GOODAT: Towards Test-time Graph Out-of-Distribution Detection Graph neural networks (GNNs) have been widely applied in various domains for modeling graph data. However, GNNs often fail to accurately predict samples from an unfamiliar distribution (out-of-distribution, OOD). To address this, the paper introduces GOODAT, a test-time graph OOD detection method that operates independently of training data and GNN architecture modifications. GOODAT uses a lightweight graph masker to learn informative subgraphs from test samples, enabling the capture of distinct graph patterns between OOD and in-distribution (ID) samples. Three unsupervised objective functions based on the graph information bottleneck principle are designed to optimize the graph masker, encouraging it to capture compact yet informative subgraphs for OOD detection. Comprehensive evaluations show that GOODAT outperforms state-of-the-art benchmarks across various real-world datasets. GOODAT is a data-centric, unsupervised, and plug-and-play solution that does not require training data or modifying the original GNN architecture. It leverages the information bottleneck principle to design three GIB-boosted losses: subgraph GIB loss, masked graph GIB loss, and graph distribution separating loss. These losses help in learning informative subgraphs and separating the distributions of ID and OOD graphs. The method is tested on multiple datasets and scenarios, demonstrating its effectiveness in detecting OOD samples. The results show that GOODAT achieves significant improvements in graph OOD detection tasks compared to baselines. The method is also effective in graph anomaly detection, showing superior performance over other test-time-oriented techniques. The paper concludes that GOODAT is a promising approach for test-time graph OOD detection, offering a lightweight, training data-independent, and plug-and-play solution for graph OOD detection.GOODAT: Towards Test-time Graph Out-of-Distribution Detection Graph neural networks (GNNs) have been widely applied in various domains for modeling graph data. However, GNNs often fail to accurately predict samples from an unfamiliar distribution (out-of-distribution, OOD). To address this, the paper introduces GOODAT, a test-time graph OOD detection method that operates independently of training data and GNN architecture modifications. GOODAT uses a lightweight graph masker to learn informative subgraphs from test samples, enabling the capture of distinct graph patterns between OOD and in-distribution (ID) samples. Three unsupervised objective functions based on the graph information bottleneck principle are designed to optimize the graph masker, encouraging it to capture compact yet informative subgraphs for OOD detection. Comprehensive evaluations show that GOODAT outperforms state-of-the-art benchmarks across various real-world datasets. GOODAT is a data-centric, unsupervised, and plug-and-play solution that does not require training data or modifying the original GNN architecture. It leverages the information bottleneck principle to design three GIB-boosted losses: subgraph GIB loss, masked graph GIB loss, and graph distribution separating loss. These losses help in learning informative subgraphs and separating the distributions of ID and OOD graphs. The method is tested on multiple datasets and scenarios, demonstrating its effectiveness in detecting OOD samples. The results show that GOODAT achieves significant improvements in graph OOD detection tasks compared to baselines. The method is also effective in graph anomaly detection, showing superior performance over other test-time-oriented techniques. The paper concludes that GOODAT is a promising approach for test-time graph OOD detection, offering a lightweight, training data-independent, and plug-and-play solution for graph OOD detection.