Enhancing Occupancy Detection with Explainable AI: Leveraging SHAP and Large Language Models

The rapid growth of the field of Artificial Intelligence has increased the utilization of predictive machine learning models in the construction industry. While their capabilities are significant, they often perform in a black-box manner, raising the need for interpretability of these models. This work focuses on the problem of occupancy detection and proposes a generalized explainability approach for common model configurations trained on public occupancy datasets. The data are collected from buildings with multiple sensor modalities, including power meters and environmental sensors. The proposed methodology involves the interpretation of the model predictions via the SHapley Additive exPlanations (SHAP) method and the employment of a Large Language Model (LLM) to provide the reasoning behind the model predictions and the feature contribution in natural language. This work provides generalized prompt instructions along with example responses on multiple scenarios, as well as a discussion on the limitations, common issues that were overcome with minimal prompt-tuning, and the potential for future development of the proposed method.