Informatics Seminar Recordings
Feature Importance Distributions
On the Discovery of Feature Importance Distributions: An Overlooked Area
Detecting feature importance (predictive power) is a key problem in Machine Learning. Previous methods have been focusing on providing a single value as the estimation of the importance. However, the meaning of such value is not always obvious. Moreover, in reality a feature's importance may vary dramatically across the feature's values. A point estimation of the importance cannot capture such variations. We propose a new definition of feature importance, which directly measures a feature's predictive power. We also propose an approach to detect a high-resolution distribution of a feature's importance across the feature's values. The key novelty is a feature importance model that allows identifying significant change of importance between adjacent feature values, and a cost function that permits separating the importance of different features. Empirical results on real-world medical datasets (Breast Cancer, Parkinson's, and Drug Consumption) show that, the proposed work could help discover better knowledge, build better models, and make better decisions.
Link to recording: