Accurate localization of oriented objects in remote sensing images (RSI) faces complex challenges, particularly the difficulty of feature modeling arising from the geometric complexity of oriented targets, manifested as arbitrary orientations and scale variations. Existing paradigms, primarily based on Convolutional Neural Networks (CNNs) or Transformers, which struggle to capture robust rotation-invariant features due to inherent structural deficiencies (feature confusion due to fixed grid sampling in CNNs, spatial misalignment due to inadequate orientation encoding in attention mechanism). Graph-based feature modeling displays robustness against geometric complexity, prompting several studies to explore the incorporation of graph structures. However, these methods primarily focus on region proposals or pixel-level rotation correlations, struggling to effectively model the rotation-invariant features of targets. In this paper, we propose the Graph-based Semantic Rerouting Interaction (GSRI), shifting the modeling of rotated semantic representation for objects from tilted spatial-domain pixel recognition to the exploration of consistent channel-level semantic responses, thereby achieving robust rotation semantic feature modeling. Specifically, GSRI introduces a dynamic dilated KNN based graph structure to enhance the rotation-invariant feature modeling through capturing the consistent activation relationships of similar targets in high-dimensional feature space. By incorporating a cross-attention mechanism with dual positional encoding supervision, GSRI facilitates the alignment and propagation of robust rotational semantics across cross-scale feature maps, thereby mitigating background noise and bolstering the perceptibility of rotated objects in complex RSI scenarios. Extensive experiments on the DOTA-v1.0 and DIOR-R datasets demonstrate the superior effectiveness of our method.

The target of video moment retrieval (VMR) is predicting temporal spans within a video that semantically match a given linguistic query. Existing VMR methods based on multimodal large language models (MLLMs) overly rely on expensive high-quality datasets and time-consuming fine-tuning. Although some recent studies introduce a zero-shot setting to avoid fine-tuning, they overlook inherent language bias in the query, leading to erroneous localization. To tackle the aforementioned challenges, this paper proposes Moment-GPT, a tuning-free pipeline for zero-shot VMR utilizing frozen MLLMs. Specifically, we first employ LLaMA-3 to correct and rephrase the query to mitigate language bias. Subsequently, we design a span generator combined with MiniGPT-v2 to produce candidate spans adaptively. Finally, to leverage the video comprehension capabilities of MLLMs, we apply Video-ChatGPT and span scorer to select the most appropriate spans. Our proposed method substantially outperforms the state-of-the-art MLLM-based and zero-shot models on several public datasets, including QVHighlights, ActivityNet-Captions, and Charades-STA.

Relation reasoning in temporal knowledge graph infers unknown or emerging relational dependencies from historical structured data. Traditional approaches face inherent limitations in capturing complex semantic correlations and structural patterns among relations. To tackle this problem, we propose the Relational Multi-path Enhancement network (RME), which primarily focuses on relation modeling to enrich relation representations through comprehensive multi-path analysis. RME consists of five key components: (1) Controlled random walk module creates multi-hop head-to-tail paths using an adaptive stopping rule that balances short- and long-term connections. (2) Shared path extraction module identifies both shared-head paths and shared-tail paths. (3) Time-decayed path encoding module processes these paths differently. (4) Gated information aggregation module combines path information to determine which parts matter most. (5) Attention decoding module makes the final prediction by focusing on the most relevant path features. Experiments on multiple TKG benchmark datasets demonstrate that RME outperforms the state-of-the-art methods in relation multi-path reasoning.

Current large vision-language models (VLMs) exhibit remarkable performance in basic video understanding tasks. However, existing VLMs are still limited to surface-level perception and lack fine-grained spatio-temporal understanding and combinatorial reasoning capabilities. Existing methods typically rely on expensive human annotations or subtitle extraction, yet they struggle to effectively model temporal relations between frames. This paper proposes a structured representation based on temporal triplets to address two major challenges in traditional approaches: temporal fragmentation and entity reference ambiguity. By modeling objects, attributes, and relationships within the video and incorporating temporal information, we convert semantic content from keyframes into a sequence of temporal triplets. This structured representation is then used as input for zero-shot video question answering (VideoQA). Experiments were conducted on four benchmark VideoQA datasets: NExT-QA, STAR, MSVD-QA, and MSRVTT-QA, showing that our method achieves competitive performance without requiring fine-tuning, validating its generality and effectiveness.

