Graphs provide a powerful basis for modeling Web-based relational data, with expressive GNNs to support the effective learning in dynamic web environments. However, real-world deployment is hindered by pervasive out-of-distribution (OOD) shifts, where evolving user activity and changing content semantics alter feature distributions and labeling criteria. These shifts often lead to unstable or overconfident predictions, undermining the trustworthiness required for Web4Good applications. Achieving reliable OOD generalization demands principled and interpretable uncertainty estimation; however, existing methods are largely post-hoc, insensitive to distribution shifts, and unable to explain where uncertainty arises especially in high-stakes settings. To address these limitations, we introduce SpIking GrapH predicTive coding (SIGHT), an uncertainty-aware plug-in graph learning module for reliable OOD Generalization. SIGHT performs iterative, error-driven correction over spiking graph states, enabling models to expose internal mismatch signals that reveal where predictions become unreliable. Across multiple graph benchmarks and diverse OOD scenarios, SIGHT consistently enhances predictive accuracy, uncertainty estimation, and interpretability when integrated with GNNs.

Knowledge graphs (KGs) effectively represent relations between real-world entities. However, many KGs suffer from a limitation in the diversity of relation types. When entities are linked with limited relation types, they tend to be highly similar. As a result, distinguishing between entities becomes particularly challenging, especially in larger KGs with numerous entities. Filling this gap, this paper proposes a novel representation model, Rich Relational Graph Convolution Network (R2GCN) for KGs. Rather than using the pair-wise relation in most general knowledge graphs, R2GCN fuses the high-order structural information and edge directions during representation learning. By this design, the types of relations are largely enriched, and similar entities can be distinguished. This can naturally alleviate the high similarity issue and thus overcome the representation difficulty. We evaluate R2GCN on two downstream tasks with both general KGs and relation-limited KGs. The consistently good performance of our R2GCN demonstrates its effectiveness on relation enrichment.

Despite notable advancements in prompting methods for Large Language Models (LLMs), such as Chain-of-Thought (CoT), existing strategies still suffer from excessive token usage and limited generalisability across diverse reasoning tasks. To address these limitations, we propose an Adaptive Causal Prompting with Sketch-of-Thought (ACPS) framework, which leverages structural causal models to infer the causal effect of a query on its answer and adaptively select an appropriate intervention (i.e., standard front-door and conditional front-door adjustments). This design enables generalisable causal reasoning across heterogeneous tasks without task-specific retraining. By replacing verbose CoT with concise Sketch-of-Thought, ACPS enables efficient reasoning that significantly reduces token usage and inference cost. Extensive experiments on multiple reasoning benchmarks and LLMs demonstrate that ACPS consistently outperforms existing prompting baselines in terms of accuracy, robustness, and computational efficiency.

Large Language Models (LLMs) are increasingly used for toxicity assessment in online moderation systems, where fairness across demographic groups is essential for equitable treatment. However, LLMs often produce inconsistent toxicity judgements for subtle expressions, particularly those involving implicit hate speech, revealing underlying biases that are difficult to correct through standard training. This raises a key question that existing approaches often overlook: when should corrective mechanisms be invoked to ensure fair and reliable assessments? To address this, we propose FairToT, an inference-time framework that enhances LLM fairness through prompt-guided toxicity assessment. FairToT identifies cases where demographic-related variation is likely to occur and determines when additional assessment should be applied. In addition, we introduce two interpretable fairness indicators that detect such cases and improve inference consistency without modifying model parameters. Experiments on benchmark datasets show that FairToT reduces group-level disparities while maintaining stable and reliable toxicity predictions, demonstrating that inference-time refinement offers an effective and practical approach for fairness improvement in LLM-based toxicity assessment systems. The source code can be found at https://aisuko.github.io/fair-tot/.

Medical image sharing in Internet of Medical Things (IoMT) environments faces risks of privacy leakage, unauthorized redistribution, and malicious manipulation, necessitating lossless copyright protection and trustworthy source traceability. Although zero-watermarking preserves the diagnostic integrity of medical images without modifying the original content, existing feature representations remain insufficiently robust for reliable watermark recovery under complex attacks. To address this issue, this paper proposes a hybrid zero-watermarking scheme based on a reparameterized efficient attention network (RepEA-Net). The proposed scheme employs a pre-trained RepEA-Net to extract robust multi-scale features from medical images, which are binarized by average hashing and combined with the encrypted watermark via an XOR operation to construct a zero-watermark. To enhance watermark security, a two-dimensional improved logistic-sine chaotic map (2D-ILSCM) is designed for watermark encryption. RepEA-Net reduces attack-induced feature shifts by strengthening local–global feature representation, thereby improving feature robustness and discriminability. Experiments are conducted on a Kaggle-derived self-built medical image dataset. The results demonstrate that the proposed scheme maintains average normalized correlation coefficient (NC) values above 0.85 under complex attacks and outperforms representative watermarking schemes in terms of robustness and stability, indicating its potential applicability to lossless copyright protection and trustworthy source traceability in IoMT-based medical image applications.

