Spatial-temporal graph neural networks have demonstrated remarkable performance in time series forecasting. However, their inherent opacity poses a significant challenge, limiting their adoption in high-stakes applications. Existing post-hoc explainers are fundamentally limited, as they operate on static physical topologies and isolated time slices, failing to capture essential semantic relationships and long-range temporal dynamics. To address this, we propose FusionSHAP, a novel model-agnostic explanation framework. FusionSHAP makes two primary contributions. First, it introduces Semantic-Physical Graph Fusion, a process that constructs a hybrid graph by augmenting the conventional physical topology with a semantic graph generated by a Large Language Model. This fusion enables the discovery of non-obvious, functionally-driven relationships between nodes. Second, the framework employs Window-Level Attribution with Causal Lag Alignment, a method that adapts KernelSHAP to a node-timestep feature space. This provides temporally-precise explanations that pinpoint not only which nodes were influential but also when. Experiments on real-world traffic forecasting datasets demonstrate that FusionSHAP generates explanations with demonstrably higher fidelity and sparsity than existing methods, paving a new path toward uncovering the complex, non-local, and temporally-specific dependencies learned by STGNNs.
Clique is a fundamental model for cohesive subgraph discovery and has been extensively studied in single-layer graphs. However, in multilayer (ML) graphs, where vertices are connected through multiple types of relationships, existing studies remain limited and primarily focus on density-based models. In this paper, we propose two novel models, namely skyline ML-clique and optimal ML-clique. Specifically, given an ML graph G, a subgraph S is a skyline ML-clique if it is not dominated by any other clique in terms of both vertex set and supporting layers, while an optimal ML-clique is defined as the subgraph that maximizes a tunable function balancing clique size and layer support. We prove that the problem of enumerating all skyline ML-cliques is NP-hard, and develop a baseline enumeration algorithm along with several optimization techniques, including categorical pruning and adaptive skyline verification. For the optimal ML-clique, we further design a layered branch-and-bound framework with progressive computation to obtain high-quality solutions efficiently. Finally, extensive experiments on 7 real-world ML graphs demonstrate the effectiveness of the proposed algorithms.
Effective spatial keyword queries require joint modeling of location constraints and textual semantics. Existing systems typically rely either on spatial indexes with keyword filters or on semantic embeddings, which treat the two dimensions separately and thus lead to rigid spatial boundaries and limited semantic flexibility. To better understand these limitations, we systematically evaluate six representative approaches on large-scale OpenStreetMap (OSM) data from Greater Melbourne. The results show that non-embedding methods deliver exact matches but fail to generalise semantically, while embedding-based methods broaden coverage but capture spatial context poorly. This trade-off is not inevitable. We contribute two fusion strategies, namely concatenation embedding and contrastive fusion, that map spatial coordinates and textual descriptions into a shared vector space, and we integrate them with PostgreSQL, PostGIS and pgvector for evaluation alongside traditional baselines. Experiments demonstrate that fused embeddings improve semantic recall while preserving spatial fidelity, and they remain stable as query intent shifts between location and meaning. These findings indicate that unified spatial-semantic embeddings provide a practical direction for advancing spatial keyword queries.
Graph neural networks (GNNs) are an emerging tool for analysing graph-structured data, yet their training and inference require access to large and complex graphs. Major organisations such as Google and Facebook hold massive graph datasets (e.g., user interactions/relationships), but cannot share these directly due to their sensitive nature. We propose a secure framework, ColabGNN, for collaborative GNN training and inference that leverages Function Secret Sharing (FSS) to protect both the structure and features of private graphs. Our method ensures that raw graph data, training/inference computations and client queries remain hidden, while still enabling joint model training and secure inference across distributed parties.
