Technical Summary: LG-Plug for OOD Detection on Text-Attributed Graphs

Problem Definition

The paper addresses the challenge of Out-of-Distribution (OOD) detection on Text-Attributed Graphs (TAGs). TAGs consist of nodes associated with both graph structure (edges) and textual descriptions. While Graph Neural Networks (GNNs) effectively model in-distribution (ID) data by jointly learning semantic and structural information, they often fail on OOD nodes—those with unseen textual patterns or structural configurations—producing overconfident but incorrect predictions.

Existing solutions face two primary limitations:

1. Topology-driven methods: These rely on structural neighbors to detect OOD nodes but typically treat node texts as shallow features (e.g., bag-of-words), failing to leverage deep semantic knowledge.
2. LLM-based methods: Recent approaches use Large Language Models (LLMs) to synthesize pseudo OOD priors. However, they suffer from a reliability-informativeness imbalance:
   • Reliability-oriented methods (prompting with ID labels) generate OOD samples that are safe from ID overlap but often semantically unrealistic.
   • Informativeness-oriented methods (sampling unlabeled nodes) generate realistic semantics but risk mislabeling true ID nodes as OOD, introducing noise.
   • Furthermore, these methods often require specialized, end-to-end architectures, limiting their compatibility with established topology-driven detectors.

The authors pose two research questions: How to generate OOD priors that are both reliable and informative? How to utilize these priors while remaining compatible with existing topology-driven detectors?

Methodology: LG-Plug

The authors propose LG-Plug (LLM-Guided Plug-and-play), a framework designed to augment existing topology-driven OOD detectors with LLM-derived semantic priors without modifying their underlying architectures. The framework consists of three main stages:

1. Topology-Text Representation Alignment

To obtain fine-grained node embeddings that distinguish ID and OOD nodes, LG-Plug aligns graph and text representations:

• Encoders: A 2-layer GCN encodes the graph structure, while a Transformer-based encoder processes raw node texts.
• Alignment Objectives:
  • Node-level Alignment: Uses a symmetric contrastive loss to align the graph embedding ($Z$) and text embedding ($H$) for each node.
  • Edge-level Alignment: Minimizes the discrepancy between the topological similarity matrix (derived from $Z$) and the textual similarity matrix (derived from $H$) for connected nodes.
• Goal: This joint training produces embeddings where ID and OOD nodes are initially separable based on both semantic and structural cues.

2. Consensus-Driven OOD Exposure with LLM

Instead of querying the LLM for individual nodes (which is noisy and costly), LG-Plug generates OOD exposure at the cluster level:

• Clustering: Unlabeled nodes are clustered in the aligned embedding space.
• Representative Selection: For each cluster, a small set of nodes closest to the centroid is selected to ensure semantic coherence.
• Iterative Prompting & Codebook:
  • A lightweight category codebook is initialized with known ID classes.
  • The LLM is queried sequentially on batches of representative nodes within a cluster. The prompt asks the LLM to assign a node to an existing category or create a new OOD category.
  • The codebook is updated dynamically based on the LLM's assignments.
• Consensus Filtering: The process repeats until a batch reaches semantic consensus (all assigned to the same category) or a trial limit is reached.
• OOD Exposure Selection: The top-$K$ categories from the codebook are selected. If these categories contain no ID classes, the nodes assigned to them are added to the OOD exposure set ($V_{exp}$). This ensures the generated OOD priors are reliable (no ID overlap) and informative (based on cluster consensus).
• Efficiency: By using cluster sampling and a codebook, the method significantly reduces LLM query costs compared to exhaustive node-level prompting.

3. Score Regularization

The generated OOD exposure set ($V_{exp}$) is used as a regularization term to enhance existing detectors:

• A margin-based loss ($L_{reg}$) is applied to the detector's OOD scores.
• It enforces a separation margin: ID nodes should have low OOD scores, while nodes in $V_{exp}$ should have high OOD scores.
• This allows LG-Plug to be seamlessly integrated as a "plug-in" to any topology-driven detector (e.g., GNNSafe, GRASP) without architectural changes.

Key Contributions

1. Insights: The paper identifies the reliability-informativeness trade-off in current LLM-based OOD generation and the incompatibility of these methods with topology-driven approaches.
2. Novel Method (LG-Plug): A plug-and-play framework that uses topology-text alignment and consensus-driven LLM prompting to generate high-quality OOD priors.
3. Performance: The method achieves state-of-the-art results by improving existing detectors while maintaining efficiency.

Experimental Results

The authors evaluated LG-Plug on six TAG benchmarks (Cora, CiteSeer, PubMed, WikiCS, Books-History, ogbn-arxiv) using FPR95 (False Positive Rate at 95% True Positive Rate) and AUROC.

• Improvement over Topology-Driven Methods: Integrating LG-Plug with topology-driven baselines (GNNSafe, GRASP) consistently reduced FPR95 by $\ge$ 7% compared to the baselines alone.
• Comparison with LLM-based Methods: LG-Plug-integrated detectors outperformed standalone LLM-based OOD methods (LLMGuard, GLIP-OOD, GOE-LLM) by $\ge$ 5% in FPR95 reduction.
• Ablation Studies: Removing either the representation alignment or the intra-cluster filtering mechanism resulted in significant performance drops, confirming the necessity of both components.
• Efficiency: LG-Plug demonstrated superior efficiency, requiring significantly fewer LLM tokens and less time (70s on Cora vs. 295s–5400s for baselines) due to its cluster-based sampling strategy.
• Qualitative Analysis: Word cloud visualizations of the generated OOD exposure showed that LG-Plug captured distinct, high-frequency OOD semantics (e.g., "evolutionary," "theoretical") with minimal overlap with ID categories, whereas baselines exhibited semantic shifts or ID noise.

Significance and Claims

The paper claims that both topology and text matter for effective OOD detection on TAGs. It argues that while topology provides structural context, deep semantic understanding is crucial for distinguishing OOD nodes. LG-Plug bridges the gap between topology-driven and LLM-based approaches by:

• Leveraging LLMs to generate reliable and informative OOD priors through consensus-driven clustering, avoiding the pitfalls of individual node labeling.
• Maintaining compatibility with the ecosystem of topology-driven detectors, allowing them to benefit from semantic supervision without retraining the entire architecture.
• Offering a computationally efficient solution that decouples LLM query costs from graph size.

The authors conclude that LG-Plug represents a robust step forward in making GNNs safer for real-world deployment where distribution shifts are common, particularly in safety-critical domains like threat intelligence and biomedical research.