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

Aras Yurtman, Daan Van Wesenbeeck, Wannes Meert, Hendrik Blockeel · 2025

Read Steering the LoCoMotif: Using Domain Knowledge in Time Series Motif Discovery by Aras Yurtman, Daan Van Wesenbeeck, Wannes Meert, Hendrik Blockeel (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-11850.txt.

Introduction

Aras Yurtman and three co-authors present a 2025 computer science paper on patterns in data streams. Many fields record values over time, from heart traces to machine sensors. Hidden in such streams are short shapes that repeat, called motifs. Software can find them on its own, but it may surface shapes no expert cares about.

The team built a frame that lets users steer the search with rules. Rules are stated in terms of traits the wanted shapes should have. Their program, LoCoMotif-DoK, runs that search fast and with care. Tests on real and synthetic data show it uses expert rules well. It beat older tools that accept only a narrow kind of guidance. The work aims to make pattern hunts match what each field actually needs.

Catalog record

CREATORAras Yurtman, Daan Van Wesenbeeck, Wannes Meert, Hendrik Blockeel
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-11850

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Steering the LoCoMotif: Using Domain Knowledge in Time Series Motif Discovery Aras Yurtman, Daan Van Wesenbeeck, Wannes Meert, Hendrik Blockeel 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. keywords pattern mining, time series, motif discovery, domain knowledge, constrained optimization 1 Introduction Time series data often contain repeating patterns, some of which might be interesting because they reveal important aspects of the underlying process. Such patterns can be manually extracted by experts, but this is a time-consuming and costly approach. An alternative is to use Time Series Motif Discovery (TSMD) , which automatically identifies (approximately) repeating patterns in a time series in an unsupervised manner. In the context of TSMD, a time segment that contains a single occurrence of a pattern is called a motif , and the set of all occurrences of the same pattern is called a motif set . Due to the unsupervised nature of TSMD, the identified motif sets might not be aligned with the user’s intentions. To demonstrate this, consider a physiotherapy use case where a subject performs certain exercises repeatedly and their body motions are captured by wearable sensors as time series data. In such a time series, the desired motifs are individual exercise executions grouped into motif sets according to their execution type. Figure 1 a depicts an example time series that contains an exercise performed in three different execution types, each type repeated ten times. In this time series, an unsupervised TSMD technique fails to discover the desired motif sets (Figure 1 b); instead, it mostly finds motif sets that consist of parts of exercise executions and idle time periods between them (Figure 1 c). We call such motif sets, which are not interesting to the user, off-target motif sets. Off-target motif sets are also discovered in the rest of the physiotherapy dataset (Section 6.1 ) as well as in other application domains such as electrocardiogram, seismology, and epileptic seizure monitoring (Section 6 and Dau and Keogh, (2017) ). 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 ( Van Wesenbeeck et al., 2024 ) in (c) and the proposed one, LoCoMotif-DoK , in (d). To discover desired motif sets in TSMD, we leverage domain knowledge , which is the available information about the desired motifs and can be defined by an expert without much effort in many applications. Domain knowledge includes properties of motifs or motif sets to be discovered, such as restrictions on and preferences for certain patterns, and does not require manually annotating occurrences of any patterns. In our physiotherapy example, we have three pieces of domain knowledge: (1) The subject was advised to perform the exercises ten times, (2) every exercise execution starts and ends when the subject is at resting position, and (3) each execution type has a specific range of duration. Leveraging these three pieces of domain knowledge enables us to exclusively discover the desired motif sets in the example time series (Figure 1 d). Although incorporating domain knowledge into TSMD is valuable, only a few existing TSMD techniques support it, and their support is quite limited. To represent domain knowledge, these techniques merely consider an auxiliary time series that specifies for every time index whether a motif can start (or end) there or how preferable such a motif is ( Dau and Keogh, 2017 ; Mohammad and Nishida, 2009 ; Van Wesenbeeck et al., 2024 ) . This approach is not expressive enough to represent a wide variety of domain knowledge; for example, to define the knowledge in our running example. Some TSMD techniques have hyperparameters (such as motif length) that can be selected based on domain knowledge, but they are also not expressive enough. For example, a hyperparameter that specifies the motif length range usually applies to all discovered motifs, whereas the domain knowledge in our running example includes different ranges for different motif sets to be discovered. Therefore, most forms of domain knowledge cannot be leveraged by existing TSMD techniques. This paper first proposes a framework for incorporating domain knowledge into TSMD in a generic and highly expressive manner. The framework supports domain knowledge expressed by hard and soft constraints that respectively represent the restrictions on and preferences for the patterns to be discovered. Constraints can be defined differently for individual motif sets and for specific pairs of motif sets. Defining constraints does not require any knowledge about the TSMD technique being used. Although constraints can easily be defined by the user, for convenience, we provide a catalogue of constraints that might be useful in multiple applications. Secondly, this paper proposes an implementation of the proposed framework based on an existing TSMD technique, LoCoMotif ( Van Wesenbeeck et al., 2024 ) , leading to LoCoMotif-DoK (LoCoMotif with Domain Knowledge). We evaluate LoCoMotif-DoK based on three different experiments with different types of domain knowledge: a physiotherapy use case, a benchmark for TSMD, and time series from other applications. In general, LoCoMotif-DoK successfully leverages domain knowledge, outperforming existing TSMD techniques that support domain knowledge. The remainder of this paper is organized as follows: After providing the notation (Section 2 ), we describe our proposed framework for TSMD with domain knowledge (Section 3 ) and its implementation LoCoMotif-DoK (Section 4 ). We then provide an overview of the related work (Section 5 ). We evaluate LoCoMotif-DoK in a physiotherapy use case, on a TSMD benchmark, and in other applications (Section 6 ). Lastly, we draw conclusions (Section 7 ). 2 Notation A time series \mathbf{x} of length n is a sequence of n samples (i.e., measurements): {\mathbf{x}=[x_{1},\ldots,x_{n}]} where x_{i}{\,\in\,}\mathbb{R} for univariate time series and x_{i}{\,\in\,}\mathbb{R}^{D} for D -dimensional multivariate time series for i=1,\ldots,n . A (time) segment \beta=[b:e]=[b,b+1,\ldots,e] is an interval of time indices with length |\beta|=e-b+1 for 1\leq b\leq e\leq n . The subsequence of \mathbf{x} that corresponds to \beta is \mathbf{x}_{\beta}=[x_{b},x_{b+1},\ldots,x_{e}] . To check whether a segment \beta overlaps with another segment \beta^{\prime} excessively or not, we define that \beta is \nu -coincident to \beta^{\prime} if |\beta\cap\beta^{\prime}|>\nu\cdot|\beta^{\prime}| where \nu is the overlap parameter. A motif in \mathbf{x} is a subsequence \mathbf{x}_{\beta} (which corresponds to the segment \beta ) that contains a pattern that occurs multiple times in \mathbf{x} . Occurrences of the same pattern constitute a motif set : \mathcal{M}=\{\beta_{1},\ldots,\beta_{k}\} , where |\mathcal{M}|=k is the cardinality of \mathcal{M} . The time coverage of \mathcal{M} is defined as the overall time duration covered by the motifs: \text{cov}(\mathcal{M})=\left|\bigcup_{\beta{\,\in\,}\mathcal{M}}\beta\right| . 3 Proposed Framework for Motif

Read the complete text