Temporal knowledge graph (TKG) reasoning has become a hot topic due to its great value in many practical tasks. The key to TKG reasoning is modeling the structural information and evolutional patterns of the TKGs. While great efforts have been devoted to TKG reasoning, the structural and evolutional characteristics of real-world networks have not been considered. In the aspect of structure, real-world networks usually exhibit clear community structure and scale-free (long-tailed distribution) properties. In the aspect of evolution, the impact of an event decays with the time elapsing. In this paper, we propose a novel TKG reasoning model called Hawkes process-based Evolutional Representation Learning Network (HERLN), which learns structural information and evolutional patterns of a TKG simultaneously, considering the characteristics of real-world networks: community structure, scale-free and temporal decaying. First, we find communities in the input TKG to make the encoding get more similar intra-community embeddings. Second, we design a Hawkes process-based relational graph convolutional network to cope with the event impact-decaying phenomenon. Third, we design a conditional decoding method to alleviate biases towards frequent entities caused by long-tailed distribution. Experimental results show that HERLN achieves significant improvements over the state-of-the-art models.

Conditional semantic textual similarity (C-STS) assesses the similarity between pairs of sentence representations under different conditions. The current method encounters the over-estimation issue of positive and negative samples. Specifically, the similarity within positive samples is excessively high, while that within negative samples is excessively low. In this paper, we focus on the C-STS task and develop a conditional contrastive learning framework that constructs positive and negative samples from two perspectives, achieving the following primary objectives: (1) adaptive selection of the optimization direction for positive and negative samples to solve the over-estimation problem, (2) fully balance of the effects of hard and false negative samples. We validate the proposed method with five models based on bi-encoder and tri-encoder architectures, the results show that our proposed method achieves state-of-the-art performance. The code is available at https://github.com/qinzeyang0919/CCL.

Online continual learning (OCL) is a strict setting of continual learning (CL), where the OCL agent faces a never-ending data stream and encounters each new sample only once. An OCL agent suffers more serious catastrophic forgetting (i.e., forgetting previous knowledge of old classes) than a CL agent. Existing OCL methods leverage the contrastive-based losses to improve the classifier’s ability against forgetting. However, almost all these methods ignore the role of hard negative samples in these losses, and these samples are difficult to distinguish from positive samples, exacerbating catastrophic forgetting. In this paper, we focus on the classification task in the online contrastive continual learning (OCCL) setting and propose a novel contrastive-based OCL method named OCCL with Hard Negative Samples (OHNS) which emphasizes the importance of hard negative samples. Concretely, OHNS designs an adaptive weight to measure the hardness of the negative sample and proposes a new contrastive-based loss function by combining this loss function with the weight to enhance the classification strength on hard negative samples. We conduct extensive experiments on three real-world benchmark datasets, and the results demonstrate the superiority of OHNS over various state-of-the-art OCL methods.

Linear attention reduces the quadratic computational complexity of the Transformer’s Softmax attention to linearity. However, it restricts the emphasis on critical regions and diminishes feature diversity. To address these limitations, we propose the Selective Sampling Softmax and Multi-Feature Enhancement Linear (S3ML) Attention mechanism. First, we design a selective sampling Softmax attention to efficiently emphasize essential regions with constrained computational costs. Second, we propose a feature enhancing module, which integrates external channel attention, focused linear attention and token interaction to enrich feature diversity. Leveraging the S3ML attention mechanism, we construct a series of Vision Transformer models named S3ML-ViT. Classification experiments on ImageNet-1K and object detection tests on COCO2017 show that S3ML-ViT balances performance and efficiency effectively, highlighting its strong potential for downstream tasks. Our code is available at https://github.com/Senqi-Guan/Linear_Attention/tree/main/S3ML-ViT.

