IBPL: Information Bottleneck-based Prompt Learning for graph out-of-distribution detection

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph out-of-distribution (OOD) detection is crucial for reliable graph learning systems when training and test data distributions differ.
• Current prompt-based graph methods are efficient but overlook overlapping features, hindering OOD detection performance.

Purpose of the Study:

• To address the limitations of existing graph prompt methods in handling overlapping features between in-distribution (ID) and OOD graphs.
• To enhance the accuracy and robustness of graph OOD detection by developing a novel prompt learning approach.

Main Methods:

• Proposed Information Bottleneck-based Prompt Learning (IBPL) with a novel graph prompt masking node features and graph structure.
• Utilized an information bottleneck (IB) objective with noise data augmentation to eliminate overlapping features.
• Maximized mutual information between prompt graph and category labels to extract ID features, mitigating negative impacts of perturbed graphs.

Main Results:

• IBPL demonstrated superior performance in graph OOD detection across multiple real-world datasets.
• The method proved effective in both supervised and unsupervised learning scenarios.
• Empirical results and analyses confirmed the effectiveness of IBPL over competitive baselines.

Conclusions:

• IBPL offers a significant advancement in graph OOD detection by effectively handling overlapping features.
• The proposed approach enhances the reliability and safety of graph learning systems.
• IBPL provides a computationally efficient and effective solution for detecting OOD graphs.