cs.AIMar 19, 2026

LLM-Enhanced Energy Contrastive Learning for Out-of-Distribution Detection in Text-Attributed Graphs

Authors: Xiaoxu Ma, Dong Li, Minglai Shao, Xintao Wu, Chen Zhao

Organizations: School of New Media and Communication, Tianjin University, China · Department of Computer Science, Baylor University, USA · Key Lab of Education Blockchain and Intelligent Technology, Ministry of Education, Guangxi Normal University, China · Electrical Engineering and Computer Science Department, University of Arkansas, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Text-attributed graphs, where nodes are enriched with textual attributes, have become a powerful tool for modeling real-world networks such as citation, social, and transaction networks. However, existing methods for learning from these graphs often assume that the distributions of training and testing data are consistent. This assumption leads to significant performance degradation when faced with out-of-distribution (OOD) data. In this paper, we address the challenge of node-level OOD detection in text-attributed graphs, with the goal of maintaining accurate node classification while simultaneously identifying OOD nodes. We propose a novel approach, LLM-Enhanced Energy Contrastive Learning for Out-of-Distribution Detection in Text-Attributed Graphs (LECT), which integrates large language models (LLMs) and energy-based contrastive learning. The proposed method involves generating high-quality OOD samples by leveraging the semantic understanding and contextual knowledge of LLMs to create dependency-aware pseudo-OOD nodes, and applying contrastive learning based on energy functions to distinguish between in-distribution (IND) and OOD nodes. The effectiveness of our method is demonstrated through extensive experiments on six benchmark datasets, where our method consistently outperforms state-of-the-art baselines, achieving both high classification accuracy and robust OOD detection capabilities.

Figures & tables

Figure 1: The overall pipeline of LECT. Given a text-attributed graph, we first construct pseudo-OOD nodes by injecting random edges and generating their textual content with an LLM. We then derive node representations using a text encoder and a GNN with a projector. Energy scores are subsequently computed to form Linked IND–OOD Pairs and Triplet Contrastive Pairs for training. Finally, the model identifies OOD samples based on the resulting energy levels.
 CoraCiteseerPubmed
 IND-Acc ↑\uparrowAUROC ↑\uparrowAUPR ↑\uparrowFPR95 ↓\downarrowIND-Acc ↑\uparrowAUROC ↑\uparrowAUPR ↑\uparrowFPR95 ↓\downarrowIND-Acc ↑\uparrowAUROC ↑\uparrowAUPR ↑\uparrowFPR95 ↓\downarrow
MSP86.72 ±\pm 0.4781.96 ±\pm 1.3843.37 ±\pm 1.2399.21 ±\pm 0.4375.36 ±\pm 0.7671.47 ±\pm 1.2930.81 ±\pm 0.5797.64 ±\pm 0.8992.68 ±\pm 0.1860.53 ±\pm 3.3848.67 ±\pm 3.7196.65 ±\pm 0.94
ODIN86.78 ±\pm 0.2180.76 ±\pm 1.9643.92 ±\pm 4.4862.74 ±\pm 11.6176.40 ±\pm 0.5070.60 ±\pm 2.2732.97 ±\pm 3.7879.77 ±\pm 3.0992.26 ±\pm 0.2959.86 ±\pm 3.0049.09 ±\pm 4.0792.39 ±\pm 1.32
OSSNC86.22 ±\pm 0.4182.32 ±\pm 1.7647.01 ±\pm 4.3497.51 ±\pm 0.6375.89 ±\pm 0.4072.14 ±\pm 1.5733.27 ±\pm 1.6796.23 ±\pm 2.4492.50 ±\pm 0.2861.14 ±\pm 2.3248.84 ±\pm 2.3897.22 ±\pm 0.36
EMP86.14 ±\pm 0.5882.12 ±\pm 2.4947.00 ±\pm 5.3060.91 ±\pm 4.5875.89 ±\pm 0.5170.16 ±\pm 2.0630.87 ±\pm 2.2779.26 ±\pm 2.4692.59 ±\pm 0.3161.44 ±\pm 4.2651.24 ±\pm 4.6392.90 ±\pm 1.86
OODGAT88.08 ±\pm 0.6580.31 ±\pm 1.4737.51 ±\pm 4.2549.09 ±\pm 5.8276.90 ±\pm 0.6568.43 ±\pm 1.2627.33 ±\pm 2.4997.72 ±\pm 2.0491.52 ±\pm 0.2670.08 ±\pm 1.3932.67 ±\pm 3.9996.55 ±\pm 1.78
Table 1: OOD detection performance in terms of Ind-Acc, AUROC, AUPR, and FPR95 (mean ±\pm std). Best results are bolded, second best are marked with a dash, and ↑\uparrow ( ↓\downarrow ) indicates that larger (smaller) values are better.
Figure 2: Ablation study results of LECT on the Cora , Citeseer , and Pubmed datasets, showing the performance without contrastive learning and without LLM-generated samples, respectively.
Figure 3: Ablation study results of LECT on the Cora , Citeseer , and Pubmed datasets, showing the performance without Lind-ood\mathcal{L}_{\text{ind-ood}} and without Ltriplet\mathcal{L}_{\text{triplet}} , respectively.
Figure 4: t-SNE visualization of node embeddings on the Cora for different baseline models and LECT.
Figure 5: t-SNE visualization and textual representation of the generated near-OOD and far-OOD samples on the Citeseer .
Figure 6: Heatmap of AUROC values for different sampling sizes of Linked IND-OOD pairs and Triplet Contrastive Pairs.
More in this topic

OOD Detection

Out-of-distribution (OOD) detection identifies inputs unlike a model’s training data so systems can flag uncertainty, reject inputs, or adapt safely.

See all 222 papers

Explore similar work

CardsList

  1. From Subtle to Significant: Prompt-Driven Self-Improving Optimization in Test-Time Graph OOD Detection

    Feb 19, 2026Luzhi Wang, Xuanshuo Fu, He Zhang +3Graph Prompt LearningOOD Detection

    College of Artificial Intelligence, Dalian Maritime University · Computer Vision Center, Universitat Aut`onoma de Barcelona · School of Computing Technologies, RMIT University +1

  2. Learning to Explore: Policy-Guided Outlier Synthesis for Graph Out-of-Distribution Detection

    Feb 28, 2026Li Sun, Lanxu Yang, Jiayu Tian +6OOD DetectionSynthetic Anomaly Generation

    North China Electric Power University · Beijing University of Posts and Telecommunications · Tsinghua University +4

  3. What Information Matters? Graph Out-of-Distribution Detection via Tri-Component Information Decomposition

    May 13, 2026Danny Wang, Ruihong Qiu, Zi HuangGraph Neural NetworksOOD Detection

    The University of Queensland, Australia.