Signal An Agentic Approach for Active Data Collection, Travel Behavior Modeling, and Weather-Sensitive Demand Prediction

Summary

Travel behavior research often develops digital data collection and predictive modeling as separate efforts that are evaluated independently of one another. This study instead proposes a three-agent workflow that ties together conversational data collection, structured data processing, and behavioral prediction. A chatbot-run, image-augmented stated-preference survey gathered mode-choice responses from student commuters across five predefined weather scenarios, producing 454 respondent-scenario observations in total. Weather-related patterns were analyzed with a multinomial logit model, while logistic regression and random forest served as machine-learning benchmarks for comparison. Nine locally run large language models, ranging from two to thirty-five billion parameters, were tested across four zero-shot conditions and then extended with persona, few-shot, and vision-based setups. Random forest reached 69.6 percent accuracy on the five-class task, the best text-only zero-shot model reached 69.9 percent without any task-specific tuning, and the top vision-based configuration, using the same weather images shown to respondents, reached 71.5 percent.