[1Yin-B] Poster 2

[1Yin-B-12] Anomaly Detection via Bayesian Causal Discovery

〇Fushihara Daichi¹, Ishibashi Kesisuke² (1. NTT, Inc., 2. ICU)

Keywords:

anomaly detection,causal discovery,Baysian inference

Anomaly detection is important for maintaining the safety and reliability of systems in industrial domains. To perform anomaly detection, it is necessary to use methods based not only on correlations among data but also on directed acyclic graphs (DAGs) that represent causal relationships between variables. Conventionalanomalu detection methods assume that a DAG can be uniquely estimated from observational data. However, when this assumption does not hold, their performance degrades. In this paper, we propose a method for performing anomaly detection using Bayesian causal discovery, which is applicable even when a DAG cannot be uniquely identified from observational data. We sample DAGs using the Gumbel–Max trick, and estimate the posterior distribution of DAGs representing the causal structure of normal data via variational inference. During anomaly detection, we sample multiple DAGs from the learned posterior distribution of DAGs, and conduct anomaly detection using the deviation of test data from the sampled DAGs. We also report experimental results using synthetic data.