A machine-learning approach for evaluating occupants’ indoor environment satisfaction in high-rise mixed-use buildings

Purpose This paper presents a pilot study of a machine learning (ML) approach to predict occupants' satisfaction with the indoor environment in high-rise mixed-use buildings, aiming to validate a proof of concept for integrating ML models into early-stage design tools to support occupant-centred performance evaluation. Design/methodology/approach Using post-occupancy evaluation data from a case study building, Random Forest and Neural Network models were trained to classify satisfaction levels–Dissatisfied, Neutral or Satisfied–for both residents and workers based on indoor environmental factors. The methodology focuses on addressing class imbalance through data resampling and cost-sensitive learning, with model performance assessed using class-specific metrics. Findings Both models achieved high overall accuracy (cross-validation score >0.80), with notable improved performance in identifying minority classes after balancing methods were employed. While limited to a single case study, future data collection across diverse buildings and occupant profiles has the potential to improve performance and enable generalisability. Originality/value This research demonstrates the feasibility of a scalable framework for predicting indoor environmental satisfaction, enabling the integration of ML models into simulation-based workflows for data-driven, occupant-centric design evaluation. It advances the field by (1) classifying satisfaction into three actionable categories while explicitly addressing class imbalance, (2) operationalising POE data to move beyond retrospective reporting and (3) establishing a proof of concept for embedding ML models into early-stage design tools.