Selected work

Representative Papers

Compositional Feature Augmentation for Unbiased Scene Graph Generation

Aug 13, 2023IEEE International Conference on Computer Vision

Scene Graph Generation (SGG) suffers from severe predicate long-tail distribution, where conventional re-sampling–based debiasing methods fail to improve tail-predicate performance—primarily due to insufficient modeling of relational triplet feature diversity. This paper introduces, for the first time, a feature disentanglement perspective: it decomposes triplet representations into intrinsic (predicate semantics) and extrinsic (contextual dependency) components. Building upon this, we propose a plug-and-play replace-mix augmentation strategy that enhances tail-predicate feature diversity without modifying the backbone architecture. The method is model-agnostic and computationally efficient. Evaluated on Visual Genome (VG) and PIC benchmarks, it achieves state-of-the-art performance across multiple metrics, notably boosting tail-predicate Recall@100 by a significant margin. Moreover, it seamlessly integrates with diverse SGG frameworks, demonstrating broad compatibility and practical utility.

27 citations4 influentialRead paper

Handling Feature Heterogeneity with Learnable Graph Patches

Jun 16, 2026

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.

0 citationsRead paper

A Unified Framework for Context-Aware and Relation-Aware Graph Retrieval-Augmented Generation

Jun 16, 2026

Existing Graph Retrieval-Augmented Generation (Graph RAG) approaches struggle to effectively integrate contextual and relational information from both entities and text segments, thereby limiting their capacity to capture emergent knowledge. To address this limitation, this work proposes HyGRAG, a novel framework that unifies contextual and relational modeling through a hybrid hierarchical graph index. HyGRAG leverages iterative clustering and large language models to generate multi-granularity knowledge summaries and introduces a cross-level, context- and relation-aware retrieval mechanism. Notably, it is the first Graph RAG framework to achieve deep fusion of these two types of information and supports dynamic, attachment-based local re-summarization for efficient updates. Experimental results demonstrate that HyGRAG improves average accuracy by 9.7% on multi-hop reasoning tasks while maintaining strong computational efficiency.

0 citationsRead paper

PTCG-Bench: Can LLM Agents Master Pokémon Trading Card Game?

May 28, 2026

Existing benchmarks for large language model (LLM) agents struggle to comprehensively evaluate their capabilities in sustained decision-making and self-improvement within complex strategic interactions. This work proposes PTCG-Bench, the first evaluation framework based on the Pokémon Trading Card Game, which systematically assesses LLM agents along two dimensions: in-game strategic decision-making and experience-driven self-evolution. We introduce an interpretable, modular testing framework that effectively decouples agent architecture from underlying model capabilities and employ behavioral ablation analyses to identify key performance factors. Experimental results demonstrate that LLM agents can achieve non-trivial performance in this environment; however, achieving stable and continuous self-evolution remains challenging and is highly sensitive to the design of the evaluation framework.

0 citationsRead paper

How to use Graph Data in the Wild to Help Graph Anomaly Detection?

Jun 04, 2025Knowledge Discovery and Data Mining

Addressing key challenges in graph anomaly detection—including label scarcity, ambiguous anomaly definitions, and difficulty in modeling normal distribution—this paper proposes Wild-GAD, the first framework to enable cross-domain knowledge transfer using large-scale, heterogeneous “in-the-wild” graph data. Methodologically, it introduces (i) a unified graph database (UniWildGraph) and a shared feature space; (ii) an external graph selection criterion balancing representativeness and diversity; and (iii) an unsupervised transfer learning detection paradigm. Evaluated on six real-world datasets, Wild-GAD achieves average improvements of +18% in AUC-ROC and +32% in AUC-PR over state-of-the-art methods. This work establishes a scalable, annotation-free general enhancement paradigm for low-resource graph anomaly detection.

0 citationsRead paper

Recent publications

Latest Papers

Handling Feature Heterogeneity with Learnable Graph Patches

Jun 16, 2026

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.

0 citationsRead paper

A Unified Framework for Context-Aware and Relation-Aware Graph Retrieval-Augmented Generation

Jun 16, 2026

Existing Graph Retrieval-Augmented Generation (Graph RAG) approaches struggle to effectively integrate contextual and relational information from both entities and text segments, thereby limiting their capacity to capture emergent knowledge. To address this limitation, this work proposes HyGRAG, a novel framework that unifies contextual and relational modeling through a hybrid hierarchical graph index. HyGRAG leverages iterative clustering and large language models to generate multi-granularity knowledge summaries and introduces a cross-level, context- and relation-aware retrieval mechanism. Notably, it is the first Graph RAG framework to achieve deep fusion of these two types of information and supports dynamic, attachment-based local re-summarization for efficient updates. Experimental results demonstrate that HyGRAG improves average accuracy by 9.7% on multi-hop reasoning tasks while maintaining strong computational efficiency.

0 citationsRead paper

PTCG-Bench: Can LLM Agents Master Pokémon Trading Card Game?

May 28, 2026

Existing benchmarks for large language model (LLM) agents struggle to comprehensively evaluate their capabilities in sustained decision-making and self-improvement within complex strategic interactions. This work proposes PTCG-Bench, the first evaluation framework based on the Pokémon Trading Card Game, which systematically assesses LLM agents along two dimensions: in-game strategic decision-making and experience-driven self-evolution. We introduce an interpretable, modular testing framework that effectively decouples agent architecture from underlying model capabilities and employ behavioral ablation analyses to identify key performance factors. Experimental results demonstrate that LLM agents can achieve non-trivial performance in this environment; however, achieving stable and continuous self-evolution remains challenging and is highly sensitive to the design of the evaluation framework.

0 citationsRead paper

How to use Graph Data in the Wild to Help Graph Anomaly Detection?

Jun 04, 2025Knowledge Discovery and Data Mining

Addressing key challenges in graph anomaly detection—including label scarcity, ambiguous anomaly definitions, and difficulty in modeling normal distribution—this paper proposes Wild-GAD, the first framework to enable cross-domain knowledge transfer using large-scale, heterogeneous “in-the-wild” graph data. Methodologically, it introduces (i) a unified graph database (UniWildGraph) and a shared feature space; (ii) an external graph selection criterion balancing representativeness and diversity; and (iii) an unsupervised transfer learning detection paradigm. Evaluated on six real-world datasets, Wild-GAD achieves average improvements of +18% in AUC-ROC and +32% in AUC-PR over state-of-the-art methods. This work establishes a scalable, annotation-free general enhancement paradigm for low-resource graph anomaly detection.

0 citationsRead paper

KAA: Kolmogorov-Arnold Attention for Enhancing Attentive Graph Neural Networks

Jan 23, 2025

Existing graph attention mechanisms suffer from limited expressivity of their scoring functions, leading to inaccurate node importance estimation and constraining GNN performance. To address this, we propose the Kolmogorov–Arnold Attention (KAA) module—the first to incorporate Kolmogorov–Arnold Networks (KANs) into graph attention scoring—employing zero-order B-spline parameterization to achieve a favorable trade-off between high expressivity and low parameter count. We introduce the Maximum Ranking Distance (MRD) metric to quantify the upper bound of scoring error and theoretically prove that KAA possesses near-universal function approximation capability under parameter constraints. KAA is plug-and-play and compatible with diverse attentive GNN backbones. Extensive experiments demonstrate its superiority across multiple node-level and graph-level benchmarks, with improvements exceeding 20% in several cases, validating both its effectiveness and generalizability.

0 citationsRead paper