Abstract
Out-of-distribution (OOD) detection remains challenging in text-richnetworks, where textual features intertwine with topological structures.Existing methods primarily address label shifts or rudimentary domain-basedsplits, overlooking the intricate textual-structural diversity. For example, insocial networks, where users represent nodes with textual features (name, bio)while edges indicate friendship status, OOD may stem from the distinct languagepatterns between bot and normal users. To address this gap, we introduce theTextTopoOOD framework for evaluating detection across diverse OOD scenarios:(1) attribute-level shifts via text augmentations and embedding perturbations;(2) structural shifts through edge rewiring and semantic connections; (3)thematically-guided label shifts; and (4) domain-based divisions. Furthermore,we propose TNT-OOD to model the complex interplay between Text aNd Topologyusing: 1) a novel cross-attention module to fuse local structure intonode-level text representations, and 2) a HyperNetwork to generatenode-specific transformation parameters. This aligns topological and semanticfeatures of ID nodes, enhancing ID/OOD distinction across structural andtextual shifts. Experiments on 11 datasets across four OOD scenariosdemonstrate the nuanced challenge of TextTopoOOD for evaluating OOD detectionin text-rich networks.