[ICLR 2022] Part 2: Time Series Anomaly Detection

Anomaly Detection is the task of finding outliers in data that do not follow normal patterns, and it is used in various fields such as risk management, security, and fraud detection. Classical approaches such as One-Class SVM or Kernel Density Estimation have limitations in learning the complex dynamics of high-dimensional time series data and generalizing to unknown scenarios.

Recently, deep learning has been applied to overcome these limitations, and since labeling abnormal data is difficult in real-world environments, research based on Unsupervised Learning has become the mainstream. The unsupervised learning approach trains a model such as an Auto-Encoder using only normal data, and then determines anomalies based on the Reconstruction Error between the input data and the reconstructed data.

However, existing deep learning models have limitations in that they reconstruct even abnormal data too well, or fail to sufficiently reflect the temporal context of time series. To address this, ICLR 2022 presented innovative models such as the Anomaly Transformer, which proposes a Transformer structure optimized for anomaly detection.