Disentangled Graph Prompting for Out-Of-Distribution Detection

DGP (Disentangled Graph Prompting) is a prompt-based framework for graph out-of-distribution (OOD) detection under the pre-training paradigm.

Instead of fine-tuning the entire GNN encoder, DGP freezes a pre-trained graph encoder and learns two complementary prompt graphs to capture fine-grained in-distribution patterns:

• Class-specific prompt graph focus on discriminative structures for classification
• Class-agnostic prompt graph capture shared patterns across ID graphs

By jointly modeling these two perspectives, DGP enhances the separation between ID and OOD samples and achieves strong performance with high efficiency.

• Paper: https://arxiv.org/abs/2603.29644
• Venue: TKDE

✨ Key Features

• 🔍 Fine-grained pattern modeling
  Disentangles ID patterns into class-specific and class-agnostic components for better OOD detection.

• 🧊 Pre-training + prompting paradigm
  Reuses pre-trained GNN encoders without full fine-tuning, reducing training cost.

• 🧠 Graph structure-aware prompting
  Learns edge-level reweighting to construct prompt graphs instead of modifying node features.

• ⚡ Efficient and scalable
  Avoids end-to-end retraining and achieves significant speedups over prior methods.

🧭 Why DGP

Most existing graph OOD methods rely on end-to-end training and implicitly learn ID patterns without explicit supervision.
DGP takes a different approach: it leverages pre-trained representations and explicitly models multiple views of ID structure, enabling more robust and interpretable OOD detection.

By combining prompt learning with graph representation learning, DGP provides a lightweight yet powerful alternative to traditional methods.

🔧 Requirements

We recommend using Python 3.9 with the following dependencies:

• torch-geometric==2.0.4
• torch-scatter==2.0.93
• torch-sparse==0.6.15
• numpy==1.21.2
• pandas==1.3.0
• python==3.9.15
• scikit-learn=1.0.2
• scipy==1.9.3
• torch==1.11.0
• torchvision==0.12.0

🚀 Training:

▶ Run DGP-GCL:

python DGP_GCL.py \
  --DS <dataset> \
  --model_type dgp-gcl \
  --lr <learning_rate> \
  --aug <augmentation_type> \
  --DS_pair <dataset_pair> \
  --lambda_ <lambda> \
  --gamma <gamma> \
  --alpha_1 <alpha1> \
  --alpha_2 <alpha2> \
  --dgp_lr <dgp_lr>

▶ Run DGP-Sim:

python DGP_Sim.py \
  --DS <dataset> \
  --model_type dgp-sim \
  --lr <learning_rate> \
  --eta <eta> \
  --DS_pair <dataset_pair> \
  --lambda_ <lambda> \
  --gamma <gamma> \
  --alpha_1 <alpha1> \
  --alpha_2 <alpha2> \
  --dgp_lr <dgp_lr>

🔁 Other Variants

To run pre-trained GNNs or their fine-tuned versions, simply modify the --model_type parameter.

🔍 Hyper-parameter Search

We also provide the code to search hyper-parameters, you can use the following command (for TOX21-SIDER dataset) to run it:

bash run_grid_search.sh

📄 Citation

@article{yang2026disentangled,
  title   = {Disentangled Graph Prompting for Out-of-Distribution Detection},
  author  = {Cheng Yang and Yu Hao and Qi Zhang and Chuan Shi},
  journal = {arXiv preprint arXiv:2603.29644},
  year    = {2026}
}