Dissimilarity-Preserving Representation Learning for One-Class Time Series Classification.

Area of Science:

• Machine Learning
• Data Science
• Time Series Analysis

Background:

• Time series classification often relies on complex dissimilarity measures like Dynamic Time Warping (DTW).
• High dimensionality of raw time series data poses computational and storage challenges.
• Existing methods may not efficiently preserve pairwise dissimilarities in reduced-dimension representations.

Purpose of the Study:

• To embed time series into a latent space that preserves pairwise Euclidean distances (EDs) corresponding to original dissimilarities.
• To develop elastic dissimilarity measures using neural networks for improved time series representation.
• To reduce dimensionality for efficient time series classification and one-class classification tasks.

Main Methods:

• Utilized auto-encoder (AE) and encoder-only neural networks to learn elastic dissimilarity measures.
• Employed dynamic time warping (DTW) as a key dissimilarity measure within the learned representations.
• Applied learned representations to one-class classification on UCR/UEA archive datasets using a 1-nearest neighbor (1NN) classifier.

Main Results:

• Achieved classification performance comparable to using raw time series data.
• Demonstrated a substantial reduction in the dimensionality of the data representation.
• Showcased significant savings in computational and storage requirements for nearest neighbor classification.

Conclusions:

• The proposed latent space embedding effectively preserves time series dissimilarities.
• Learned representations offer a computationally efficient alternative to raw data for time series classification.
• This method provides practical advantages for large-scale time series analysis and nearest neighbor classification.