graph patch extraction

This work addresses the challenges of feature heterogeneity and cross-domain transfer in graph data caused by the absence of textual information. It proposes “learnable graphlets” as the minimal semantic units of graphs, enabling for the first time a text-free cross-domain graph pre-training framework. By designing graphlet decomposition, a graphlet encoder, and an aggregator, the approach constructs a domain-agnostic architecture that extracts transferable knowledge from multi-domain graph data. The method supports joint pre-training across multiple domains and consistently achieves significant performance gains on diverse downstream tasks and datasets. Moreover, its effectiveness scales with the volume of pre-training data, and it reveals intrinsic connections between graphlet representations, existing graph models, and the transferability of node embeddings.

Existing visual graph recognition methods are often confined to specific tasks and lack generalizability and cross-scenario transferability. This work proposes GraSP, an end-to-end framework based on subgraph prediction that jointly models graph structure and visual features to enable unified recognition of diverse graph types and rendering styles. GraSP achieves cross-task transfer without task-specific fine-tuning, representing the first general-purpose and transferable approach for visual graph recognition. Evaluated on multiple synthetic benchmarks and a real-world application, GraSP demonstrates exceptional generalization and adaptability, advancing the field toward a unified paradigm for graph recognition.

While graph prompting has demonstrated strong empirical performance in graph data manipulation, its effectiveness has long lacked rigorous theoretical foundations. Method: This paper establishes the first theoretical framework for graph prompting from a data manipulation perspective, proving that graph prompting can approximate any graph transformation operator—thereby bridging pretraining and downstream tasks. We propose the first graph prompting approximation theorem, derive error upper bounds for both single-graph and batched-graph settings, and extend the analysis from linear models (e.g., GCN) to nonlinear models (e.g., GAT). Contribution/Results: The theoretical results are verifiable and generalizable; extensive experiments across multiple graph learning tasks empirically validate the reliability of prompting operations and confirm consistency with our theoretical predictions.

Graph Neural Networks (GNNs) suffer from limited expressive power due to their reliance on local, pairwise message passing, hindering effective modeling of higher-order subgraph structures. Existing random-walk-based kernel methods are designed for graph-level tasks, exhibit poor generalization, and employ fixed kernel configurations, lacking flexibility in structural modeling. Method: We propose Mixture of Subgraph Experts (MoSE), a novel framework that extracts informative subgraphs via anonymous walks and employs a gated routing mechanism to dynamically assign them to semantically specialized subgraph experts, enabling flexible and interpretable higher-order structural modeling. Contribution/Results: MoSE is the first to integrate subgraph expert ensembling with dynamic routing into multi-task graph learning. It theoretically surpasses the Subgraph Weisfeiler-Lehman (SWL) test in expressive power. Empirically, MoSE achieves significant improvements over state-of-the-art methods on node and graph classification tasks, while providing strong structural interpretability.

This work addresses the scalability bottleneck in subgraph pattern detection within large-scale graphs, a challenge rooted in the NP-completeness of the problem, by introducing the DETR paradigm to this task for the first time. The proposed method formulates subgraph detection as a set prediction problem, leveraging a graph neural network to encode the target graph, learnable query embeddings, and a Transformer decoder to jointly predict all pattern instances in an end-to-end manner via bipartite matching loss. This framework supports both exact and approximate pattern matching, thereby overcoming the limitation of traditional approaches that are restricted to exact structural matches. Experiments demonstrate that the method efficiently detects diverse patterns of up to 50 nodes in graphs containing 1,000 nodes, achieving an AP₁₀₀ of 91.2 on functional group detection in the ChEMBL molecular dataset.

This work addresses the limitation in standard graph-based semi-supervised node classification, where the training objective disregards predictive information from unlabeled nodes and thus fails to fully exploit the complete graph structure inherent in transductive learning. The authors propose a model-agnostic loss function enhancement that decomposes cross-entropy to incorporate prediction confidence from unlabeled nodes as an auxiliary learning signal. Specifically, the method minimizes the entropy of predictions on unlabeled nodes while preserving the supervised signal from labeled nodes through a balanced optimization objective. This approach seamlessly integrates with existing graph neural network (GNN) architectures without requiring architectural modifications. Extensive experiments across multiple benchmark datasets demonstrate consistent performance improvements, confirming the method’s effectiveness and broad applicability.

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.

Graph neural networks (GNNs) often suffer from semantic information loss during pooling operations in graph classification, which hinders their ability to provide interpretability at both subgraph and graph levels. To address this limitation, this work proposes the Subgraph Concept Network (SCN), which employs soft clustering of node concept embeddings to jointly and end-to-end distill semantic concepts at both subgraph and graph granularities. SCN is the first method to enable collaborative learning of multi-level concepts within GNNs, thereby overcoming the conventional reliance on node embeddings alone for interpretation. The approach achieves competitive graph classification performance while significantly enhancing model interpretability through explicit, hierarchical concept discovery.

This work addresses the subgraph alignment problem, which arises in domains such as computer vision, social network analysis, and bioinformatics, and entails locating an embedded small graph pattern within a larger graph and recovering the corresponding vertex mapping. The paper presents the first formal definition of this problem, establishes a theoretical framework based on the Erdős–Rényi subgraph-pair model, and derives near-tight information-theoretic limits for exact recovery using tools from information theory. These results offer a statistical perspective on the recoverability of the NP-hard subgraph isomorphism problem and characterize the fundamental information-theoretic limits of subgraph alignment, thereby providing a rigorous theoretical foundation and performance benchmark for future algorithmic development.