In deep reinforcement learning (DRL), the presence of dormant neurons leads to a significant reduction in network capacity, which results in sub-optimal performance and limited sample efficiency. Existing training techniques, especially those relying on periodic resetting (PR), exacerbate this issue. We propose the Full Network Capacity (FNC) framework based on PR, which consists of two novel modules: Dormant Neuron Reactivation (DNR) and Stable Policy Update (SPU). DNR continuously reactivates dormant neurons, thereby enhancing network capacity. SPU mitigates perturbation from DNR and PR and stabilizes the Q-values for the actor, ensuring smooth training and reliable policy updates. Our experimental evaluations on the Atari 100K and DMControl 100K benchmarks demonstrate the remarkable sample efficiency of FNC. On Atari 100K, FNC achieves a superhuman IQM HNS of 107.3%, outperforming the previous state-of-the-art method BBF by 13.3%. On DMControl 100K, FNC excels in 5 out of 6 tasks in terms of episodic return and attains the highest median and mean aggregated scores. FNC not only maximizes network capacity but also provides a practical solution for real-world applications where data collection is costly and time-consuming. Our implementation is publicly accessible at \url{https://github.com/tlyy/FNC}.

Video question answering involves understanding video content to generate accurate answers to questions. Recent studies have successfully modeled video features and achieved diverse multimodal interaction, yielding impressive outcomes. However, they have overlooked the fact that the video contains richer instances and events beyond the scope of the stated question. Extremely imbalanced alignment of information from both sides leads to significant instability in reasoning. To address this concern, we propose the Video-Context Aligned Transformer (V-CAT), which leverages the context to achieve semantic and content alignment between video and question. Specifically, the video and text are encoded into a shared semantic space initially. We apply contrastive learning to global video token and context token to enhance the semantic alignment. Then, the pooled context feature is utilized to obtain corresponding visual content. Finally, the answer is decoded by integrating the refined video and question features. We evaluate the effectiveness of V-CAT on MSVD-QA and MSRVTT-QA dataset, both achieving state-of-the-art performance. Extended experiments further analyze and demonstrate the effectiveness of each proposed module.

Reasoning over the Temporal Knowledge Graph (TKG) that predicts facts in the future has received much attention. Most previous works attempt to model temporal dynamics with knowledge graphs and graph convolution networks. However, these methods lack the consideration of high-order interactions between objects in TKG, which is an important factor to predict future facts. To address this problem, we introduce dynamic hypergraph embedding for temporal knowledge graph reasoning. Specifically, we obtain high-order interactions by constructing hypergraphs based on temporal knowledge graphs at different timestamps. Besides, we integrate the differences caused by time into the hypergraph representation in order to fit TKG. Then, we adapt dynamic meta-embedding for temporal hypergraph representation that allows our model to choose the appropriate high-order interactions for downstream reasoning. Experimental results on public TKG datasets show that our method outperforms the baselines. Furthermore, the analysis part demonstrates that the proposed method brings good interpretation for the predicted results.

Zero-shot node classification aims to predict nodes belonging to novel classes that have not been seen in the training. Existing studies focus on transferring knowledge from seen classes to unseen classes, which have achieved good performance in most cases. However, they do not fully leverage the relationships between nodes and overlook the issue of domain bias, affecting overall performance. In this paper, we propose a novel dependency capture and discriminative feature learning (DCDFL) model for zero-shot node classification. First, we address the issue of underutilized relationships between nodes by designing a relation-aware network to capture potentially long-range dependency, which can effectively utilize the available data information. Next, we introduce a domain-invariant adversarial loss to mitigate domain bias, encouraging the model to learn domain-insensitive feature representations. Furthermore, we refine and adapt the representations by leveraging inter-class separability within the metric space. Finally, extensive experiments on three well-known benchmark datasets demonstrate the superiority of our method over the strong baselines.

Graph neural networks have achieved tremendous success in various tasks over the past decade. However, recent studies have shown their vulnerabilities to well-designed adversarial attacks, in which even tiny perturbations can lead to the misclassification of the model. In this paper, we investigate the global evasion injection attack, aiming to degrade model performance on test nodes by injecting additional nodes into the graph. We propose the Graph Local Sparsity Attack(GLSA), where we introduce the concept of graph local sparsity, defining a node’s vulnerability by incorporating its neighborhood. Our method pre-computes scores to select vulnerable nodes using the sparsity and the gradual consistency condition. Subsequently, an iterative selection strategy is employed to inject the nodes. Finally, scores are dynamically updated, nodes are injected and features are finely generated to effectively degrade model performance. Experiments on several large-scale datasets demonstrate the effectiveness of our proposed method compared with the state-of-the-art approaches.

