PhD: "Learning Explainable Constrained Representation for Time Series"

PhD defence : Hussein El Amouri

Team: SDC

Date & time : September 22nd, at 09:00 AM at ICube, Illkirch (Room A301)

Title : "Learning Explainable Constrained Representation for Time Series"

Abstract : This PhD research introduces a novel approach to learning explainable constrained representations for time series data, with a primary focus on addressing two critical challenges: the lack of labeled data and time series distortions. Labeling time series data is a resource-intensive process; therefore, clustering methods are widely used. However, these clustering methods may not produce results aligned with expert requirements and can benefit from further improvement through constrained clustering. Second, time series data often contain distortions, and capturing these distortions is essential for accurate clustering. Our proposed approach empowers experts to guide the learning process by providing information (constraints) about pairs of samples, indicating whether they are similar or dissimilar. This concept stems from the framework of constrained learning, where experts provide information to guide the clustering process. This approach is rooted in the concept of 'shapelets,' which are discriminative subsequences capable of distinguishing between different patterns within time series data. We extend our work from a previous study on unsupervised shapelet learning, referred to as  'Learning DTW Preserving Shapelets,' into a semi-supervised approach. In this extension, we leverage both types of constraints while approximating Dynamic Time Warping (DTW) distances. DTW is a distance measure well-suited for accommodating the various distortions commonly encountered in time series data. We term this extended approach 'Constrained DTW Preserving Shapelets,' as it effectively incorporates the constraints provided by experts into the learning process while approximating DTW in the transformed space. Furthermore, our research delves into the interpretability of shapelets and explores their potential for providing descriptions of clustering results. This interpretability enables experts to gain a deeper understanding of the clustering results, facilitating an interactive process that ultimately leads to improved clustering outcomes. Additionally, by utilizing the feature space, we can investigate the possibility of proposing further constraints to experts, enabling an active and informed clustering process.

The jury will be composed of:

• President — Mr. Cédric Wemmert (Professor, University of Strasbourg, France)
• Reporter — Mr. Romain Tavenard (Professor, University of Rennes 2, France)
• Reporter — Ms. Seham Amer-Yehia (Director of Research, CNRS, France)
• Examinator — Ms. Thi-Bich-Hanh Dao (Assistant Professor, University of Orléans, France )
• Supervisor — Mr. Pierre Gançarski (Professor, University of Strasbourg, France)
• Supervisor — Mr. Thomas Lampert (Assistant Professor, University of Strasbourg, France)
• Supervisor — Mr. Clément Mallet (Director of Research, IGN, University Gustave Eiffel, France)