra-modal relationships between rumors and non-rumors in every modality and the contrastive pairwise learning to capture the inter-modal relationships across multimodal. Extensive experiments on two real-world multimodal datasets show the superior performance of our rumor detection method.

Due to the unsupervised nature of anomaly detection, the key to fueling deep models is finding supervisory signals. Different from current reconstruction-guided generative models and transformation-based contrastive models, we devise novel data-driven supervision for tabular data by introducing a characteristic -- scale -- as data labels. By representing varied sub-vectors of data instances, we define scale as the relationship between the dimensionality of original sub-vectors and that of representations. Scales serve as labels attached to transformed representations, thus offering ample labeled data for neural network training. This paper further proposes a scale learning-based anomaly detection method. Supervised by the learning objective of scale distribution alignment, our approach learns the ranking of representations converted from varied subspaces of each data instance. Through this proxy task, our approach models inherent regularities and patterns within data, which well describes data "normality". Abnormal degrees of testing instances are obtained by measuring whether they fit these learned patterns. Extensive experiments show that our approach leads to significant improvement over state-of-the-art generative/contrastive anomaly detection methods.

Log-based anomaly detection plays an essential role in various system reliability-related fields including software reliability, network reliability, and so on. System log data is a kind of semi-structured heterogeneous data that contains both semantic parts and numeric variables which both reflect the abnormal behavior of the system. However, existing log-based anomaly detection methods fail to capture the numeric information in raw data which makes them degrade a lot when only limited labeled data is available. To comprehensively capture the semantic and numeric information to enhance anomaly detection, we propose LogRep, a novel representation-based log anomaly detection method that captures both semantic and numeric information in the learned representations. The newly proposed position-aware numeric representation learning module and the attention-based representation fusion module in LogRep solve the heterogeneity problem well in log data. Due to the high quality of learned log representation, LogRep can achieve a comparable anomaly detection performance with SOTA methods while the training data used in LogRep is two orders of magnitude less than that used in SOTA methods. When reducing the training data scale, the performance of SOTA methods drops a lot, while LogRep keeps a stable good performance on two public HDFS dataset, BGL dataset, and one self-collected dataset. Specifically, LogRep achieves the 10.6% and 5.8% improvements over the second-best method in terms of F1 score on the BGL and HDFS datasets when only 1% training data are available respectively.

Graph neural networks (GNN) have been proven to be mature enough for handling graph-structured data on node-level graph representation learning tasks. However, the graph pooling technique for learning expressive graph-level representation is critical yet still challenging. Existing pooling methods either struggle to capture the local substructure or fail to effectively utilize high-order dependency, thus diminishing the expression capability. In this paper we propose HAP, a hierarchical graph-level representation learning framework, which is adaptively sensitive to graph structures, i.e., HAP clusters local substructures incorporating with high-order dependencies. HAP utilizes a novel cross-level attention mechanism MOA to naturally focus more on close neighborhood while effectively capture higher-order dependency that may contain crucial information. It also learns a global graph content GCont that extracts the graph pattern properties to make the pre- and post-coarsening graph content maintain stable, thus providing global guidance in graph coarsening. This novel innovation also facilitates generalization across graphs with the same form of features. Extensive experiments on ten datasets show that HAP significantly outperforms twelve popular graph pooling methods on graph classification task with an maximum accuracy improvement of 20.18%, and exceeds the performance of state-of-the-art graph matching and graph similarity learning algorithms by over 3.42% and 16%.

Graph pooling plays a vital role in learning graph embeddings. Due to the lack of label information, unsupervised graph pooling has received much attention, primarily via mutual information (MI). However, most existing MI-based pooling methods only preserve node features while overlooking the hierarchical substructural information. In this paper, we propose SMIP, a novel unsupervised hierarchical graph pooling method based on substructure-sensitive MI maximization. SMIP reconstructs a hard-style substructure encoder based on cluster-based pooling paradigm, and trains it with two substructure-sensitive MI-based objectives, i.e., node-substructure MI and node-node MI. The node-substructure MI guides to transfer maximum node feature information into corresponded substructures and the node-node MI guarantees a more accurate node allocation. Moreover, to avoid extra computation of augmented graphs and prevent noise information during MI estimation, we propose a local-scope contrastive MI estimation method, making SMIP more potent in capturing intrinsic features of the input graph. Experiments on six benchmark graph classification datasets demonstrate that our hierarchical deep learning approach outperforms all state-of-the-art unsupervised GNN-based methods and even surpasses the performance of nine supervised ones. Generalization study shows that the proposed substructure-sensitive MI objective can be successfully embedded into other cluster-based pooling methods to improve their performance.

