Protocol to calculate and compare exact Shapley values for different kernels in support vector machine models using...

Area of Science:

• Computational chemistry
• Machine learning
• Artificial intelligence

Background:

• Machine learning models, particularly support vector machines (SVMs), are increasingly used in scientific research.
• Explaining the predictions of these complex models is crucial for trust and validation.
• The Shapley value formalism offers a theoretically sound method for attributing contributions to model predictions.

Purpose of the Study:

• To present a detailed protocol for calculating exact Shapley values for SVM models.
• To enable comparison of Shapley values across different kernels and binary input features.
• To provide a framework for analyzing and interpreting feature importance in SVM predictions.

Main Methods:

• Adaptation of the Shapley value formalism from cooperative game theory.
• Development of a protocol for calculating exact Shapley values for SVMs.
• Utilization of customizable Python scripts for data preparation and value computation.
• Implementation of correlation analysis and feature mapping for results interpretation.

Main Results:

• A reproducible protocol for computing exact Shapley values for SVM models was established.
• The protocol facilitates the comparison of feature importance across various SVM configurations.
• Methods for analyzing and visualizing Shapley value results were detailed.

Conclusions:

• The presented protocol provides a robust method for enhancing the interpretability of SVM models.
• This approach allows for a deeper understanding of feature contributions to model predictions.
• The Shapley value formalism, when applied to SVMs, offers valuable insights into AI model behavior.