LTSpAUC: Learning Time-Series Shapelets for Partial AUC Maximization.

Area of Science:

• Machine Learning
• Time Series Analysis
• Pattern Recognition

Background:

• Shapelets are key discriminative segments for time-series classification, offering interpretability and accuracy.
• Jointly learning shapelets and classifiers improves performance but requires efficient optimization.
• Partial area under the ROC curve (pAUC) is a critical metric for imbalanced datasets in industrial applications.

Purpose of the Study:

• To propose a novel method for jointly learning shapelets and classifiers optimized for pAUC across all false-positive rate (FPR) ranges.
• To introduce extensions for linear time complexity and explicit class-shapelet matching.
• To demonstrate the method's effectiveness and superiority over existing approaches in diverse industrial case studies.

Main Methods:

• Developed a joint learning framework for shapelets and classifiers targeting pAUC optimization.
• Introduced algorithmic complexity reduction to linear time complexity concerning time-series length.
• Implemented explicit determination of class-shapelet associations for enhanced interpretability.

Main Results:

• The proposed method demonstrated superior pAUC performance on the UCR time-series datasets compared to state-of-the-art methods.
• Achieved linear time complexity, significantly improving efficiency for long time series.
• Validated effectiveness through successful industrial case studies in medicine, manufacturing, and maintenance.

Conclusions:

• The proposed joint shapelet and classifier learning method effectively optimizes pAUC for time-series classification.
• Extensions provide significant algorithmic advantages and improved interpretability.
• The method shows strong practical utility and superior performance in real-world industrial scenarios.