時系列データに対する異常検知のための教師なし学習方法の提案

Abstract

時系列データを対象とした異常検知を行う際に、異常データの事前情報が収集困難な場合や、未知の異常検知を目的として、教師なし学習による異常検知を行う場合でも、教師なし異常検知モデルに依存しない定量的な性能評価が可能となる手法を検討した。これにより、教師なし異常検知モデルのハイパーパラメータの最適化や異常判定のための閾値の設定が容易になることが期待できる。本報では、この手法について紹介するとともに、この手法を使って実際に試行した異常検知結果を示す。

In this paper, we propose an unsupervised learning method for time series anomaly detection, which is independent of unsupervised anomaly detection models. Although unsupervised anomaly detection is effective when it is difficult to collect anomaly data beforehand or when the goal is to detect unknown anomalies, there is a problem that there is no model-independent performance metrics such as the normal/abnormal sample accuracy in supervised learning. We use Kullback-Leibler divergence between the predictive distribution of time-series anomaly scores based on Gaussian process regression and the actual distribution of calculated anomaly scores as a metrics for evaluating the performance of anomaly detectors using unsupervised learning. This method is expected to facilitate the optimization of hyperparameters for unsupervised anomaly detection models and the setting of thresholds for anomaly score. In this paper, we introduce the proposed method and show the actual anomaly detection results of the proposed method.