Discovering Multiple Subdimensional Motifs in Multivariate Time Series

Motif Discovery aims at identifying repeated patterns in time series. It is a fundamental task in time series analysis, with applications across numerous fields where recurring phenomena are observed. In multivariate time series, motifs may appear only on a subset of dimensions, making their identification more challenging. Existing approaches to multivariate Motif Discovery face limitations when many motifs are present. In particular, current methods struggle to handle ambiguities arising when motifs overlap in time but span disjoint subsets of dimensions. To address this issue, we formalize these ambiguous cases and propose a method based on a more expressive notion of overlap that accounts for both temporal alignment and dimensional involvement. Finally, we introduce a new evaluation framework that jointly assesses temporal and dimensional accuracy, and demonstrate its relevance through carefully designed experiments.