UCLA

This thesis provides a detailed comparative analysis of non-linear and linear modeling techniques in feature correlation analysis. In this work, I explore various testing methods on non-linear models, which are often challenging to interpret statistically. This challenge is particularly prevalent in the medical field where the complexity of models can obscure their interpretability. To address this issue, I introduce SHapley Additive exPlanations (SHAP) value as a means to enhance clarity and precision in the interpretation of results. Throughout this comparison, I will demonstrate that integrating SHAP value allows us to merge the interpretability of linear models with the accuracy of non-linear models, thus leveraging the strengths of both approaches. The difficulty in interpreting non-linear models stems from their complexity and the intricate relationships they encapsulate, which do not easily translate into direct, understandable insights as linear models do.