A New Algorithm to Optimize Maximal Information Coefficient.

Area of Science:

• Statistics
• Machine Learning
• Data Mining

Background:

• The Maximal Information Coefficient (MIC) is a metric used to quantify the dependence between paired variables, encompassing both functional and non-functional relationships.
• Existing methods for MIC calculation, such as ApproxMaxMI, have limitations including grid size restrictions and computational costs.

Purpose of the Study:

• To develop a new, efficient, and accurate method for calculating MIC values.
• To improve upon the limitations of existing MIC algorithms, specifically ApproxMaxMI.
• To evaluate the performance of the new method in terms of accuracy, statistical power, and computational cost, particularly in feature selection applications.

Main Methods:

• Development of the ChiMIC algorithm, which utilizes the chi-square test to optimize grid selection.
• Modification of the grid optimization process to remove the maximal grid size limitation inherent in ApproxMaxMI.
• Comparative computational experiments to assess MIC values for noiseless and noisy functional relationships, as well as for independent variables.

Main Results:

• ChiMIC maintains equivalent MIC values for noiseless functional relationships compared to ApproxMaxMI.
• ChiMIC yields significantly smaller MIC values for independent variables, reducing false positives.
• For noisy functional relationships, ChiMIC achieves optimal partitioning faster and demonstrates higher statistical power.
• The Mutual Information Coefficient (MCN) values derived from ChiMIC better capture functional relationship complexity.
• Feature selection using ChiMIC-derived MIC values resulted in improved classification accuracy.

Conclusions:

• ChiMIC offers a superior approach to calculating MIC, providing enhanced accuracy, statistical power, and computational efficiency.
• The method effectively distinguishes between dependent and independent variables and excels in noisy data scenarios.
• ChiMIC-based feature selection demonstrates practical utility by improving classification performance.