generalized subgraph attribution

Existing higher-order subgraph attribution methods, such as GNN-LRP, suffer from exponential computational complexity, which hinders their scalability to deep graph neural networks. This work proposes an efficient algorithm grounded in the message-passing mechanism and the distributive property, achieving—for the first time—linear-time complexity for higher-order subgraph attribution, with computational overhead scaling linearly with network depth. The proposed method not only substantially accelerates the attribution process but also naturally incorporates neighborhood graph structural information, thereby establishing a generalized subgraph attribution framework. Experimental results demonstrate that the approach significantly enhances scalability and computational efficiency while preserving attribution fidelity.

This work addresses the challenge of effectively estimating the importance of subgraphs for graph-level tasks without access to ground-truth labels or reliance on the specific architecture of a graph neural network’s output layer. To this end, the authors formulate the problem as a linear Group Lasso regression in the embedding space and introduce structural priors—used here for the first time—to guide the solution. The proposed approach eliminates dependence on both task-specific labels and model architectures while naturally extending to the identification of critical nodes. Extensive experiments on multiple real-world graph datasets demonstrate that the method significantly outperforms existing baselines, confirming its effectiveness and strong generalization capability in assessing both subgraph and node importance.

This study investigates how graph topology influences the distribution of node attributes in attributed graphs. To this end, the authors propose an algebraic framework grounded in category theory to formally model how nodes perceive graph topology and to construct a probabilistic model of attribute distributions conditioned on topological structure. The model is theoretically sound, as it recovers the original attribute distribution in the limiting case of a complete graph. By integrating category theory, probabilistic modeling, and algebraic graph methods, the work achieves a unified representation of topology-aware attribute distributions. Empirical evaluation through unsupervised graph anomaly detection on a custom ID testing benchmark demonstrates that the proposed approach effectively captures the structural influence of topology on attribute distributions.

This work proposes Influence Propagation-based Graph Neural Network (IGNP), a novel framework for link prediction in multi-relational graphs. The approach uniquely models relational interactions between nodes as an influence diffusion process, leveraging an extended SIR epidemic model to capture global structural information within large-scale subgraphs. To enhance scalability, IGNP incorporates a virtual edge compression technique that substantially reduces computational complexity. By effectively integrating local node features with global topological patterns, the method achieves significant performance gains over strong existing baselines across multiple real-world datasets, thereby demonstrating the efficacy and superiority of modeling multi-relational link prediction through the lens of influence propagation.

To address the limited expressive power and high computational overhead of subgraph-based GNNs, this paper proposes HyMN—a unified framework that leverages random-walk-based centrality measures (e.g., PageRank, k-step centrality) for both subgraph sampling and structural encoding. Centrality scores guide the selection of informative subgraphs and serve as structural feature embeddings, thereby alleviating the expressivity bottlenecks of message-passing neural networks (MPNNs) and reducing redundant subgraph computations. Theoretical analysis via perturbation sensitivity demonstrates the robustness and effectiveness of centrality-guided sampling. HyMN adopts a lightweight hybrid architecture, achieving performance on par with full-subgraph GNNs and state-of-the-art models on both synthetic and real-world graph benchmarks, while significantly reducing runtime—thus balancing high discriminative power with computational efficiency.

This paper addresses feature redundancy in node representations within graph neural networks (GNNs) by proposing an adaptive, interpretable feature selection method. The approach dynamically evaluates node feature importance during training: it measures performance degradation on the validation set under interventional feature perturbations, theoretically models the coupling between GNN performance, node features, and graph structure, and progressively prunes irrelevant features based on their evolving correlation trajectories. Its core innovation lies in establishing a model- and task-agnostic online feature selection framework compatible with end-to-end joint optimization. Extensive experiments across multiple GNN architectures—including GCN, GAT, and GIN—on real-world graph datasets demonstrate that the method consistently improves classification accuracy (average gain of +1.2%), reduces input dimensionality (up to 40% reduction), and enhances model interpretability.

This work addresses the excessive computational and memory overhead in existing graph structure learning methods, often caused by redundant edges. To mitigate this issue, the study introduces diversity into graph structure learning for the first time, proposing a novel edge construction strategy that jointly leverages node similarity and diversity. The resulting graph structure is optimized under mutual information guidance, enabling the method to function as a plug-and-play module compatible with prevailing frameworks. This approach significantly reduces the number of edges while simultaneously enhancing model performance. Extensive experiments demonstrate consistent and substantial performance gains across six state-of-the-art graph structure learning methods, validating the effectiveness and generalizability of the proposed technique.

Existing graph representation learning methods generally lack interpretability and struggle to reveal community structures along with the rationale behind predictions. This work proposes GraphHull, the first approach to incorporate a bilevel convex hull geometric structure into graph generative models: global convex hull vertices serve as pure community prototypes, while local convex hulls capture intra-community variation, with a node’s position within these hulls directly explaining its linking behavior. By integrating convex hull representations, a determinantal point process prior, MAP estimation, and scalable subsampling, GraphHull ensures mutually exclusive local communities and a clear global structure while achieving multiscale interpretability. Experiments demonstrate that GraphHull effectively recovers hierarchical community structures in real-world networks and achieves state-of-the-art or competitive performance in link prediction and community detection, all while providing inherently interpretable predictions.

We propose SWING: Space Walks for Implicit Network Graphs, a new class of algorithms for computations involving Graph Random Features on graphs given by implicit representations (i-graphs), where edge-weights are defined as bi-variate functions of feature vectors in the corresponding nodes. Those classes of graphs include several prominent examples, such as: $\epsilon$-neighborhood graphs, used on regular basis in machine learning. Rather than conducting walks on graphs'nodes, those methods rely on walks in continuous spaces, in which those graphs are embedded. To accurately and efficiently approximate original combinatorial calculations, SWING applies customized Gumbel-softmax sampling mechanism with linearized kernels, obtained via random features coupled with importance sampling techniques. This algorithm is of its own interest. SWING relies on the deep connection between implicitly defined graphs and Fourier analysis, presented in this paper. SWING is accelerator-friendly and does not require input graph materialization. We provide detailed analysis of SWING and complement it with thorough experiments on different classes of i-graphs.

The absence of a unified theoretical framework for identifying core entities in higher-order interaction networks hinders systematic analysis of hypergraph centrality. Method: This paper systematically reviews 39 hypergraph centrality measures and proposes the first structured taxonomy—categorizing them into structural, functional, and contextual classes. Leveraging hypergraph modeling, network dynamical analysis, and empirical evaluation, we characterize systematic differences in similarity patterns and computational complexity across categories. We further construct a reproducible, comparable benchmark suite. Contribution/Results: Our work bridges dual gaps in the field: theoretical integration and empirical validation. The taxonomy provides a methodological guide and technical roadmap for hypergraph analysis, enabling principled, scalable advancement of higher-order network centrality research.