A continual learning agent, which faces a never-ending stream of data, suffers from severe catastrophic forgetting. To prevent forgetting, memory-based methods have shown more effective performance by retaining fractional previous data in a fixed-size memory buffer to maintain the observed category information. Nevertheless, which samples should be kept in the memory buffer is still an open question, and existing methods rarely address this issue from the perspective of sample information. In this work, we contribute a concise yet effective memory update method, Maximum Entropy Memory Update (MEMU). MEMU retains the samples adjacent to the decision boundaries since we observe that these samples have higher entropy. To this end, we design an indicator to score each sample and retain higher-score samples in the buffer. Compared to the state-of-the-art benchmarks, the experiments demonstrate that MEMU improves performance on five data streams with three metrics in the online continual learning setting.

This study focuses on the Badui region in eastern Tibet, an area with complex topography featuring numerous valleys, ravines, and frequent geological hazards. Given the economic expansion in this region, advanced techniques are essential for analyzing the distribution of geological hazards and developing early warnings of geological hazards. The research employs enhanced small baseline subset interferometric synthetic aperture radar (ESBAS-InSAR) technology, which provides more ascending and descending data than traditional small baseline subset-InSAR (SBAS-InSAR), allowing reprojection into vertical and horizontal components. Following dimensionality reduction through principal component analysis (PCA) and k-means clustering, the horizontal displacements were categorized into four clusters, and the vertical displacements were categorized into five clusters. Time-series data of vertical and horizontal displacements, rainfall, and normalized difference vegetation index (NDVI) were then used to assess 16 displacement prediction models. The temporal fusion transformer (TFT) model demonstrated the best predictive performance. To further improve accuracy, 11 static variables such as clusters, elevation, slope, aspect, distance from faults, time-varying known categorical variable, and earthquake times, were added as the TFT input variables. Results indicate that the optimized TFT model reduces the root-mean-square error (RMSE) from 3.4842 to 2.1707, the mean absolute percentage error (MAPE) from 2625.6399 to 2154.5505, and the mean absolute error (MAE) from 2.4392 to 2.3731. Overall, this study provides a framework for multivariate, multistep forecasting of diverse deformation modes across large areas and identifies distinct landslide deformation patterns through clustering, thereby enhancing the prediction of landslide deformation.

Multiview spectral clustering has received considerable attention in the past decades and still has great potential due to its unsupervised integration manner. It is well known that pairwise constraints boost the clustering process to a great extent. Nevertheless, the constraints are usually marked by human beings. To ameliorate the performance of multiview spectral clustering and alleviate the consumption of human resources, we propose self-supervised multiview spectral clustering with a small number of automatically retrieved pairwise constraints. First, the fused multiple autoencoders are used to extract the latent consistent feature of multiple views. Second, the pairwise constraints are achieved based on the commonality among multiple views. Then, the pairwise constraints are propagated through the neural network with historical memory. Finally, the propagated constraints are used to optimize the fused affinity matrix of spectral clustering. Our experiments on four benchmark datasets show the effectiveness of our proposed approach.

Early depression detection is essential to enable healthcare professionals to effectively intervene and treat the depressive conditions. In existing research, an increasing number of psychological manifestations of depression are incorporated, and demonstrated effective by integrating them into computational models, particularly the emotional cues of depression. Although emotional factors are effective in depression detection, few studies have paid attention to the underlying depressive causal factors hidden within social media text for depression detection. In this paper, we propose a novel partitioned filtering network model to extract emotional and causal features for predicting the depressive users on social media. Experimental results demonstrate the proposed model achieves superior performance over recent baseline models on the dataset by highlighting the emotional and causal factors.