This paper proposes a novel method for wheat disease image segmentation under limited labeled data. Rather than relying on traditional fine-tuning strategies, we address the challenge through visual parameter-efficient tuning, enabling pre-trained vision models to adapt effectively to the segmentation of wheat diseases with scarce data. Specifically, we freeze the backbone parameters of the segmentation network to mitigate the need for large-scale training samples. However, wheat disease symptoms exhibit distinct properties compared to natural images, such as randomly distributed lesions and interfering symptoms. To address these domain-specific characteristics, we design a disease-aware adapter that enhances the model’s ability to capture lesion-specific patterns and suppress irrelevant features via introducing class guidance, thereby improving disease segmentation accuracy under limited data conditions. Extensive experiments on wheat disease benchmarks demonstrate that the proposed parameter-efficient wheat disease segmenter achieves superior performance compared with state-of-the-art algorithms.
Reinforcement learning (RL)-based recommender systems have gained significant attention in recent years. However, the design of an effective reward function, which guides the optimization of the recommendation policy, is often challenging. Instead of relying solely on the reward function, exploring the causal factors underlying user behavior can be a promising approach to capturing dynamic user interests. Additionally, the limitations of simulation environments, such as data inefficiency, hinder the widespread application of existing methods in large-scale scenarios. Although some attempts have been made to convert offline datasets into simulators, the learning process becomes slower due to data inefficiency. Moreover, traditional RL algorithms lack the ability to learn directly from offline datasets, unlike supervised learning methods. In this paper, we propose a novel model called the Deep Distributional Offline Reinforcement Learning for Recommendation (DDRL4Rec). DDRL4Rec is an offline RL system that learns the distribution from datasets and does fine-turn during online interactions. In order to demonstrate the superiority of our model, we conducted extensive experiments on six real-world offline datasets and one online simulator.
Mobile crowdsourcing leverages mobile devices to enable effective truth discovery by recruiting workers to collect data. However, most existing incentive mechanisms overlook workers’ privacy. Although a few studies incorporate workers’ privacy, they often impose constraints on workers’ privacy demands, which may result in budget wastage. Moreover, none of the existing privacy-aware incentives account for the influence of malicious workers on truth discovery. This paper proposes Bid-TD, a personalized incentive mechanism for privacy-aware truth discovery. Bid-TD incorporates the influence of malicious workers to enhance the truth discovery accuracy and accurately captures workers’ privacy demands to provide fair compensation. Specifically, the platform publishes tasks to workers, who then submit bids reflecting their privacy demands. Taking into account each worker’s trust and privacy demands, the platform selects participants and designs the optimal contract that specifies the privacy budget and the corresponding apportionment ratio. We analyze and design an optimal contract function, which maximizes accuracy while satisfying budget feasibility and individual rationality. Experiments on both synthetic and real datasets demonstrate the effectiveness and superiority of Bid-TD.
Program synthesis from user intent remains a core challenge in databases and programming languages. Prior neural and symbolic NL2SQL systems can generate executable queries from natural language (NL) or input-output (IO) examples. However, they often struggle with robustness and generalizability in real-world scenarios. Addressing these limitations, we propose the first multimodal pipeline for SQL query synthesis that jointly leverages NL intent and multiple IO example pairs. Our framework fuses large language models with a counter example guided inductive synthesis (CEGIS) inspired, backtracking pipeline—spanning user intent interpretation, schema mapping, and iterative generate-validate-repair loops. Empirical results on TPC-H benchmarks show our method outperforms leading NL2SQL and Multi-Modal SQL synthesis baselines, correctly synthesizing 19/22 parameterized queries demonstrating a twofold improvement over SOTA. Our results show that combining natural language with input-output examples, along with iterative counterexample-driven repair, significantly improves reliability. This approach increases the automation potential of data-centric applications and supports robust query synthesis in complex, real-world scenarios.
