Graph Energy Variety Network-based Out-of-distribution Detection

Graph relational structures are ubiquitous and prediction problems on graphs are very popular, such as node prediction vs. edge prediction. However, current models concentrating on improving test performance on intra-distributed data and largely ignoring the potential risks of out-of-distribution (OOD) test samples. In some cases, models may lead to negative results if they misclassify anomalous or out-of-distribution data. In this paper, we investigate the problem of OOD detection for graph-structured data and identify an effective OOD recogniser based on the loss of an energy function extracted directly from a graph neural network trained using standard classification losses, and constructed a way to use graph-based neural network learning in the context of energy theory. More importantly, it can further enhance the recognition of out-of-distribution data through unlearned hierarchical energy transfer mechanisms and energy attenuation schemes. For a comprehensive evaluation, we have conducted experiments using a recognised benchmark setup and our proposed method has achieved very good results in various experiments.