A graph attention-based framework for explainable thermal comfort prediction using infrared facial landmarks

Individualized thermal comfort prediction is critical for intelligent building control, yet existing methods are constrained by the invasiveness of wearable sensors and privacy and robustness limitations of visible-light imaging. We propose a non-invasive and interpretable framework based on thermal infrared facial keypoints and a multi-head attention graph neural network, operating solely on grayscale thermal infrared images. Facial landmarks are used to extract localized skin temperature, forming node features that include skin temperature, ambient temperature differentials, and spatial coordinates, integrated with global descriptors such as personal attributes, environmental conditions, and head pose. These heterogeneous features are structured into an anatomically informed graph and processed using attention mechanisms to model facial spatial relationships, followed by feature aggregation for thermal sensation classification. The proposed method achieves a mixed-sample four-class accuracy of 0.97. In the more stringent leave-one-subject-out validation, which better reflects real world generalization to unseen individuals, the model attains an accuracy of 0.69, outperforming the Predicted Mean Vote model (0.43), XGBoost (0.61), and a fully connected graph neural network baseline (0.64). Ablation studies demonstrate consistent gains from temperature-difference features, anatomical constraints, and multi-head attention mechanisms. Importance of facial keypoint temperatures is derived from attention weights and validated with generalized ordered logistic regression, revealing consistent patterns: nasal dominance under frontal views; with head pose deviation, ipsilateral cheek and jaw regions become more important, while periocular and chin regions remain stable contributors. The model identifies key facial thermal regions across pose variations, supporting real-time personalized thermal comfort prediction and environmental control in intelligent buildings.