Functional partitioning through competitive learning.

Area of Science:

• Data Science
• Machine Learning
• Algorithm Development

Background:

• Datasets frequently contain diverse functional patterns representing different aspects or regimes.
• These patterns are often unevenly distributed, posing challenges for analysis.

Purpose of the Study:

• To develop a novel partitioning algorithm for detecting and separating functional patterns in datasets.
• To demonstrate the utility of this algorithm in improving model performance through specialization.

Main Methods:

• A competitive learning approach where multiple models predict data points.
• A reward mechanism that trains models on data points where their predictions are best, fostering specialization.
• Validation using datasets with distinct patterns (e.g., mechanical stress/strain) and application to regression problems.

Main Results:

• The algorithm successfully detects and separates functional patterns, providing valuable dataset insights.
• Modular models, each specialized in a partition, significantly outperformed single models learning all partitions simultaneously.
• Up to 56% reduction in loss was observed in regression tasks using the proposed method.

Conclusions:

• The proposed partitioning algorithm effectively leverages model competition to uncover hidden data structures.
• Specialized modular models derived from the partitioning scheme offer superior performance compared to monolithic models.
• This approach has broad applicability for analyzing complex datasets and enhancing predictive modeling.