In dynamic networks, links are annotated with timestamps showing the emerging time and the link prediction problem is to infer the future links in networks. Universal link prediction methods are highly demanded in various applications, which require universal link features that are feasible for multiple kinds of network topological structures and capable to address the difference of links with different timestamps. In this paper, we propose a novel link feature called Structure Subgraph Feature (SSF). The SSF is an outstanding link feature that is feasible to various dynamic networks due to the following superiorities: (1) the proposed structure subgraph is so far the most effective manner to represent surrounding topological features of target link and (2) the normalized influence well specifies the influence of multiple links and different timestamps in structure subgraph. We finally propose two link prediction methods by applying SSF to a linear regression model and a neural machine. Experimental results on real-world dynamic network datasets indicate that the SSF-based methods consistently provide top-class performance on various dynamic networks.

Many kinds of real world data can be modeled by a heterogeneous information network (HIN) which consists of multiple types of objects. Clustering plays an important role in mining knowledge from HIN. Several HIN clustering algorithms have been proposed in recent years. However, these algorithms suffer from one or moreof the following problems: (1) inability to model general HINs, (2) inability to simultaneously generate clusters for all types of objects, (3) inability to use similarity information of the objects with the same type. In this paper, we propose a powerful HIN clustering algorithm which can handle general HINs, simultaneously generate clusters for all types of objects, and use the similarity information of the same type of objects. First, we transform a general HIN into a meta-path-encoded relationship set. Second, we propose a nonnegative matrix tri-factorization multi-type co-clustering method, HMFClus, to cluster all types of objects in HIN simultaneously. Third, we integrate the information between the objects with the same type into HMFClus by using a similarity regularization. Extensive experiments on real world datasets show that the proposed algorithm outperforms the state-of-the-art methods.

Twitter has become a target platform for both promoters and spammers to disseminate their messages, which are more harmful than traditional spamming methods, such as email spamming. Recently, large amounts of campaigns that contain lots of spam or promotion accounts have emerged in Twitter. The campaigns cooperatively post unwanted information, and thus they can infect more normal users than individual spam or promotion accounts. Organizing or participating in campaigns has become the main technique to spread spam or promotion information in Twitter. Since traditional solutions focus on checking individual accounts or messages, efficient techniques for detecting spam and promotion campaigns in Twitter are urgently needed. In this article, we propose a framework to detect both spam and promotion campaigns. Our framework consists of three steps: the first step links accounts who post URLs for similar purposes; the second step extracts candidate campaigns that may be for spam or promotion purposes; and the third step classifies the candidate campaigns into normal, spam, and promotion groups. The key point of the framework is how to measure the similarity between accounts' purposes of posting URLs. We present two measure methods based on Shannon information theory: the first one uses the URLs posted by the users, and the second one considers both URLs and timestamps. Experimental results demonstrate that the proposed methods can extract the majority of the candidate campaigns correctly, and detect promotion and spam campaigns with high precision and recall.

Semi-automatic anti-spam algorithms propagate either trust through links from a good seed set (e.g., TrustRank) or distrust through inverse links from a bad seed set (e.g., Anti-TrustRank) to the entire Web. These kinds of algorithms have shown their powers in combating link-based Web spam since they integrate both human judgement and machine intelligence. Nevertheless, there is still much space for improvement. One issue of most existing trust/distust propagation algorithms is that only trust or distrust is propagated and only a good seed set or a bad seed set is used. According to Wu et al. [2006a], a combined usage of both trust and distrust propagation can lead to better results, and an effective framework is needed to realize this insight. Another more serious issue of existing algorithms is that trust or distrust is propagated in nondifferential ways, that is, a page propagates its trust or distrust score uniformly to its neighbors, without considering whether each neighbor should be trusted or distrusted. Such kinds of blind propagating schemes are inconsistent with the original intention of trust/distrust propagation. However, it seems impossible to implement differential propagation if only trust or distrust is propagated. In this article, we take the view that each Web page has both a trustworthy side and an untrustworthy side, and we thusly assign two scores to each Web page: T-Rank, scoring the trustworthiness of the page, and D-Rank, scoring the untrustworthiness of the page. We then propose an integrated framework that propagates both trust and distrust. In the framework, the propagation of T-Rank/D-Rank is penalized by the target's current D-Rank/T-Rank. In other words, the propagation of T-Rank/D-Rank is decided by the target's current (generalized) probability of being trustworthy/untrustworthy; thus a page propagates more trust/distrust to a trustworthy/untrustworthy neighbor than to an untrustworthy/trustworthy neighbor. In this way, propagating both trust and distrust with target differentiation is implemented. We use T-Rank scores to realize spam demotion and D-Rank scores to accomplish spam detection. The proposed Trust-DistrustRank (TDR) algorithm regresses to TrustRank and Anti-TrustRank when the penalty factor is set to 1 and 0, respectively. Thus TDR could be seen as a combinatorial generalization of both TrustRank and Anti-TrustRank. TDR not only makes full use of both trust and distrust propagation, but also overcomes the disadvantages of both TrustRank and Anti-TrustRank. Experimental results on benchmark datasets show that TDR outperforms other semi-automatic anti-spam algorithms for both spam demotion and spam detection tasks under various criteria.

