Interpretability analysis for thermal sensation machine learning models: An exploration based on the SHAP approach

Area of Science:

• Environmental Science
• Computer Science
• Human-Computer Interaction

Background:

• Machine learning models are prevalent in thermal sensation studies.
• Existing models often suffer from a lack of transparency due to their black-box nature.
• Current interpretability methods for thermal sensation models are often inadequate.

Purpose of the Study:

• To enhance the transparency of machine learning models used for thermal sensation analysis.
• To investigate feature importance and decision pathways within these models using a game theory approach.
• To provide a more robust understanding of how various factors influence thermal perception.

Main Methods:

• Applied SHapley Additive exPlanation (SHAP) for interpretability analysis of thermal sensation models.
• Examined feature effects on thermal sensations from both local and global perspectives.
• Analyzed correlations between features and thermal sensation outcomes.

Main Results:

• Identified that feature combinations (2-4 features) significantly determine thermal sensations, not solely individual feature magnitudes.
• Demonstrated that feature effects vary across samples, highlighting differentiation alongside magnitude.
• Revealed complex interactions between features influencing thermal perception.

Conclusions:

• The SHAP analysis provides crucial insights into the decision-making processes of thermal sensation models.
• Thermal sensation is influenced by dynamic, high-dimensional feature spaces rather than static conditions.
• This research contributes to developing more interpretable and reliable thermal comfort models.