Steering the LoCoMotif: Using Domain Knowledge in Time Series Motif Discovery

TL;DR

Steering the LoCoMotif advances Time Series Motif Discovery by enabling user supplied domain knowledge through a formal constrained optimization framework. The LoCoMotif-DoK algorithm integrates hard and soft constraints with a greedy search that leverages LoCoMotif candidate sets, yielding motif discoveries that align with application specific needs. Across physiotherapy, a TSMD benchmark, and ECG-like data, LoCoMotif-DoK consistently outperforms traditional TSMD methods that lack expressive domain knowledge, demonstrating practical impact for more relevant pattern discovery in real-world time series. The framework’s generality and its catalogue of constraints position it as a versatile tool for domain-guided motif discovery in diverse applications.

Abstract

Time Series Motif Discovery (TSMD) identifies repeating patterns in time series data, but its unsupervised nature might result in motifs that are not interesting to the user. To address this, we propose a framework that allows the user to impose constraints on the motifs to be discovered, where constraints can easily be defined according to the properties of the desired motifs in the application domain. We also propose an efficient implementation of the framework, the LoCoMotif-DoK algorithm. We demonstrate that LoCoMotif-DoK can effectively leverage domain knowledge in real and synthetic data, outperforming other TSMD techniques which only support a limited form of domain knowledge.

Figures (11)

• Figure 1: (a) From the time series that captures the acceleration of the lower arm, (b) executions of a physical therapy exercise are (d) successfully discovered by using domain knowledge, (c) but not without using domain knowledge. We use the existing TSMD technique LoCoMotif locomotif in (c) and the proposed one, LoCoMotif-DoK, in (d).
• Figure 2: Taxonomy of constraints. Dashed rectangles indicate special cases of constraints.
• Figure 3: Flowchart of LoCoMotif-DoK.
• Figure 4: Left: LoCo relates similar time segments of $\mathbf{x}$ with each other by local warping paths $\mathcal{P}$, where one of the paths is highlighted. Right: From a given representative segment $\alpha$, a candidate motif set $\mathcal{C} = \{\beta_1, \beta_2, \beta_3\}$ is obtained efficiently by identifying segments that are similar to $\alpha$ according to $\mathcal{P}$.
• Figure 5: In the physiotherapy use case, LoCoMotif-DoK can leverage comprehensive domain knowledge and outperforms LoCoMotif that supports a much more limited form of domain knowledge in terms of different evaluation metrics for most strictness ($\rho$) values.