The Twitter social network has become a target platform for both promoters and stammers to disseminate their target messages. There are a large number of campaigns containing coordinated spam or promoting accounts in Twitter, which are more harmful than the traditional methods, such as email spamming. Since traditional solutions mainly check individual accounts or messages, it is an urgent task to detect spam and promoting campaigns in Twitter. In this paper, we propose a scalable framework to detect both spam and promoting campaigns. Our framework consists of three steps: firstly linking accounts who post URLs for similar purposes, secondly extracting candidate campaigns which may exist for spam or promoting purpose and finally distinguishing their intents. One salient aspect of the framework is introducing a URL-driven estimation method to measure the similarity between accounts' purposes of posting URLs, the other one is proposing multiple features to distinguish the candidate campaigns based on a machine learning method. Over a large-scale dataset from Twitter, we can extract the actual campaigns with high precision and recall and distinguish the majority of the candidate campaigns correctly.

Propagating trust/distrust from a set of seed (good/bad) pages to the entire Web has been widely used to combat Web spam. It has been mentioned that a combined use of good and bad seeds can lead to better results. However, little work has been known to realize this insight successfully. A serious issue of existing algorithms is that trust/distrust is propagated in non-differential ways. However, it seems to be impossible to implement differential propagation if only trust or distrust is propagated. In this paper, we view that each Web page has both a trustworthy side and an untrustworthy side, and assign two scores to each Web page: T-Rank, scoring the trustworthiness, and D-Rank, scoring the untrustworthiness. We then propose an integrated framework which propagates both trust and distrust. In the framework, the propagation of T-Rank/D-Rank is penalized by the target's current D-Rank/T-Rank. In this way, propagating both trust and distrust with target differentiation is implemented. The proposed Trust-Distrust Rank (TDR) algorithm not only makes full use of both good seeds and bad seeds, but also overcomes the disadvantages of both existing trust propagation and distrust propagation algorithms. Experimental results show that TDR outperforms other typical anti-spam algorithms under various criteria.

Graph-based semi-supervised learning algorithms have attracted a lot of attention. Constructing a good graph is playing an essential role for all these algorithms. Many existing graph construction methods(e.g. Gaussian Kernel etc.) require user input parameter, which is hard to configure manually. In this paper, we propose a parameter-free similarity measure Adaptive Similarity Estimation (ASE), which constructs the graph by adaptively optimizing linear combination of its neighbors. Experimental results show the effectiveness of our proposed method.

In this paper, we propose two methods exploiting path information, direct-parent based method and full-path based method for syntax-based XML subtree matching in RDBs. In each proposed method, we discuss two ways of using the path information. The one is utilizing the path information after matching the leaf nodes. The other is using the path information together with the PCDATA value of leaf node as the join object. We perform experiments using the real bibliography XML documents stored in RDBs to evaluate the execution time, precision and recall of subtree matching. The experimental results indicate that both the two proposed path-based methods can effectively improve the precision and recall of subtree matching comparing with the original SLAX algorithm.

A number of XML labeling methods have been proposed to store XML documents in relational databases. However, they have a vulnerable point, in insertion operations. We propose the variable length endless insertable (VLEI) code and apply it to XML labeling to reduce the cost of insertion operations. Results of our experiments indicate that a combination of the VLEI code and Dewey order is effective for handling skewed insertions.