LLMGOOD

Code for Out-of-Distribution Detection via LLM-Guided Outlier Generation for Text-attributed Graph.

Rebuttal:

Embedding Visualization:

• ID: ID node embedding
• Our: Our generated high-quality OOD node
• Random: Randomly generated OOD node

Case Study:

• $\color{black}{Bold Black}$: Phrases showing semantic similarity between ID and OOD content
• $\color{red}{Red}$: Content exhibiting distinctive OOD characteristics

---

Run

# 1. Generate Hard OOD Nodes 
bash scripts/get_ood_nodes.sh

# 2. Obtain OOD-ID, OOD-OOD edge dataset
bash scripts/get_new_edges.sh

# 3. fintune local LLM for node distinction
bash scripts/train_lora.sh

# 4. train light-weight edge predictors and predict other potential edges
bash scripts/edge_train_and_predcit.sh

# 5. Combine OOD sample with ID graph to enhance OOD detection
cd gnnsafe
bash ood_detect.sh