Graph transformers are a recent advancement in machine learning, offering a new class of neural network models for graph-structured data. The synergy between transformers and graph learning demonstrates strong performance and versatility across various graph-related tasks. This survey provides an in-depth review of recent progress and challenges in graph transformer research. We begin with foundational concepts of graphs and transformers. We then explore design perspectives of graph transformers, focusing on how they integrate graph inductive biases and graph attention mechanisms into the transformer architecture. Furthermore, we propose a taxonomy classifying graph transformers based on depth, scalability, and pre-training strategies, summarizing key principles for effective development of graph transformer models. Beyond technical analysis, we discuss the applications of graph transformer models for node-level, edge-level, and graph-level tasks, exploring their potential in other application scenarios as well. Finally, we identify remaining challenges in the field, such as scalability and efficiency, generalization and robustness, interpretability and explainability, dynamic and complex graphs, as well as data quality and diversity, charting future directions for graph transformer research.

To celebrate the 50th Anniversary of the Information Systems Journal, we are delighted to share our research reflections on the article "Community-diversified influence maximization in social networks" published at Information Systems in 2020. Our reflections will highlight the impact of this article on the authors' research trajectories, its influence on the broader research community, and its contributions to industry practice.

Epigenetic clocks based on DNA methylation have emerged as powerful tools for estimating biological age, with broad applications in aging research, age-related disease studies, and longevity science. Despite advances across machine learning approaches to epigenetic age prediction, spanning penalised linear regression, deep feedforward networks, residual architectures, and graph neural networks, no existing method jointly models co-methylation graph structure and site-specific DNA sequence context within a unified framework. We propose a unified sequence–graph integration framework for epigenetic age prediction that addresses this gap, integrating eight-dimensional DNA sequence statistical features through a lightweight gated modulation mechanism that adaptively scales each site's methylation signal according to its sequence-determined biological relevance prior to graph convolution. Evaluated on 3,707 blood methylation samples against a comprehensive set of baselines, our method achieves a test MAE of 3.149 years, a 12.8% improvement over the strongest graph-based baseline. Biologically informed statistical features outperform CNN-based sequence encoding, demonstrating that handcrafted sequence features are more effective than end-to-end learned representations in this data regime. Post-hoc interpretability analysis identifies CpG density and local adenine frequency as features with age-dependent importance shifts, consistent with known mechanisms of age-related hypermethylation at CpG-dense promoter regions. Our code is at https://github.com/yaoli2022/graphage-seq.

Private federated learning (FL) facilitates collaborative machine learning by leveraging cryptographic mechanisms to enhance the protection of local gradients. However, private FL faces two challenges in heterogeneous wireless networks with distributed devices. (1) Latency issue: participants are dynamic and have different characteristics or capabilities, and performance is determined by the node with the lowest computing and communication capabilities. (2) Data-owner dropout: participants may arbitrarily drop out of the training process, which degrades the training performance. Previous works address these issues using Shamir secret sharing or Lagrange-coded computing, which burdens the overhead of communication and computation. To address these challenges, we introduce a proxy-assisted secure aggregation protocol (PASA) for heterogeneous wireless networks. Particularly, a subset of distributed nodes is selected as proxy devices to conduct hierarchical aggregation. The hierarchical framework alleviates the computation and communication burden of participants, incurring only linear overheads. Beyond improving efficiency, the implementation of proxy devices also achieves dropout tolerance and provides privacy properties. Our experimental results validate that the aggregation overhead increases linearly and leads to faster training compared to previous studies. With a 10% dropout rate, our method achieves approximately 2.8 & times; speedup over SA (CCS'17) and 5 & times; over TJL (ACSAC'22).

