Explainable Long- and Short-term Pattern Detection in Projected Sequential Data

Abstract

Combining explainable artificial intelligence and information visualization holds great potential for users to understand and reason about complex multidimensional sequential data. This work proposes a semi-supervised two-step approach for extracting long- and short-term patterns in low-dimensional representations of sequential data. First, unsupervised sequence clustering is used to identify long-term patterns. Second, these long-term patterns serve as supervisory information for training a self-attention-based sequence classification model. The resulting feature embedding is used to identify short-term patterns. The approach is validated on a self-generated dataset consisting of heart-shaped paths with different sampling rates, rotations, scales, and translations. The results demonstrate the approach's effectiveness for clustering semantically similar paths and/or path sequences. This detection of both global long-term patterns and local short-term patterns facilitates the understanding and reasoning about complex multidimensional sequential data.

Citation

Matthias Bittner, Andreas Hinterreiter, Klaus Eckelt, Marc Streit
Explainable Long- and Short-term Pattern Detection in Projected Sequential Data
ECML PKDD Workshop on Explainable AI for Time Series: Advances and Applications (XAI-TS '23), 2023.

BibTeX

@inproceedings{2023_pattern_detection_xai,
    title = {Explainable Long- and Short-term Pattern Detection in Projected Sequential Data},
    author = {Matthias Bittner and Andreas Hinterreiter and Klaus Eckelt and Marc Streit},
    journal = {ECML PKDD Workshop on Explainable AI for Time Series: Advances and Applications (XAI-TS '23)},
    year = {2023}
}