The rapid proliferation of location-based social media platforms has greatly accelerated the dissemination of geo-tagged information, but it has also facilitated the widespread propagation of localized rumors. Geo-textual Rumor Detection (GRD) has therefore become an important research topic in geoinformatics aimed at automatically identifying deceptive content tied to specific geographical contexts. However, most existing GRD methods rely on learning static patterns from offline datasets, which limits their ability to generalize to emergent local events characterized by rapidly evolving spatial-temporal information distributions. To better understand this limitation, we conduct preliminary analyses of model fitting behaviors during training and identify two critical issues: imbalanced fitting between real and fake classes, and low-rank feature representations caused by the model’s tendency to overfit to homogeneous real patterns. These phenomena directly lead to the severe loss of vital spatial information, which significantly constrains the model’s capacity to capture the diverse spatial and textual patterns inherent in localized rumors. To address these challenges, we propose a novel framework named Decomposing Orthogonal Spatial Subspaces for Emergent Geo-textual rumor detection (Doseg). Our approach decomposes model transformation matrices via singular value decomposition, explicitly separating linguistic semantic, geographical spatial information-aligned, and localized event-specific spatial components while enforcing orthogonality constraints to enhance spatial feature diversity. Extensive experiments on benchmark geo-textual datasets with strict spatial-temporal splits demonstrate that our method substantially improves detection performance and increases the number of dominant principal components in feature representations, leading to stronger generalization for emergent geo-textual rumor scenarios within the geospatial ecosystem.

Text steganography is a widely used method for covert communication, which primarily relies on natural language processing (NLP) techniques to embed hidden information using linguistic features within text. However, existing generative text encoding methods often face significant information capacity loss during the rejection process. In this paper, we propose a novel framework for text encoding: the Reject-Split Steganography Framework (RS-Stega). This framework leverages existing models to perform quality control on generated text, minimizing the impact on embedding capacity. The key innovation of our approach is its ability to adaptively select tokens based on context using the rejection mechanism, ensuring both fluency and coherence in the cover text, while optimizing embedding capacity with the split mechanism. Extensive comparative and ablation studies on three benchmark datasets demonstrate the effectiveness of our proposed framework. The results show that RS-Stega significantly improves both information capacity and cover text quality, achieving a 52.22 https://github.com/Poshang-Taoist/RS-Stega )

Retrieval-Augmented Generation (RAG) has emerged as a standard for knowledge-intensive tasks. However, synthesizing evidence across disjointed documents for multi-hop reasoning remains challenging. Existing entity-centric paradigms discard contextual nuances essential for disambiguation, causing severe information loss, noise sensitivity, and cascading error propagation. To address these limitations, we propose LECG, a Layered Evidence-Centric Graph framework designed to preserve semantic integrity through a multi-level hierarchical structure. LECG integrates structured reasoning with unstructured evidence via an Anchor-Expand strategy, which utilizes graph topology to prune irrelevant branches and employs a dynamic fallback mechanism to maintain reasoning continuity under uncertainty. We evaluated our framework on three standard multi-hop benchmarks. It achieves an average Exact Match rate of 49.39

Large language models (LLMs) exhibit strong abilities in multi-document reasoning, yet their evidence identification is highly sensitive to input order. We trace this limitation to attention mechanisms, where many heads overemphasize sequence boundaries and neglect central content. We systematically analyze attention distributions under document permutations and discover a small subset of heads that consistently prioritize task-relevant documents regardless of position. We formalize these as Expert Heads, identified via activation frequency and stability across permutations. Experiments on LLaMA, Mistral, and Qwen reveal architecture-specific patterns: mid-layer heads in LLaMA and Mistral dominate semantic integration, while deeper-layer heads in Qwen specialize in evidence selection. Moreover, Expert Heads exhibit concentrated focus during understanding and more distributed engagement during generation. Their activation strongly correlates with answer correctness, providing diagnostic signals for hallucination detection. Leveraging Expert Heads for document voting significantly improves retrieval and ranking on HotpotQA, 2WikiMultiHopQA, and MuSiQue, outperforming dense retrievers and LLM-based ranking with minimal overhead. Ablations confirm that even a small subset achieves robust gains. Our findings establish Expert Heads as a stable and interpretable mechanism for evidence integration, offering new directions for context pruning, hallucination mitigation, and head-guided training of LLMs