Search space, searching method, and candidate evaluation scheme are critical to the success of Neural Architecture Search (NAS), especially the evaluation strategy. An effective and efficient neural architecture performance evaluator could successfully save computing costs and search time while guiding the NAS process to the optimal solution as fast as possible. Most existing NAS algorithms attempt to compute the absolute accuracy of the candidate architecture, which is almost impossible to achieve and meaningless to the final performance improvement. In this paper, we propose ERNAS, a novel neural architecture performance evaluation approach that optimizes the ranking of the candidate architecture performance, rather than the absolute accuracy itself. With the help of ERNAS, many existing NAS methods could achieve better performance without further evaluation. The experimental results demonstrate that ERNAS can be trained effectively enough with extremely limited training data (423 neural architectures randomly sampled form NAS-Bench-101, which is only 0.1% of the entire search space). The accuracy of the neural architecture search result produced by ERNAS is greater than that of the SOTA methods.

The majority of existing few-shot learning methods describe image relations with binary labels. However, such binary relations are insufficient to teach the network complicated real-world relations, due to the lack of decision smoothness. Furthermore, current few-shot learning models capture only the similarity via relation labels, but they are not exposed to class concepts associated with objects, which is likely detrimental to the classification performance due to underutilization of the available class labels. For instance, children learn the concept of tiger from a few of actual examples as well as from comparisons of tiger to other animals. Thus, we hypothesize that both similarity and class concept learning must be occurring simultaneously. With these observations at hand, we study the fundamental problem of simplistic class modeling in current few-shot learning methods. We rethink the relations between class concepts, and propose a novel Absolute-relative Learning paradigm to fully take advantage of label information to refine the image an relation representations in both supervised and unsupervised scenarios. Our proposed paradigm improves the performance of several state-of-the-art models on publicly available datasets.

Script event prediction is a significant task to extract event evolution rules from text. It requires to capture the intra-rationality and inter-relationship of events. The previous methods have made great progress, but they can not fully capture the inner features of events. This paper proposes an event representation layer with argument-level attention to capture the relationship between predicate and arguments, which benefit to assess the rationality of the event. Besides, we integrate the enhanced event representation with the multi-level event similarity scores to obtain more robust performance. The multi-level similarity consists of scene similarity and semantic consistency. We utilize the similarity between candidates and individual events to evaluate the scene similarity. Semantic consistency is evaluated by the hidden state of Gate Recurrent Unit (GRU). Unlike the previous work, we only use the last hidden state of GRU instead of all hidden states to reduce training time. In addition, we adopt adversarial training to improve the robustness and generalization of the model. The experimental results verify the effectiveness of our model compared to these baseline methods. This paper also demonstrates what the attention mechanism has learned to enhance the interpretability of our model.

Cloud computing is developing rapidly. Data centers are important infrastructures of cloud service and JointCloud structure. DRAM failure is one of the main causes which can lead to node outage in data centers. This paper proposes a decision-tree-based DRAM failure prediction method for large-scale data centers of cloud service. We utilize the first public-available DRAM failure prediction dataset released in PAKDD 2021 AIOps competition. We construct a suite of handcrafted features based on the system kernel log data and MCA log data. Feature engineering is detailedly introduced in this paper, which can inspire and foster future research in this field. Harnessing the power of a state-of-the-art classifier (i.e., XGBoost), our method can effectively and timely predict DRAM failures. Our solution has good performance on the PAKDD 2021 dataset, it can generally achieve more than 60% precision in the validation phase. Extensive experiments investigate the performance of variants of our method to validate the significance of different strategies in the proposed solution.

