Matrix Profile Indexing for Multivariate Time Series in Impulsive Environments

Abstract

Real-world time series from environmental sensors and industrial systems are frequently corrupted by impulsive noise, where extreme events occur with far greater probability than what Gaussian models predict. In such settings, the Euclidean distance underlying standard matrix profile computation becomes unreliable, as squared differences amplify the effect of heavy-tailed outliers on similarity search. In this talk, we introduce a generalization of the matrix profile framework that extends the Euclidean distance with an Lp (pseudo-)norm, enabling robust motif discovery and discord detection under heavy-tailed noise. Experimental evaluation on real data demonstrates that the Lp-based matrix profile consistently outperforms the Euclidean baseline in preserving the ranking of true anomalies, with the advantage increasing as noise becomes more impulsive. The approach extends an existing open-source Python toolkit for irregular multivariate time series, offering a robust alternative for anomaly detection in non-Gaussian environments.

Bio: Georgia M. Tsanta is a postgraduate student in the Computer Science and Engineering program at the University of Crete and a research fellow at the Signal Processing Laboratory of the Institute of Computer Science at FORTH. She earned her BSc in Computer Science from the University of Crete in 2025. Her broader scientific interests encompass Machine Learning, Data Science, Time Series Analysis, and Artificial Intelligence applications.