Exploring Deep Learning and Behavioural Segmentation for Predicting Activity Schedules

Activity-based models (ABMs) are essential for travel demand forecasting, but traditional rule-based and econometric approaches for activity scheduling are often complex, rigid and difficult to scale. This paper presents a data-driven framework to predict 24-h human activity schedules using deep learning and behavioural segmentation. A novel, cluster-informed bidirectional long short-term memory network is proposed, which learns directly from daily activity sequences represented as 96 discrete 15-min time steps. The model’s key innovations include the integration of unsupervised k-prototypes clustering to create a latent behavioural signature for each user and a progressive minority training strategy with focal loss to address severe class imbalance. Trained on data from the Halifax Travel Activity Survey, the model’s effectiveness is validated through extensive ablation studies. Results show that the cluster-informed, bidirectional architecture significantly outperforms alternatives. While achieving high time step-level accuracy (macro F1: 59.9%), the model’s primary strength lies in generating functionally plausible schedules. A novel grouped sequence alignment method distance metric reveals that many prediction errors are semantically minor substitutions within functionally similar activity categories (e.g., ‘home chore’ for ‘staying at home’). The framework successfully replicates aggregate temporal patterns for critical activities such as work and school, making it a robust and practical tool for transportation planners to simulate policy impacts with greater behavioural realism.