Outlier detection is an important task in many domains and is intensively studied in the past decade. Further, how to explain outliers, i.e., outlier interpretation, is more significant, which can provide valuable insights for analysts to better understand, solve, and prevent these detected outliers. However, only limited studies consider this problem. Most of the existing methods are based on the score-and-search manner. They select a feature subspace as interpretation per queried outlier by estimating outlying scores of the outlier in searched subspaces. Due to the tremendous searching space, they have to utilize pruning strategies and set a maximum subspace length, often resulting in suboptimal interpretation results. Accordingly, this paper proposes a novel Attention-guided Triplet deviation network for Outlier interpretatioN (ATON). Instead of searching a subspace, ATON directly learns an embedding space and learns how to attach attention to each embedding dimension (i.e., capturing the contribution of each dimension to the outlierness of the queried outlier). Specifically, ATON consists of a feature embedding module and a customized self-attention learning module, which are optimized by a triplet deviation-based loss function. We obtain an optimal attention-guided embedding space with expanded high-level information and rich semantics, and thus outlying behaviors of the queried outlier can be better unfolded. ATON finally distills a subspace of original features from the embedding module and the attention coefficient. With the good generality, ATON can be employed as an additional step of any black-box outlier detector. A comprehensive suite of experiments is conducted to evaluate the effectiveness and efficiency of ATON. The proposed ATON significantly outperforms state-of-the-art competitors on 12 real-world datasets and obtains good scalability w.r.t. both data dimensionality and data size.

The cross-domain representation learning plays an important role in tasks including domain adaptation and transfer learning. However, existing cross-domain representation learning focuses on building one shared space and ignores the unlabeled data in the source domain, which cannot effectively capture the distribution and structure heterogeneities in cross-domain data. To address this challenge, we propose a new cross-domain representation learning approach: MUltiple Lipschitz-constrained AligNments (MULAN) on partially-labeled cross-domain data. MULAN produces two representation spaces: a common representation space to incorporate knowledge from the source domain and a complementary representation space to complement the common representation with target local topological information by Lipschitz-constrained representation transformation. MULAN utilizes both unlabeled and labeled data in the source and target domains to address distribution heterogeneity by Lipschitz-constrained adversarial distribution alignment and structure heterogeneity by cluster assumption-based class alignment while keeping the target local topological information in complementary representation by self alignment. Moreover, MULAN is effectively equipped with a customized learning process and an iterative parameter updating process. MULAN shows its superior performance on partially-labeled semi-supervised domain adaptation and few-shot domain adaptation and outperforms the state-of-the-art visual domain adaptation models by up to 12.1%.

Classic recommender systems face challenges in addressing the data sparsity and cold-start problems with only modeling the user-item relation. An essential direction is to incorporate and understand the additional heterogeneous relations, e.g., user-user and item-item relations, since each user-item interaction is often influenced by other users and items, which form the user’s/item’s influential contexts. This induces important yet challenging issues, including modeling heterogeneous relations, interactions, and the strength of the influence from users/items in the influential contexts. To this end, we design Influential-Context Aggregation Units (ICAU) to aggregate the user-user/item-item relations within a given context as the influential context embeddings. Accordingly, we propose a Heterogeneous relations-Embedded Recommender System (HERS) based on ICAUs to model and interpret the underlying motivation of user-item interactions by considering user-user and item-item influences. The experiments on two real-world datasets show the highly improved recommendation quality made by HERS and its superiority in handling the cold-start problem. In addition, we demonstrate the interpretability of modeling influential contexts in explaining the recommendation results.

The formation of a complex network is highly driven by multi-aspect node influences and interactions, reflected on network structures and the content embodied in network nodes. Limited work has jointly modeled all these aspects, which typically focuses on topological structures but overlooks the heterogeneous interactions behind node linkage and contributions of node content to the interactive heterogeneities. Here, we propose a multi-aspect interaction and influence-unified evolutionary coupled system (MAI-ECS) for network representation by involving node content and linkage-based network structure. MAI-ECS jointly and iteratively learns two systems: a multi-aspect interaction learning system to capture heterogeneous hidden interactions between nodes and an influence propagation system to capture multiaspect node influences and their propagation between nodes. MAI-ECS couples, unifies and optimizes the two systems toward an effective representation of explicit node content and network structure, and implicit node interactions and influences. MAI-ECS shows superior performance in node classification and link prediction in comparison with the stateof-the-art methods on two real-world datasets. Further, we demonstrate the semantic interpretability of the results generated by MAI-ECS.

