Automatic induction of projection pursuit indices

Area of Science:

• Machine Learning
• Data Science
• Computational Statistics

Background:

• Projection techniques are crucial for feature extraction and dimensionality reduction in machine learning.
• Projection Pursuit (PP) is a key method, relying on a projection index to guide transformation.
• Designing effective PP indices is challenging and often relies on expert intuition.

Purpose of the Study:

• To address the design of Projection Pursuit (PP) index functions for linear feature extraction.
• To develop an automated method for creating new PP indices tailored to specific datasets.
• To compare the performance of automatically generated indices against expert-designed ones.

Main Methods:

• An evolutionary search framework was employed to generate new PP index functions.
• The framework utilizes a rich set of function primitives for high expressive power.
• The generated indices were evaluated on classification tasks.

Main Results:

• Automatically generated PP indices demonstrated improved performance compared to existing human-designed indices.
• A decrease in classification errors was observed when using the novel, data-driven indices.
• The evolutionary approach successfully created effective indices suited to dataset properties.

Conclusions:

• Evolutionary search provides a powerful mechanism for designing effective Projection Pursuit indices.
• Automated index generation can enhance feature extraction and classification performance in machine learning.
• This data-driven approach offers a flexible alternative to manual index design.