Commonsense Knowledge Graph Completion (CKGC) aims to infer missing facts from the known commonsense knowledge. However, existing approaches still face the following two challenges, i.e., graph sparsity and noisy facts. First, entities in CKGs are generally sparse, which causes a vast of low-degree nodes lack enough contextual information. Second, CKGs are generally constructed by crowdsourcing, so there exist low-quality facts which degrades inference accuracy. To address these issues, we propose a novel framework CSGD for commonsense KG completion. The framework dynamically generates synthetics edges via graph densification strategy for low-degree entities. Extract high quality subgraph for reasoning. Additionally, we propose an adaptive dynamic weighted sampling strategy to improve the training efficiency. Extensive experiments on three benchmark datasets demonstrate the effectiveness of our proposed framework. The codes are publicly available.

Recently, transformer has been widely used for sequential recommendation due to its superior sequence modeling and information sensing capabilities. Meanwhile, some studies capture high-order cooperative signals between sequences by graph structure. However, the general graph structure is not enough to capture nonlinear high-order cooperative signals and there are no detailed studies to balance the sequence-level information and the global graph-level higher-order information in sequential recommendation. To solve these challenges, we propose a model called Contrasting Transformer and Hypergraph Network for Cooperative Sequential Recommendation (THCSRec) to coordinate sequence-level information with global graph-level information. Specifically, our model uses a transformer network to capture the information of the sequence itself, and a hypergraph neural network to capture the global graph-level high-order information. Furthermore, the two networks cooperate through a contrastive learning task to maximize mutual information. Finally, the representations of the two networks are aggregated for prediction. In the experiments, we conducted extensive evaluation and ablation studies to verify the effectiveness of THCSRec ^1 on three real datasets, which exceeded the existing SOTA performance lines. ^1 (Our code is available on https://github.com/Elina-wu/THCSRec )

The Knowledge Graph logical reasoning task faces a dual challenge of insufficient semantic coverage from type information and missing structural information from relations. Although type annotations provide semantic priors for entities, their coarse-grained features cannot comprehensively characterize entity attributes; conversely, relational structure can enhance semantic representation, but the incompleteness of edges in real-world graphs limits modeling when relying on a single information source. To address these issues, we propose MRCLQR (Multi-information Relation Constraint-based Logical Query Reasoning), a framework with three core innovations: (1) an Information Semantic Alignment module based on contrastive learning, which achieves cross-modal semantic collaboration via entity-type-structure pairing; (2) a Constraint-aware Relation Encoding method that decomposes relation semantics into domain aggregation features, relation ontology semantics, and range constraint features; and (3) Neural-Symbolic Operators guided by domain constraints, which narrow the reasoning space through a constraint-aware attention mechanism. Experiments on FB15k, FB15k-237, and NELL-995 demonstrate that MRCLQR achieves average MRR scores of 35.8%, 16.2%, and 19.6%, respectively improving over the strongest baselines by 0.5%, 0.2%, and 0.2% - and exhibits an 8.0% average gain on complex queries involving negation. Ablation studies validate the effectiveness of multi-source collaboration and the curriculum learning strategy. This work offers a novel paradigm for heterogeneous knowledge fusion and logical query reasoning.

Next POI recommendation aids users in predicting their destinations of interest and plays an increasingly vital role in location-based social services. Recent works focus on analyzing both long-term and short-term interests in POI recommendation to gain a deeper understanding of user profiles. However, these methods for modeling long-term user’s sequences primarily rely on the Transformer model, which functions as a low-pass filter, often leading to the loss of high-frequency information. Additionally, long-term and short-term sequences are typically modeled independently, with short-term sequences often defined solely by the most recent check-ins, overlooking their interactions and dependencies. Therefore, we propose Enhancing Long-and Short-Term Representations for Next POI Recommendations via Frequency and Hierarchical Contrastive Learning (FHCRec). FHCRec captures both high-frequency and low-frequency information in long-term sequences to model richer long-term user’s preference representations. Moreover, it harnesses the characteristics of the short-term subsequences embedded within long-term sequences to enhance short-term preference characterization via local and global hierarchical contrastive learning, resulting in more personalized short-term preferences. The enhanced long-term and short-term preferences are integrated to improve model recommendation performance. Extensive experiments on three real-world datasets demonstrate the effectiveness of our method.