Finding the maximum common induced subgraph between two graphs is a fundamental problem with diverse applications, ranging from chemistry to social network analysis. The McSplit algorithm, which is based on a partitioning-based branch-and-bound approach, has demonstrated remarkable efficiency in solving this problem by leveraging an upper bound to prune unproductive search paths. However, existing improvements to the McSplit algorithm focus primarily on upper bound refinements, neglecting the potential of strong lower bounds for early-stage pruning. In this work, we propose a two-stage framework that bridges exploratory heuristics with exact search to address this limitation. The first stage employs a fast heuristic to identify a strong initial lower bound, while the second stage uses this bound to guide an exact backtracking search, significantly improving pruning efficiency. Our dynamic transition mechanism between the stages ensures an effective balance between exploration and exhaustive search. Extensive experimental results demonstrate that the proposed method significantly enhances McSplit and its reinforcement learning-based extensions, achieving faster computation times and better solutions across various benchmarks.
Instruction tuning has revolutionized large language models (LLMs), yet their development remains bottlenecked by the need for extensive human-annotated data. While synthetic data generation methods like Self-Instruct attempt to address this limitation, they merely recycle knowledge already embedded in pre-trained models, failing to inject genuinely new information. We present Structure-to-Question (S2Q), a novel framework that transforms knowledge graphs (KGs) into high-quality instruction data without human annotation. S2Q leverages a key insight: the topological structure of KGs—from single edges to multi-hop paths—naturally corresponds to question complexity patterns, from simple factual queries to complex reasoning tasks. Our pipeline systematically samples subgraphs, converts structural patterns into natural language questions via LLM prompting, and ensures linguistic diversity through controlled generation, all while maintaining rigorous answer validation to prevent hallucination. Unlike existing approaches that use KGs solely for retrieval or verification, S2Q directly teaches structured knowledge to model parameters. Experiments demonstrate that Llama-3.1-8B models trained exclusively on S2Q-generated data achieve 78.5
In real-world scenarios, social networks play a vital role in daily communication and serve as powerful channels for rapid information dissemination. However, this high degree of connectivity also facilitates the spread of rumors, underscoring the urgent need for effective rumor control mechanisms. Most existing approaches fail to adequately account for two critical aspects of real-world rumor propagation: first, that rumors typically unfold over multiple rounds, and second, that users vary in priority within practical social contexts. To address this gap, we propose a new problem formulation termed Online Adaptive Rumor Blocking with Pertinence Set (OARBP), and introduce a corresponding solution framework–the Multi-round Hybrid Greedy Framework (MHGF). This method employs an adaptive and intelligent strategy to combat multi-round rumor propagation through three core stages: a local phase, a global phase, and a realization phase. The primary objective is to select a set of truth-seed nodes that disseminate truthful information, thereby maximizing the spread of truth while ensuring sufficient coverage within high-priority user groups. Experimental evaluations conducted on eight real-world social network datasets demonstrate the efficiency and effectiveness of the proposed framework. The results indicate that our approach not only facilitates rapid truth propagation but also successfully achieves targeted protection of key users, outperforming several existing methods.
Temporal graphs capture time-stamped interactions in domains such as finance and social networks. Prior work formalized historical k-core queries and the PHC index; constructing it reduces to computing, for every vertex and start time, its core time–the earliest end time at which the vertex enters the k-core of the windowed snapshot. The existing method computes start times serially and propagates neighbor updates, introducing dependencies within each round of computation, which hinder parallelization and scalability. We revisit the task through a parallel lens and propose a vertex-centric algorithm that computes core times for multiple start times in parallel, then updates them incrementally for subsequent start times, preserving exactness while mitigating these dependencies. Experiments on real temporal graphs show consistent speedups over a single-threaded baseline and a naive lock-based parallelization.
