MotifDisco: Motif Causal Discovery For Time Series Motifs
Josephine Lamp, Mark Derdzinski, Christopher Hannemann, Sam Hatfield, Joost van der Linden
TL;DR
MotifDisco is a novel framework that discovers causal relationships among motifs in time series data, especially health data like glucose traces, improving understanding and enabling better downstream applications.
Contribution
We introduce MotifDisco, the first causal discovery method for time series motifs, formalizing Motif Causality and leveraging Graph Neural Networks for unsupervised learning.
Findings
Significant performance improvements in forecasting, anomaly detection, and clustering tasks.
Effective identification of causal motifs in health data streams.
Framework applicable to various time series domains.
Abstract
Many time series, particularly health data streams, can be best understood as a sequence of phenomenon or events, which we call \textit{motifs}. A time series motif is a short trace segment which may implicitly capture an underlying phenomenon within the time series. Specifically, we focus on glucose traces collected from continuous glucose monitors (CGMs), which inherently contain motifs representing underlying human behaviors such as eating and exercise. The ability to identify and quantify \textit{causal} relationships amongst motifs can provide a mechanism to better understand and represent these patterns, useful for improving deep learning and generative models and for advanced technology development (e.g., personalized coaching and artificial insulin delivery systems). However, no previous work has developed causal discovery methods for time series motifs. Therefore, in this paper…
Peer Reviews
Decision·Submitted to ICLR 2025
* Empirically , using the motif causal graphs, improvements in the 3 downstream tasks are identified demonstrate the use of motifs * Comparison against chunked time series prediction should be presented * Chunking provides a simple way of making all motifs of the same length
* Comparing against existing granger causal techniques applied directly on time lagged variables needs to be done to validate the necessity of motifs * Identifiability of motifs is left as out of scope, but should be discussed as that defines the nodes used in the causal graph construction. * Metadata such as time of occurrence and frequency of occurrence of motifs is not well presented in an interpretable manner in the link prediction task, how this might be captured in the motif representation
1. Novelty of Causal Discovery Framework: The introduction of Motif Causality (MC) for time series motifs and the development of MotifDisco fills an important gap in time series analysis, especially for health-related data. No prior work has explicitly targeted causal discovery among motifs within time series, which makes this a novel contribution. 2. Flexible Application Scope: Integrating MC into multiple use cases, namely forecasting, anomaly detection, and clustering, demonstrates the propos
1. Motif Construction Limitation: The method for constructing motifs is largely dependent on heuristic techniques (e.g., chopping or sliding windows). This may lead to arbitrary definitions of motifs that do not always correspond to well-defined physiological phenomena. The authors could consider more dynamic motif extraction methods. 2. Lack of Personalization: The majority strategy for causal inference used in the GNN might overlook personalized differences across individuals, which could limi
The paper is well written. The literature survey is good. Work is mathematically sound and the author shows run time requirements. The idea of showing the model performance on three tasks was also good.
Marginal technical novelty. What is the contribution compared to Lamp et el (2024) needs to be discussed. Pan et al., 2024; Lowe ¨ et al., 2022 Bonetti et al., 2024; Najafi et al., 2023) are already using Granger Causality and Transfer Entropy then what is contribution compared to this work is not clear to me. Please explicitly state key technical contributions of this paper and how it differ from or improve upon the cited works, particularly in the context of motif-based causal discovery for
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsTime Series Analysis and Forecasting · Neural Networks and Applications · Music and Audio Processing
MethodsFocus