The representation of categorical data with hierarchical value coupling relationships (i.e., various value-to-value cluster interactions) is very critical yet challenging for capturing complex data characteristics in learning tasks. This paper proposes a novel and flexible coupled unsupervised categorical data representation (CURE) framework, which not only captures the hierarchical couplings but is also flexible enough to be instantiated for contrastive learning tasks. CURE first learns the value clusters of different granularities based on multiple value coupling functions and then learns the value representation from the couplings between the obtained value clusters. With two complementary value coupling functions, CURE is instantiated into two models: coupled data embedding (CDE) for clustering and coupled outlier scoring of high-dimensional data (COSH) for outlier detection. These show that CURE is flexible for value clustering and coupling learning between value clusters for different learning tasks. CDE embeds categorical data into a new space in which features are independent and semantics are rich. COSH represents data w.r.t. an outlying vector to capture complex outlying behaviors of objects in high-dimensional data. Substantial experiments show that CDE significantly outperforms three popular unsupervised encoding methods and three state-of-the-art similarity measures, and COSH performs significantly better than five state-of-the-art outlier detection methods on high-dimensional data. CDE and COSH are scalable and stable, linear to data size and quadratic to the number of features, and are insensitive to their parameters.

Mixed data with both categorical and continuous features are ubiquitous in real-world applications. Learning a good representation of mixed data is critical yet challenging for further learning tasks. Existing methods for representing mixed data often overlook the heterogeneous coupling relationships between categorical and continuous features as well as the discrimination between objects. To address these issues, we propose an auto-instructive representation learning scheme to enable margin-enhanced distance metric learning for a discrimination-enhanced representation. Accordingly, we design a metric-based auto-instructor (MAI) model which consists of two collaborative instructors. Each instructor captures the feature-level couplings in mixed data with fully connected networks, and guides the infinite-margin metric learning for the peer instructor with a contrastive order. By feeding the learned representation into both partition-based and density-based clustering methods, our experiments on eight UCI datasets show highly significant learning performance improvement and much more distinguishable visualization outcomes over the baseline methods.

New contents like blogs and online videos are produced in every second in the new media age. We argue that attraction is one of the decisive factors for user selection of new contents. However, collaborative filtering cannot work without user feedback; and the existing content-based recommender systems are ineligible to capture and interpret the attractive points on new contents. Accordingly, we propose attraction modeling to learn and interpret user attractiveness. Specially, we build a multilevel attraction model (MLAM) over the content features -- the story (textual data) and cast members (categorical data) of movies. In particular, we design multilevel personal filters to calculate users' attractiveness on words, sentences and cast members at different levels. The experimental results show the superiority of MLAM over the state-of-the-art methods. In addition, a case study is provided to demonstrate the interpretability of MLAM by visualizing user attractiveness on a movie.

Appropriate similarity measures always play a critical role in data analytics, learning, and processing. Measuring the intrinsic similarity of categorical data for unsupervised learning has not been substantially addressed, and even less effort has been made for the similarity analysis of categorical data that is not independent and identically distributed (non-IID). In this work, a Coupled Metric Similarity (CMS) is defined for unsupervised learning which flexibly captures the value-to-attribute-to-object heterogeneous coupling relationships. CMS learns the similarities in terms of intrinsic heterogeneous intra- and inter-attribute couplings and attribute-to-object couplings in categorical data. The CMS validity is guaranteed by satisfying metric properties and conditions, and CMS can flexibly adapt to IID to non-IID data. CMS is incorporated into spectral clustering and k-modes clustering and compared with relevant state-of-the-art similarity measures that are not necessarily metrics. The experimental results and theoretical analysis show the CMS effectiveness of capturing independent and coupled data characteristics, which significantly outperforms other similarity measures on most datasets.

Learning the representation of categorical data with hierarchical value coupling relationships is very challenging but critical for the effective analysis and learning of such data. This paper proposes a novel coupled unsupervised categorical data representation (CURE) framework and its instantiation, i.e., a coupled data embedding (CDE) method, for representing categorical data by hierarchical value-to-value cluster coupling learning. Unlike existing embedding- and similarity-based representation methods which can capture only a part or none of these complex couplings, CDE explicitly incorporates the hierarchical couplings into its embedding representation. CDE first learns two complementary feature value couplings which are then used to cluster values with different granularities. It further models the couplings in value clusters within the same granularity and with different granularities to embed feature values into a new numerical space with independent dimensions. Substantial experiments show that CDE significantly outperforms three popular unsupervised embedding methods and three state-of-the-art similarity-based representation methods.