Logical query reasoning over knowledge graphs (KGs) is an important task for querying some information upon specified conditions. Despite recent advancements, existing methods typically focus on the inherent structure of logical queries and fail to capture the commonality among entities and relations, resulting in cascading errors during multi-hop inference. To mitigate this issue, we resort to inferring relations' domain constraints based on the commonality of their connected entities implicitly. Specifically, to capture the domain constraints of relations, we treat the set of relations emitted by an entity as its implicit concept information and derive a relation's domain constraint by aggregating the implicit concept information of its head entities. Employing a geometric-based embedding strategy, we enrich the representations of entities in the query with their implicit concept information. Additionally, we design a straightforward yet effective curriculum learning strategy to refine its reasoning skills. Notably, our model can be integrated into any existing query embedding-based logical query reasoning methods in a plug-and-play manner, enhancing their understanding of the entities as well as relations in queries. Experiments on three widely used datasets show that our model can achieve comparable outcomes and improve the performance of existing logical query reasoning models. Particularly, as a plug-in, it achieves an absolute improvement of the maximum 8.4% Hits@3 compared to the original model on the FB15k dataset, and it surpasses the former state-of-the-art plug-and- play logical query reasoning model in most scenes, exceeding it by up to 2.1% average Hits@3 results.

Continuous Integration (CI) is a cornerstone of modern collaborative software development, and numerous CI platforms are available. Differences in maintenance overhead, reliability, and integration depth with code-hosting platforms make migration between CI platforms a common practice. A central step in migration is translating CI configurations, which is challenging due to the intrinsic complexity of CI configurations and the need to understand semantic differences and relationships across CI platforms. With the advent of large language models (LLMs), recent advances in software engineering highlight their potential for CI configuration translation. In this paper, we present a study on LLM-based CI configuration translation, focusing on the migration from Travis CI to GitHub Actions. First, using 811 migration records, we quantify the effort involved and find that developers read an average of 38 lines of Travis configuration and write 58 lines of GitHub Actions configuration, with nearly half of the migrations requiring multiple commits. We further analyze translations produced by each of the four LLMs and identify 1,121 issues grouped into four categories: logic inconsistencies (38

Nowadays, misinformation articles, especially multimodal ones, are widely spread on social media platforms and cause serious negative effects. To control their propagation, Multimodal Misinformation Detection (MMD) becomes an active topic in the community to automatically identify misinformation. Previous MMD methods focus on supervising detectors by collecting offline data. However, in real-world scenarios, new events always continually emerge, making MMD models trained on offline data consistently outdated and ineffective. To address this issue, training MMD models under online data streams is an alternative, inducing an emerging task named continual MMD. Unfortunately, it is hindered by two major challenges. First, training on new data consistently decreases the detection performance on past data, named past knowledge forgetting. Second, the social environment constantly evolves over time, affecting the generalization on future data. To alleviate these challenges, we propose to remember past knowledge by isolating interference between event-specific parameters with a Dirichlet process-based mixture-of-expert structure, and anticipate future environmental distributions by learning a continuous-time dynamics model. Accordingly, we induce a new continual MMD method DAEDCMD. Extensive experiments demonstrate that DAEDCMD can consistently and significantly outperform the compared methods, including six MMD baselines and three continual learning methods.

