Estimating Node Abnormalities From Imprecise Subgraph-Level Supervision

Although existing node anomaly detection techniques following supervised or unsupervised paradigms have achieved empirical success, the expensive annotation cost and the high possibility of spotting uninteresting outliers are still two major limitations. Interestingly, with limited time, workforce, and expertise, it seems easier for regulators to roughly screen out a suspicious region likely to contain anomalies rather than precisely locate the problematic individual. Thus, it is appealing to investigate how to learn node abnormalities from such coarse-grained imprecise subgraph-level supervision, related research however has received little scrutiny. In this paper, we generalize classic multiple instance learning to graph data and propose a novel architecture ASSESS which regards subgraphs as bags and nodes as bag instances. By pulling apart the score gap between anomalies and normal nodes through inter-bag loss and intra-bag loss, ASSESS tends to assign higher scores to abnormal nodes so that node anomalies can be detected. To further enhance the adaptability to low labeling rates, the self-training mechanism is introduced to ASSESS++, which automatically explores subgraphs that may cover abnormal nodes as pseudo-supervision by measuring the node score distribution in the subgraph. Experiments on real-world benchmark datasets corroborate the superiority of our proposed model w.r.t. AUC and AP.