TiCTok: Time-Series Anomaly Detection With Contrastive Tokenization

Abstract

Detecting anomalies in multivariate time-series data is an important task in various real world applications. Recent advances using deep learning have shown promising results in this area. Nowadays, Transformer-based models have shown outstanding performance and contrastive learning has emerged as a powerful technique for representation learning, however, it may not be directly applicable to the time-series domain. Here, we propose a time-series anomaly detection model with contrastive tokenization (TiCTok). We propose a time-series token encoder to transform raw time-series data into latent embeddings containing high-level wide-range temporal information. We exploit both token encoder and contrastive learning to produce high quality latent embeddings. In addition, we propose a novel anomaly scoring method simply utilizing the contrastive loss used in the training phase. According to our experimental results, the proposed model achieved better or comparable performance compared to the previous state-of-the-art on five widely used benchmark datasets in terms of F1-score.