Trajectory metric learning, which supports the trajectory similarity search, is one of the most fundamental tasks in spatial-temporal data analysis. However, existing trajectory metric learning methods rely on massive labels of pairwise trajectory distance, and thus cannot be applied to few-shot scenarios frequently occurring in real-world applications. Though performance drops caused by insufficient labels can be alleviated by knowledge distillation, we demonstrate that they cannot be directly applied to few-shot trajectory metric learning due to the domain shift problem. To this end, this paper proposes invariant and relaxed learning enhanced knowledge distillation method TMLKD for few-shot trajectory metric learning, such that domain-invariant representation and rank knowledge can be distilled. Specifically, in the representation learning phase, it first employs an adversarial sub-network to distinguish domain-specific and domain-invariant information, so as to distill transferable representation knowledge from teacher models. To mitigate the few-shot problem in student model training, we further enrich sparse labels of the target domain by utilizing the rank knowledge revealed in teachers' predictions. Particularly, TMLKD employs a list-wise learning-to-rank approach to learn the relaxed trajectory ranking orders instead of focusing on all the samples inefficiently. Finally, to guide accurate distillation, we adaptively assign reliability of teacher prediction by utilizing the ground-truth labels, to avoid misleading the student model with low-quality teacher predictions. Extensive experiments on three real-world datasets demonstrate the superiority of our model.

Positive and Unlabeled (PU) learning is a special case of binary classification with weak supervision, where only positive labeled and unlabeled data are available. Previous studies suggest several specific risk estimators of PU learning such as non-negative PU (nnPU), which are unbiased and consistent with the expected risk of supervised binary classification. In nnPU, the negative-class empirical risk is estimated by positive labeled and unlabeled data with a non-negativity constraint. However, its negative-class empirical risk estimator approaches 0, so the negative class is over-played, resulting in imbalanced error rates between positive and negative classes. To solve this problem, we suppose that the expected risks of the positive-class and negative-class should be close. Accordingly, we constrain that the negative-class empirical risk estimator is lower bounded by the positive-class empirical risk, instead of 0; and also incorporate an explicit equality constraint between them. we suggest a risk estimator of PU learning that balances positive and negative classification error rates, named $\mathrm{D{\small C-PU} }$, and suggest an efficient training method for $\mathrm{D{\small C-PU} }$ based on the augmented Lagrange multiplier framework. We theoretically analyze the estimation error of $\mathrm{D{\small C-PU} }$ and empirically validate that $\mathrm{D{\small C-PU} }$ achieves higher accuracy and converges more stable than other risk estimators of PU learning. Additionally, $\mathrm{D{\small C-PU} }$ also performs competitive accuracy performance with practical PU learning methods.

Fake news video detection is a challenge for social media and content platforms. There are two limitations in current method. (1) they only analyze video frames or video objects separately, which lacks information integration; and (2) the key role of semantic knowledge in recognition is ignored, and the video context information is not fully utilized. To solve these problems, we propose the Visual-Semantic Fake News Detection (VS-FND) framework, which aims to achieve deep relational inference of video content. Two types of graph memory modules are designed: (a) visual graph memory module, which focuses on mining key cues in video visual information. (b) Semantic graph memory module, which can use the semantic knowledge of videos to construct a rich semantic space and identify the semantic features related to fake news. Through the collaboration of these two modules, VS-FND build a hierarchical framework to enable visual-semantic relational reasoning from object level to frame level. We experiment on two benchmark datasets and achieve competitive performance compared to state-of-the-art methods, while also achieving significant advantages in the number of parameters and inference speed.

The rapid increase in the parameter counts of Large Language Models (LLMs), which often reach into the billions or even trillions, presents significant challenges for their practical deployment, particularly in resource-constrained environments. To address this issue, we propose PIP (Perturbation-based Iterative Pruning), a novel double-view structured pruning method to optimize LLMs, which combines information from two different views: the unperturbed view and the perturbed view. With the calculation of gradient differences, PIP iteratively prunes those that struggle to distinguish between these two views. Our experiments show that PIP reduces the parameter count by approximately 20% while retaining over 85% of the original model’s accuracy across varied benchmarks. In some cases, the performance of the pruned model is within 5% of the unpruned version, demonstrating PIP’s ability to preserve key aspects of model effectiveness. Moreover, PIP consistently outperforms existing state-of-the-art (SOTA) structured pruning methods, establishing it as a leading technique for optimizing LLMs in constrained environments.

