Reproduce “Clustering-based Anomaly Detection in Multivariate Time Series Data” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jinbo Li, Hesam Izakian, Witold Pedrycz, Iqbal Jamal
Researchers at the University of Alberta introduce an extended Fuzzy C-Means framework, optimized with Particle Swarm Optimization, for unsupervised anomaly detection in multivariate time series data. This method dynamically weights variable contributions, enabling it to detect both amplitude and shape anomalies and achieving superior performance over HMM-based techniques, with accuracies up to 0.9983 on real-world datasets.
Key insights
- Dynamically weighting individual variables within the clustering process (via λ_i in extended FCM) significantly enhances the accuracy of cluster formation for multivariate time series.
- Reconstruction error, derived from the optimized cluster structure, serves as a robust and unsupervised anomaly score capable of identifying both amplitude and shape deviations.
- Transforming subsequences into their autocorrelation coefficient representation effectively normalizes patterns, allowing for accurate shape anomaly detection by removing time shift dependencies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.