A Domain-Informed Deep Learning Framework for Trip-Purpose Prediction from Household Travel Survey Data
TL;DR
This study benchmarks domain-informed deep learning for 20-class trip-purpose prediction using the 2015 Southeast Michigan household travel survey using the 2015 Southeast Michigan (SEMCOG) household travel survey, underscoring the necessity of leakage-controlled evaluation, class-sensitive metrics, imbalance-aware training, and temporal-transferability testing when machine learning informs travel-demand analysis.
Abstract
Trip-purpose identification is fundamental to travel-demand modelling because it links observed movements to the behavioral motivations behind them. This study benchmarks domain-informed deep learning for 20-class trip-purpose prediction using the 2015 Southeast Michigan (SEMCOG) household travel survey (89,740 trips from 9940 households), under a leakage-controlled protocol: households are assigned in their entirety to training, validation, and test sets; preprocessing is fitted on training data only; and every model is retrained across 10 random seeds, with results reported as means with standard deviations. Six generic neural architectures, a parameter-matched single-stream ablation, and four classical benchmarks—including a multinomial logit (MNL) specification—are compared with TravelNet, a domain-informed network that encodes temporal, socio-economic, and mode-use characteristics in separate streams before fusion. TravelNet attains the highest accuracy (57.67 ± 0.51%) and macro F1 (0.264 ± 0.006), exceeding the parameter-matched ablation by 1.10 accuracy points (paired p = 0.002), which attributes a small but consistent gain to the domain-informed grouping itself. Macro-averaged metrics remain far below accuracy for every model, and eight low-frequency purposes show near-zero recall under standard training; inverse-frequency class weighting recovers recall for all eight (to 0.12–0.44) at a mean cost of 14.4 accuracy points with essentially unchanged macro F1, revealing a precision–recall redistribution rather than a free balanced-performance gain. Models transferred across a decade to the harmonized 2005 SEMCOG survey retain above-baseline but substantially degraded accuracy (37.0 ± 1.7% versus 61.1 ± 0.9% in-domain). The findings underscore the necessity of leakage-controlled evaluation, class-sensitive metrics, imbalance-aware training, and temporal-transferability testing when machine learning informs travel-demand analysis.