Trip Purpose Prediction with Minimal Sequential Context: A Parsimonious Machine Learning Approach

ABSTRACT The prediction of trip purpose has received growing attention in recent years. However, many existing approaches either rely on complex sequence models that require fully labelled daily trip chains or focus primarily on destination‐based spatial attributes, limiting their practical applicability and interpretability. This study proposes a machine learning framework for trip purpose prediction that incorporates minimal sequential context, defined as the immediately preceding and subsequent trips, while spatial composition indicators serve primarily as supplementary contextual signals. Using Seoul as a case study, this study integrates data from the 2021 Household Travel Survey (HTS), public transportation OD data, and point of interest (POI) data from Kakao Map to construct a multi‐class classification model implemented with XGBoost and LightGBM. The proposed model achieves approximately 71% overall accuracy and a weighted ROC‐AUC of approximately 0.95 across eleven trip purpose categories. Ablation results indicate that minimal sequential context accounts for most of the performance improvement, while POI‐based spatial indicators provide only modest and supplementary incremental gains relative to sequential features. Overall, the findings suggest that substantial predictive power can be obtained without modelling full daily trip chains. By leveraging limited temporal continuity within an interpretable machine learning framework, the proposed approach provides a scalable and deployable solution for trip purpose prediction in data‐constrained urban environments, demonstrating that localised temporal structure captures the dominant behavioural signal without requiring full‐day trip chains.