Abstract
Open-set graph learning is a practical task that aims to classify the knownclass nodes and to identify unknown class samples as unknowns. Conventionalnode classification methods usually perform unsatisfactorily in open-setscenarios due to the complex data they encounter, such as out-of-distribution(OOD) data and in-distribution (IND) noise. OOD data are samples that do notbelong to any known classes. They are outliers if they occur in training (OODnoise), and open-set samples if they occur in testing. IND noise are trainingsamples which are assigned incorrect labels. The existence of IND noise and OODnoise is prevalent, which usually cause the ambiguity problem, including theintra-class variety problem and the inter-class confusion problem. Thus, toexplore robust open-set learning methods is necessary and difficult, and itbecomes even more difficult for non-IID graph data.To this end, we propose aunified framework named ROG$_{PL}$ to achieve robust open-set learning oncomplex noisy graph data, by introducing prototype learning. In specific,ROG$_{PL}$ consists of two modules, i.e., denoising via label propagation andopen-set prototype learning via regions. The first module corrects noisy labelsthrough similarity-based label propagation and removes low-confidence samples,to solve the intra-class variety problem caused by noise. The second modulelearns open-set prototypes for each known class via non-overlapped regions andremains both interior and border prototypes to remedy the inter-class confusionproblem.The two modules are iteratively updated under the constraints ofclassification loss and prototype diversity loss. To the best of our knowledge,the proposed ROG$_{PL}$ is the first robust open-set node classification methodfor graph data with complex noise.