Fun-TSG: A Function-Driven Multivariate Time Series Generator with Variable-Level Anomaly Labeling

A team of researchers from IRIT and the University of Toulouse has introduced Fun-TSG, a novel tool designed to solve a persistent problem in AI research: the lack of high-quality, transparent benchmark datasets for evaluating anomaly detection in multivariate time series. Current datasets often lack fine-grained labels and obscure the underlying data-generating processes, making it difficult to compare models or understand why they fail. Fun-TSG addresses this by generating synthetic time series data where every anomaly is precisely labeled at the variable and timestamp level, and the mathematical functions governing variable dependencies are fully known.

The tool offers two modes: a fully automated mode that randomly samples dependency structures and anomaly types, and a manual mode where researchers can define their own equations and anomaly configurations. This dual approach supports diverse and reproducible benchmarking scenarios, from testing classical statistical models to modern deep learning approaches. By providing 'ground truth' for every generated datapoint, Fun-TSG enables fine-grained performance analysis, helping developers pinpoint whether a model fails to detect a specific type of anomaly or misattributes it to the wrong variable.

This transparency is critical for advancing interpretable AI. The ability to rigorously test models on data with known properties accelerates the development of more reliable and explainable anomaly detection systems for critical applications like industrial monitoring, financial fraud detection, and healthcare diagnostics.