Issue with Feature Selection Algorithms in Matlab Classification Learner Except for MRMR

Hi, I am relatively new to Matlab and currently facing a challenge with the feature selection process in the Classification Learner app. My dataset consists of speech parameter values extracted from voice records, stored in a CSV file, with 61 columns and 700,000 rows. Parameters include features like MFCC, delta, delta delta, LPCC, HNR, ZCR.

After applying the feature selection, I've noticed that only the MRMR algorithm produces interpretable importance scores. Other algorithms such as Chi2, ANOVA, and Kruskal-Wallis return an importance score of infinity for most parameters.

I'm wondering if theres a straightforward explanation for this. Could be this related to the type of data each algorithm expects as input?

Thank you for your help.

Answer 1

[Drew]

This appears to be as designed. As seen at https://www.mathworks.com/help/stats/feature-selection-and-feature-transformation.html, the algorithms Chi2, ANOVA, and Kruskal-Wallis produce scores that correspond to -log(p). When the estimated p-value is zero, the score is infinite, or "Inf". A score of "Inf" indicates that, according to that feature ranking algorithm, that feature definitely is useful for determining response class labels. Of course, having lots of features with score "Inf" doesn't provide any discrimination among those features.

A different approach which will provide more discrimination among features is to train some models, and then use model interpretability techniques such as permutation importance or shapley importance to rank the features. The ranking will then depend on factors such as the type of model trained, and the type of model interpretability algorithm that is used. The R2024a Classification Learner app includes Shapley Importance plots.

If this answer helps you, please remember to accept the answer.