Facilitating time series classification by linear law-based feature space transformation.

Area of Science:

• Machine Learning
• Data Science
• Signal Processing

Background:

• Time series classification is crucial for pattern recognition in sequential data.
• Existing methods often struggle with high dimensionality and computational cost.
• Feature engineering significantly impacts classifier performance.

Purpose of the Study:

• To introduce and evaluate a novel Linear Law-based Transformation (LLT) for time series classification.
• To assess LLT's effectiveness in improving accuracy and computational efficiency.
• To demonstrate LLT's capability in handling both uni- and multivariate time series.

Main Methods:

• Linear Law-based Transformation (LLT) for feature space transformation.
• Time-delay embedding and spectral decomposition to identify sequence patterns.
• Application of LLT to separate training and test sets and transform test data.
• Empirical evaluation using the AReM human activity recognition database.

Main Results:

• LLT significantly boosts the accuracy of traditional time series classifiers.
• The proposed LLT method outperforms current state-of-the-art techniques.
• Combining LLT with k-nearest neighbors (KNN) achieved the fastest error-free classification.
• Fivefold cross-validation confirmed the robustness of the results.

Conclusions:

• LLT offers a computationally efficient and highly effective approach to time series classification.
• The method demonstrates strong potential for developing advanced learning algorithms.
• LLT represents a significant advancement in feature engineering for sequential data analysis.