Unsupervised Feature Construction for Anomaly Detection in Time Series -- An Evaluation

TL;DR

This work tackles unsupervised anomaly detection in univariate time series by shifting from raw temporal data to a rich tabular feature space using TSFRESH, followed by standard tabular detectors. It systematically compares Isolation Forest and Local Outlier Factor across five diverse datasets, showing that feature construction notably enhances IF performance but yields limited or conditional benefits for LOF. The findings highlight the value of representation learning through feature extraction for TSAD and suggest that tree-based, rank-focused detectors are more receptive to high-dimensional feature spaces. The study paves the way for broader evaluator extensions, including more detectors and alternative feature pipelines, under offline analysis scenarios.

Abstract

To detect anomalies with precision and without prior knowledge in time series, is it better to build a detector from the initial temporal representation, or to compute a new (tabular) representation using an existing automatic variable construction library? In this article, we address this question by conducting an in-depth experimental study for two popular detectors (Isolation Forest and Local Outlier Factor). The obtained results, for 5 different datasets, show that the new representation, computed using the tsfresh library, allows Isolation Forest to significantly improve its performance.

Figures (2)

• Figure 1: Proposed processing pipeline.
• Figure 2: Critical diagrams for each dataset, averaged over the four window sizes.