Topic modeling enables automated identification of latent thematic patterns within unstructured text corpora, facilitating efficient knowledge discovery. However, traditional cluster-based topic models often suffer from poor topic coherence due to the unsupervised nature of clustering algorithms and lack the capability to dynamically adjust clustering granularity. To address these limitations, we propose LLM-Guided Hierarchy-Aware Topic Modeling (LHATM), a hybrid framework that integrates hierarchical clustering with LLM feedback. LHATM begins with density-sensitive hierarchical clustering to build a stable topic tree, then leverages LLMs to refine the hierarchy through dynamic split-and-merge operations. In addition, ambiguous points arising from cluster adjustments are reallocated using a hybrid semantic-geometric strategy, ensuring both robustness and interpretability. Experiments on seven benchmark datasets show that our approach consistently produces more coherent and adaptable topics than existing methods.
In this paper, we study the maximum α -quasi-clique (MQC) problem, which seeks the largest subgraph with edge density at least α , where α∈ (0,1] is a user-specified parameter. Quasi-cliques generalize cliques, with 1-quasi-cliques corresponding exactly to cliques. The MQC problem is NP-hard, and existing heuristic algorithms such as , and suffer from two major limitations: (i) their solution quality is often suboptimal, and (ii) they must rerun costly algorithms to discovery alternative dense subgraphs for different α values. We propose an index-based approach for efficiently querying large quasi-cliques. Our index stores size-density pairs {(k,ρ _k) | k ∈ [2,n]} together with their corresponding subgraphs, where ρ _k is the edge density of a k-vertex subgraph obtained by our index construction algorithms. Given a query parameter α , we return the largest k with ρ _k ≥α along with its associated subgraph. We introduce two index construction algorithms, and , which run in 𝒪(m+n) and 𝒪(δ (G)· m) time, respectively, and produce indexes of size 𝒪(n) and 𝒪(m+n) . Both indexes satisfy the monotonicity property, enabling query processing in time linear in the result size. Extensive experiments on nine large real-world graphs demonstrate the efficiency and effectiveness of our algorithms.
Traditional spatial databases and map services handle deterministic queries well, but struggle to handle intent-rich and complex spatial queries. Even a simple request such as “In Melbourne, find a convenient petrol stop along the Tullamarine Freeway to the airport, avoiding tolls and minimizing detour” often breaks current systems, revealing a gap between natural-language intent and executable spatial operations. To fill this gap, we introduce an LLM-driven framework that parses user intent, decomposes complex spatial requests into executable sub-queries, schedules the necessary Google Maps API calls, and synthesizes the final answer. To enable systematic study, we curate a dataset of complex spatial queries with selected answers and provide a concrete implementation that compiles intents into routes and places calls while coordinating routing and POI services. We evaluate on the generated dataset with Google Maps and find that our system substantially outperforms a zero-shot LLM: it retrieves more feasible, on-route candidates and produces schema-valid, verifiable answers, while trading modest extra latency.
This paper tackles catastrophic forgetting in continual learning for multimodal knowledge graphs, where evolving structures and heterogeneous modalities challenge existing methods. We propose ARTD, a novel framework integrating adaptive memory replay and topology-aware knowledge distillation. Our approach dynamically balances new knowledge acquisition with historical knowledge preservation through: task-sensitive memory sampling guided by real-time forgetting metrics, multi-level distillation preserving structural relationships via contrastive alignment and parameter regularization, and cross-modal gating for unified entity representations. Comprehensive evaluation across academic and industrial benchmarks demonstrates superior performance over state-of-the-art baselines in link prediction tasks. Ablation studies confirm the complementary nature of replay and distillation mechanisms, while temporal analysis shows effective mitigation of forgetting across incremental learning phases. The framework’s robustness to domain-specific heterogeneity establishes its viability for real-world dynamic knowledge systems.
