Optimal Parameter Selection Using Explainable AI for Time-Series Anomaly Detection

Time-series anomaly detection is a technique for detecting unusual values, changes, or movements in a large amount of data arranged in time-series. It is primarily used in the fields of intrusion detection, medical diagnosis, and industrial defect damage detection and necessary to realize agents that operate intelligently and autonomously, such as changing system behavior based on detected anomalies. SALAD is a real-time time-series anomaly detection method based on deep learning. It is lightweight and determines anomaly detection threshold flexibly; however, experts need to determine an appropriate value for a parameter so that it suits any given recurrent time series, and this inhibits the realization of the agent. In this study, we propose a method to determine automatically the optimal parameter value in SALAD's prediction model by utilizing XAI. We use SHAP, which provides interpretability to the prediction by the deep learning model. Through evaluation experiment, we demonstrate that our method is effective and provide an example of the use of XAI for time-series anomaly detection.