Causanom: Anomaly Detection With Flexible Causal Graphs

Abstract

Causality-based anomaly detection methods provide at least two significant theoretical benefits over purely statistical methods: 1. Improved robustness to non-anomalous out-of-distribution data, which implies a reduction in false-alarms; 2. A potential for failure localization due to the topological ordering of the causal graph. Recent studies have considered the utilization of causality-based methods for time series anomaly detection, however, these methods require the causal graph to be fixed; resultingly, such methods are not robust to incorrectly estimated causal graphs and are not able to natively model counterfactual scenarios. To address these limitations, we introduce Causanom: a graph-based encoder-decoder neural network for time series anomaly detection. Causanom utilizes a node conditional data-stream representation in conjunction with a weighted graph aggregation function in order to efficiently capture heterogeneous node dynamics whilst allowing for a flexible graphical structure. We show that Causanom can be trained along with auxiliary constraints in order to tune the causal graph and improve performance. Additionally, we show that Causanom can be used to produce counterfactual data, which we leverage to identify violated causal relationships. Using real and synthetic time series data respectively, we show that Causanom performs at least as well as state-of-the-art baselines in the anomaly detection task and outperforms existing methods in a causal attribution task.