Novel Node Category Detection under Subpopulation Shift

In real-world graph data, distribution shifts can manifest in various ways,such as the emergence of new categories and changes in the relative proportionsof existing categories. It is often important to detect nodes of novelcategories under such distribution shifts for safety or insight discoverypurposes. We introduce a new approach, Recall-Constrained Optimization withSelective Link Prediction (RECO-SLIP), to detect nodes belonging to novelcategories in attributed graphs under subpopulation shifts. By integrating arecall-constrained learning framework with a sample-efficient link predictionmechanism, RECO-SLIP addresses the dual challenges of resilience againstsubpopulation shifts and the effective exploitation of graph structure. Ourextensive empirical evaluation across multiple graph datasets demonstrates thesuperior performance of RECO-SLIP over existing methods.