GENDIS: Genetic Discovery of Shapelets.

Area of Science:

• Machine Learning
• Data Mining
• Time Series Analysis

Background:

• Shapelets are discriminative subsequences crucial for time series classification.
• Existing methods often require shapelets to be actual subsequences and can be computationally intensive.
• Interpretability of shapelets is valuable in critical domains with longitudinal data.

Purpose of the Study:

• To propose a novel paradigm for shapelet discovery using evolutionary computation.
• To overcome limitations of existing shapelet discovery methods, such as gradient dependency and brute-force search.
• To develop a method that jointly evolves shapelet sets, their lengths, and quantities for improved classification.

Main Methods:

• Utilizing evolutionary computation for gradient-free shapelet discovery.
• Developing an algorithm that jointly evolves shapelet sets, their lengths, and quantities.
• Evaluating entire sets of shapelets simultaneously to reduce redundancy.
• Allowing discovered shapelets to not be direct subsequences of the input time series.

Main Results:

• The proposed evolutionary approach is gradient-free, aiding in escaping local optima.
• The algorithm demonstrates scalability by avoiding brute-force search.
• Joint evolution of shapelet characteristics leads to smaller, more effective sets.
• The method produces shapelets that are not constrained to be subsequences of the input data, enhancing flexibility.

Conclusions:

• Evolutionary computation offers a powerful, flexible, and scalable alternative for shapelet discovery in time series classification.
• The proposed method enhances interpretability and predictive performance.
• This approach addresses key limitations of traditional shapelet discovery techniques.