Executing Federated Learning (FL) efficiently within heterogeneous Computing Power Networks (CPNs) presents a critical challenge. Traditional FL methods, designed for homogeneous environments, utilize static, resource-blind scheduling. This leads to severe performance bottlenecks and prohibitive training latency when confronted with the multi-dimensional resource (CPU, memory, bandwidth) and data (Non-IID) heterogeneity inherent in CPNs, as system efficiency is dictated by the slowest clients or “stragglers.“ To address this, we propose FedDRL, a novel FL framework featuring a resource-aware intelligent scheduler based on Deep Reinforcement Learning (DRL). We model the complex scheduling task as a sequential decision-making process. A DRL agent on the control plane perceives the real-time global state of the CPN, including client-side computational capabilities and data characteristics, to learn a dynamic task assignment policy. This policy tailors the computational load for each client in every round, aiming to minimize long-term average training latency without sacrificing model performance. Extensive experiments on benchmark datasets, including CIFAR-10 and the more challenging CIFAR-100, demonstrate that FedDRL significantly outperforms baseline algorithms. Notably, under extreme data heterogeneity (α = 0.1) on CIFAR-10, our method reduces end-to-end training latency by an order of magnitude while achieving state-of-the-art accuracy, a performance advantage that is consistently validated on CIFAR-100. This work provides an adaptive, self-optimizing solution for deploying efficient Federated Learning in complex, real-world Computing Power Network environments.
The increasing number of new datasets and anomaly detection (AD) algorithms has highlighted the need for a scalable benchmarking framework to showcase algorithms’ performance portfolio and potentials of datasets in an interpretable way. Existing benchmarking methods typically report raw performance metrics without offering interpretable insights into the relationships between algorithms and datasets. In this work, we propose XEvalAD, a novel and explainable evaluation framework that applies Item Response Theory (IRT) to AD benchmarking. By jointly modelling the latent traits of algorithms (e.g., capability) and datasets (e.g., difficulty), XEvalAD enables performance ranking and reveals underlying factors that influence algorithm behaviour. The framework produces intuitive visualisations that facilitate deeper understanding of algorithm-dataset interactions, uncovering robustness patterns and dataset-specific challenges. We demonstrate the effectiveness of XEvalAD through extensive experiments on a large-scale benchmark suite covering diverse AD algorithms and datasets. Results show that our approach provides a more interpretable perspective on AD evaluation. The source code is publicly available at https://github.com/SabaFathi/XEvalAD .
Temporal reachability prediction aims to determine whether a target node can be reached from a source node within a given time interval in a temporal graph. This problem is challenging due to the need to jointly capture evolving structural patterns, temporal dependencies, and multi-hop propagation. In this work, we propose an enhanced temporal graph embedding framework tailored for temporal reachability prediction. Previously, there was work on reachability prediction using random walk incorporating continuous time through constraints on transition probabilities and path frequency bias on starting node selection to solve this problem. Our method ReaCH-TGN (Reachability prediction with Contrastive Hop-Time aware Temporal Graph Network) integrates three key optimizations: (1) Hop-aware contrastive loss, which extends the NT-Xent objective by incorporating hop-dependent weights to emphasize more challenging long-range dependencies; (2) Time-gap penalty, which regularizes embeddings based on the temporal distance between events, improving recency sensitivity; and (3) Temporal data augmentation, which combines random event dropout with timestamp jitter to enhance robustness against temporal noise. Experiments on three real-world temporal networks show that our approach achieves significant improvements in accuracy across multiple hop distances, particularly in long-range reachability scenarios.
The transferability of adversarial examples under the black-box setting piques the interest of deep learning practitioners, which also offers an effective direction to identify the deficiencies of DNNs, especially for safety-critical applications. Due to its wide range of expertise, ensemble adversarial attack shows promise in synthesizing more transferable adversarial samples while being under-explored, as existing ensemble methods simply sum the contributions of sub-models or just reduce their gradient variance. Therefore, this paper develops a novel Gradients Reweighing method (dubbed GREW) for ensemble attack to steer adversarial example generation through reweighing the importance of gradients from different sub-models. Specifically, the proposed GREW method can adaptively adjust the importance of the gradient from different sub-models in each iteration, thus helping the model escape from the local optimum. Comprehensive experimental results on several widely used datasets including CIFAR100 and ImageNet confirm the superiority and effectiveness of our transferability and stability when compared to the recent methods.