Knowledge-intensive reasoning requires Large Language Models (LLMs) to ground answers in provided evidence. When evidence is insufficient, it is desirable that models abstain rather than confidently generating unsupported answers. Existing abstention methods rely on uncertainty estimation or evidence sufficiency checks, but neither tests whether the reasoning process for generation, driven by the interaction of provided evidence and the model's internal memory parameters, is actually grounded in the evidence. A key contributing factor is that entity mentions in context activate memorised associations, causing models to generate plausible responses ungrounded in evidence. We propose Twin Worlds (TW), a framework for improving reliability in knowledge-intensive reasoning through equivariance-based abstention: unlike invariance, which requires outputs to remain unchanged, equivariance requires outputs to transform correspondingly under entity substitutions. A model grounded in the evidence should produce answers that shift consistently when entities are substituted while their relations are preserved. TW constructs multiple worlds via typed substitutions of the original input that preserve relational structure while reducing parametric priors, and uses equivariance violations as an abstention signal. Across four benchmarks and three model backbones, TW identifies when answers are not reliably grounded in the provided evidence and outperforms uncertainty- and sufficiency-based baselines.

User preference modeling in service recommendation systems fundamentally relies on propagating signals over user-service interaction data to support effective service discovery and selection. However, the stochastic nature of user behavior introduces substantial noise into interaction data, which is further amplified during propagation and can degrade recommendation quality. Existing approaches either perform diffusion over interaction graphs based on the assumption that co-occurrence implies similarity, inevitably encoding and accumulating noise, or treat Knowledge Graphs (KGs) as auxiliary features without integrating them into the propagation process, thus failing to leverage structured service semantics for noise mitigation. To address these challenges, we propose DEEP, a Dual-diffusion framework for sErvice recommendation that incorporates External knowledge graPhs into the propagation process. Specifically, beyond collaborative diffusion over user-service interactions, DEEP introduces semantic diffusion grounded in service KGs to provide structured semantic guidance. Moreover, a semantic adaptive perception mechanism is designed to dynamically filter informative semantic signals during propagation, enabling more robust and semantics-aware service recommendation. Experimental results on three benchmark datasets demonstrate that DEEP consistently outperforms state-of-the-art methods, leading to significant improvements in recommendation performance.

Deep learning has been widely studied to develop computer-aided diagnosis methods for autism spectrum disorder (ASD) due to abnormalities in functional connectivity (FC) caused by neural activations. However, FC dysconnectivity occur in limited regions of the brain and the intricacy of large datasets for aggrandizing generalizability and variability are also responsible for an increase in heterogeneities. Thus, the crucial challenge for a classification framework is to identify and consolidate the distinctive features that can elevate the class separation. To resolve this predicament, we propose a Dual Attention Multi-Instance Learning Framework (DAMILF) for the diagnosis of ASD and its subtypes. The DAMILF comprises four modules: i) identification of specific regions in functional magnetic resonance imaging (fMRI), (ii) patch-level attention sub-network for extracting and strengthening patch-based localized discriminatory features, (iii) attention-aware feature map for balancing the corresponding share of each patch and generating a global feature space for the whole brain, and (iv) subject-level sub-network for supplementary whole brain global feature learning and ASD-related label prediction. Our proposed DAMILF was evaluated on large-scale baseline fMRI data collected from 408 ASD and 476 typical control subjects from the Autism Brain Imaging Data Exchange (ABIDE-1) repository, and exhibited superior diagnosis performance and a better generalization ability compared with several state-of-the-art methodologies.

Collusion attacks, where the server and malicious clients collaborate to bypass gradient source confirmation or tamper with aggregation results, cause a fundamental damage to the training process in federated learning (FL). However, existing verifiable FL frameworks typically adopt a split-verification model-clients can independently validate the correctness of aggregation results, while the server is responsible for confirming the legitimacy of gradient sources. Thus, the collusion attack has emerged as a critical and intractable vulnerability, as it completely collapses this split-verification model, thereby invalidating such verification mechanisms. To tackle this fundamental issue, we propose an innovative dual-verifiable FL framework. Specifically, by leveraging vector commitments, our scheme first integrates both gradient source confirmation and aggregation result verification into a unified framework. Based on this unified design, our scheme implements two targeted strategies to defend against collusion attacks. To prevent collusion-enabled gradient source spoofing, our scheme introduces a semi-trusted verification cluster in place of unreliable server-side validation and embeds an anonymized identity-check strategy to collaboratively confirm gradient source legitimacy. To counter collusion-driven manipulation of gradient aggregation results, our scheme customizes auxiliary verification proofs with a computational one-wayness for client-uploaded gradients. This renders it infeasible for adversaries to tamper with the aggregation result through reverse engineering. Under experiments and security analyses, our scheme achieves reliable dual-verification and robust resistance to collusion attacks. Moreover, it reduces computation and communication overhead by at least 40.83% and 50.47%, respectively, compared to state-of-the-art verifiable FL schemes.

