Improving the explainability of Random Forest classifier - user centered approach.

Area of Science:

• Computational Biology
• Machine Learning in Medicine
• Data Science

Background:

• Machine Learning (ML) methods are increasingly vital in healthcare but face adoption barriers due to their complexity and lack of explainability.
• Understanding ML model decisions is crucial for validation and trust, especially for non-expert users in clinical settings.

Purpose of the Study:

• To enhance the explainability of Random Forest (RF) classifiers using a novel method called RFEX.
• To develop user-friendly, interpretable summary reports for trained RF models to improve adoption and validation.

Main Methods:

• RFEX was developed using a user-centered approach, incorporating feedback from practitioners on explainability requirements.
• The method was implemented and tested on the Stanford FEATURE dataset, using RF to predict functional sites in 3D molecules.
• Formal usability testing was conducted with 13 expert and non-expert users to assess RFEX's utility.

Main Results:

• RFEX significantly increased the explainability and user confidence in RF classification for the FEATURE dataset.
• Analysis revealed that a small number of top-ranked features (2-6) are sufficient for achieving over 90% accuracy, even when using all 480 features.

Conclusions:

• RFEX provides a valuable tool for demystifying complex ML models, particularly RF classifiers, for a broader range of users.
• The findings highlight the potential for simplified feature sets to achieve high accuracy, aiding in the interpretation and application of ML in molecular biology and medicine.