Trajectory similarity search is one of the most fundamental tasks in spatial-temporal data analysis. Classical methods are based on predefined trajectory similarity measures, consuming high time and space costs. To accelerate similarity computation, some deep metric learning methods have recently been proposed to approximate predefined measures based on the learned representation of trajectories. However, instead of predefined measures, real applications may require personalized measures, which cannot be effectively learned by existing models due to insufficient labels. Thus, this paper proposes a transfer-learning-based model FTL-Traj, which addresses this problem by effectively transferring knowledge from several existing measures as source measures. Particularly, a ProbSparse self-attention-based GRU unit is designed to extract the spatial and structural information of each trajectory. Confronted with diverse source measures, the priority modeling assists the model for the rational ensemble. Then, sparse labels are enriched with rank knowledge and collaboration knowledge via transfer learning. Extensive experiments on two real-world datasets demonstrate the superiority of our model.

Author name disambiguation (AND) is a central task in academic search, which has received more attention recently accompanied by the increase of authors and academic publications. To tackle the AND problem, existing studies have proposed various approaches based on different types of information, such as raw document features (e.g., co-authors, titles, and keywords), the fusion feature (e.g., a hybrid publication embedding based on multiple raw document features), the local structural information (e.g., a publication’s neighborhood information on a graph), and the global structural information (e.g., interactive information between a node and others on a graph). However, there has been no work taking all the above-mentioned information into account and taking full advantage of the contributions of each raw document feature for the AND problem so far. To fill the gap, we propose a novel framework named EAND (Towards Effective Author Name Disambiguation by Hybrid Attention). Specifically, we design a novel feature extraction model, which consists of three hybrid attention mechanism layers, to extract key information from the global structural information and the local structural information that are generated from six similarity graphs constructed based on different similarity coefficients, raw document features, and the fusion feature. Each hybrid attention mechanism layer contains three key modules: a local structural perception, a global structural perception, and a feature extractor. Additionally, the mean absolute error function in the joint loss function is used to introduce the structural information loss of the vector space. Experimental results on two real-world datasets demonstrate that EAND achieves superior performance, outperforming state-of-the-art methods by at least +2.74% in terms of the micro-F1 score and +3.31% in terms of the macro-F1 score.

Entity alignment aims to find equivalent entities in two heterogeneous knowledge graphs(KGs), serving as a crucial process in the integration of KGs. As for educational KGs, different from traditional KGs, they exhibit a hierarchical structure, including one-to-one, one-to-many, and many-to-one inclusion relationships among entities. However, current research is incapable of handling hierarchical structure, resulting in entity redundancy within the educational KGs. In this study, we create two educational KGs focused on junior mathematics from different textbook versions, providing a valuable benchmark for this research community. Subsequently, we design a hierarchy-aware entity alignment approach for educational KGs across different versions. Our approach employs Graph Convolutional Networks to iteratively learn entity embeddings using seed alignment. Additionally, we harness domain-specific data such as questions, to identify potentially aligned entities. Finally, we consider alignment status within the parent and child nodes and employ a BERT binary classification model to detect inclusion relationships among candidate entity pairs, which is crucial for identifying inclusion relationships between non-independent equivalent entities and entities within the target KG. The effectiveness of the proposed approach is demonstrated through comprehensive experiments and analyses.

Question answering over temporal knowledge graphs (TKGQA) is an emerging topic, which has attracted increasing interest since it considers the dynamic knowledge in the world. Several datasets along with model developments are proposed in the TKGQA research field. However, existing studies generally focus on fact-centered reasoning, with limited attention to temporal reasoning. To tackle the intricate and comprehensive nature of temporal reasoning, we propose a new TKGQA dataset, MusTQ, which contains 666K multistep temporal reasoning questions as well as a TKG. The multi-step temporal reasoning is established based on six basic temporal reasoning types derived from a well-established measure theory. Using MusTQ, we evaluate previous TKGQA methods and find that they typically fall short in multi-step temporal reasoning. Furthermore, we propose a TKGQA model, MusTKGQA, which enhances multistep reasoning ability with entity-time attention mechanism and optimized temporal knowledge graph representation. Extensive experiments on MusTQ show that our model achieves stateof-the-art multi-step temporal reasoning performance.