In social recommenders, the inherent nonlinearity and opacity of synergistic effects across multiple social networks hinders users from understanding how diverse information is leveraged for recommendations, consequently diminishing explainability. However, existing explainers can only identify the topological information in social networks that significantly influences recommendations, failing to further explain the synergistic effects among this information. Inspired by existing findings that synergistic effects enhance mutual information between inputs and predictions to generate information gain, we extend this discovery to graph data. We quantify graph information gain to identify subgraphs embodying synergistic effects. Based on the theoretical insights, we propose SemExplainer, which explains synergistic effects by identifying subgraphs that embody them. SemExplainer first extracts explanatory subgraphs from multi-view social networks to generate preliminary importance explanations for recommendations. A conditional entropy optimization strategy to maximize information gain is developed, thereby further identifying subgraphs that embody synergistic effects from explanatory subgraphs. Finally, SemExplainer searches for paths from users to recommended items within the synergistic subgraphs to generate explanations for the recommendations. Extensive experiments on three datasets demonstrate the superiority of SemExplainer over baseline methods, providing superior explanations of synergistic effects.

Large Language Models (LLMs) often produce fluent but factually incorrect responses, a phenomenon known as hallucination. Abstention, where the model chooses not to answer and instead outputs phrases such as "I don't know", is a common safeguard. However, existing abstention methods typically rely on post-generation signals, such as generation variations or feedback, which limits their ability to prevent unreliable responses in advance. In this paper, we introduce Aspect-Based Causal Abstention (ABCA), a new framework that enables early abstention by analysing the internal diversity of LLM knowledge through causal inference. This diversity reflects the multifaceted nature of parametric knowledge acquired from various sources, representing diverse aspects such as disciplines, legal contexts, or temporal frames. ABCA estimates causal effects conditioned on these aspects to assess the reliability of knowledge relevant to a given query. Based on these estimates, we enable two types of abstention: Type-1, where aspect effects are inconsistent (knowledge conflict), and Type-2, where aspect effects consistently support abstention (knowledge insufficiency). Experiments on standard benchmarks demonstrate that ABCA improves abstention reliability, achieves state-of-the-art performance, and enhances the interpretability of abstention decisions.

Emotions are integral to communication, and understanding emotions is critically important for healthcare applications such as mental well-being and empathetic support. However, in conversational exchanges, emotions constantly shift as conversations unfold. Existing studies on reasoning about emotion shift triggers (ESR) require extensive manual annotation of both utterance-level emotion labels and emotion shift triggers. In this research, we propose a novel approach, the Wayfinding Instruction Tuning (WIT) framework, to tackle the ESR challenge in a more efficient manner with minimal dependency on emotion labels. WIT employs a large language model (LLM) to generate sequences of dynamic, context-sensitive wayfinding questions designed to guide the instruction tuning of another LLM on the ESR task. In contrast to conventional prompting, our key innovation is a structured wayfinding prompting strategy, which systematically decomposes the ESR process into carefully designed sub-questions. Together with their ground-truth-aligned answers, these sub-questions guide the model to learn how to detect changes in emotion and identify their triggers. Our experiments show that, even when fine-tuned on a medium-sized dataset of 100 training instances, using only pre- and post-shift emotion labels, WIT can directly predict emotion shifts without first identifying the emotion of each utterance. This design avoids intermediate emotion classification and its associated error propagation, a common issue in two-step pipelines. WIT achieves comparable performance (F1: 0.749 vs 0.760) to state-of-the-art supervised models while requiring only shift-boundary labels during training and zero annotation at inference, providing a practical alternative when complete utterance labels are unavailable.

Graphs are prevalent in numerous real-world applications. Previous methods directly model graph structures and achieve significant success. However, these methods encounter bottlenecks due to the inherent irregularity of graphs. An innovative solution is converting graphs into textual representations, thereby harnessing the powerful capabilities of Large Language Models (LLMs) to process and comprehend graphs. In this article, we present a comprehensive review of methodologies for applying LLMs to graphs, termed LLM4graph. The core of LLM4graph lies in transforming graphs into texts for LLMs to understand and analyze. Thus, we propose a novel taxonomy of LLM4graph methods from the view of the transformation. Specifically, existing methods can be divided into two paradigms: Graph2text and Graph2token, which transform graphs into texts or tokens as the input of LLMs, respectively. We point out four challenges during the transformation to systematically present existing methods from a problem-oriented perspective. For practical concerns, we provide a guideline for researchers on selecting appropriate models and LLMs for different graphs and hardware constraints. To empirically evaluate our taxonomy and different technical choices, we conduct experiments with representative methods in Graph2text and Graph2token. We also identify five future research directions for LLM4graph.