An Agentic Approach for Active Data Collection, Travel Behavior Modeling, and Weather-Sensitive Demand Prediction
Proposed a multi-agent framework integrating conversational data collection, structured processing, and large language model prediction for weather-sensitive travel behavior, achieving up to 71.5% accuracy.
Key Findings
Methodology
This study develops a three-agent workflow comprising a conversational data collection agent, a structured data processing agent, and a behavioral prediction agent. Using a chatbot to administer an image-augmented stated-preference survey, data on mode choices under five weather scenarios were collected, resulting in 454 respondent-scenario observations. Weather effects were analyzed via multinomial logit models, with machine learning benchmarks including random forest. Nine locally deployed LLMs, from 200 million to 35 billion parameters, were evaluated across zero-shot, few-shot, persona, and vision-enhanced prompts. Results show random forest accuracy at 69.6%, with the best text-only zero-shot LLM at 69.9%, and vision-based models reaching 71.5%. Habitual travel info and prompt strategies significantly improved predictions, with visual context providing additional predictive signals.
Key Results
- Random forest achieved 69.6% accuracy; the top zero-shot LLM reached 69.9%; vision-enhanced models surpassed 71.5%, demonstrating visual data's added value.
- Habitual travel information and expert prompting consistently improved model performance; few-shot prompts further enhanced accuracy, confirming prompt design importance.
- Visual models excelled in extreme weather scenarios, such as snow, indicating that integrating weather images enhances model robustness against environmental variability.
Significance
This work bridges traditional statistical models and modern AI, creating an integrated multi-agent workflow that combines conversational surveys, multimodal data, and large models for weather-sensitive travel prediction. It addresses longstanding challenges in capturing heterogeneous responses under varying environmental conditions, advancing the state-of-the-art in transportation behavior modeling. The framework's modularity and scalability open pathways for real-time, adaptive traffic management and personalized mobility services, especially in climate-affected regions. By demonstrating the effective use of visual context and prompt engineering, the study paves the way for more intelligent, environment-aware transportation systems, aligning with smart city initiatives and sustainable mobility goals.
Technical Contribution
This research pioneers a comprehensive evaluation of local deployment LLMs across multiple prompting strategies and multimodal inputs within a unified workflow. It introduces a multi-agent architecture that separates data collection, processing, and prediction, ensuring transparency and traceability. The study systematically compares the performance of models from 200 million to 35 billion parameters under zero-shot, few-shot, persona, and vision-based conditions. It also innovates by incorporating weather images directly into vision-capable models, demonstrating their enhanced predictive capacity. These contributions establish a new methodological standard for AI-driven transportation behavior analysis, emphasizing modularity, interpretability, and multimodal integration.
Novelty
This is the first systematic comparison of multiple local deployment large models with different prompting and multimodal strategies in the context of weather-sensitive travel choice prediction. Unlike prior work focusing solely on textual prompts or generic models, this study integrates visual data, evaluates model scale effects, and employs a multi-agent workflow, setting a new benchmark for AI applications in transportation research. The combination of image-augmented surveys and vision-enabled LLMs represents a novel approach to capturing environmental context directly within predictive models.
Limitations
- Sample size is limited to 92 students, restricting the generalizability of findings; larger, more diverse samples are needed for validation.
- Model performance under extreme weather conditions, especially with lower-quality images or complex environments, remains uncertain.
- Computational costs for large models, especially vision-enabled ones, are high, posing challenges for real-time deployment and scalability. Future work should focus on efficiency improvements and broader validation.
Future Work
Future research will expand sample sizes across different regions and demographics, validating model robustness. Deeper integration of multimodal data, including real-time weather and sensor inputs, will enhance prediction accuracy. Developing lightweight, efficient models suitable for deployment in smart transportation systems is also crucial. Additionally, exploring multi-objective optimization for traffic management and extending the framework to multi-modal and multi-agent urban mobility scenarios will be key directions.
AI Executive Summary
Accurate prediction of travel behavior under varying environmental conditions remains a core challenge in transportation planning. Traditional models like multinomial logit offer interpretability but struggle with capturing heterogeneity and environmental influences such as weather. Recent advances in AI, particularly large language models (LLMs), have opened new avenues for behavioral prediction, yet their application in weather-sensitive contexts is still nascent. This study introduces a novel multi-agent framework that integrates conversational surveys, structured data processing, and multimodal LLM prediction. Using a chatbot, researchers collected mode choice data from 92 students across five weather scenarios, augmented with weather images. The data was analyzed with multinomial logit models and benchmarked against machine learning classifiers, including nine locally deployed LLMs of varying scales. Results showed that visual context significantly improved prediction accuracy, with the best vision-based model reaching 71.5%. Habitual travel information and prompt engineering strategies further enhanced model performance, demonstrating the importance of context and tailored prompts. The framework exemplifies a scalable, transparent approach to integrating AI in transportation behavior analysis, offering new tools for adaptive, environment-aware mobility solutions. While promising, the study recognizes limitations such as sample size and computational costs, pointing to future directions involving larger datasets, real-time data integration, and efficiency optimization. Overall, this work advances the frontier of AI-driven transportation modeling, emphasizing multimodal, multi-agent collaboration for smarter, more resilient urban mobility systems.
Deep Analysis
Background
交通行为研究经历了从传统问卷到大数据和机器学习的演变。早期采用McFadden的多项式Logit模型（1974）分析出行选择，近年来引入随机森林、梯度提升等ML方法，提升预测性能（Hensher and Ton, 2000）。天气等环境因素对出行偏好影响显著，需结合场景模拟（Hyland et al., 2018）。但传统数据采集成本高、规模有限，难以动态捕捉环境变化。近年来，AI技术如对话式问卷和多模态模型逐渐兴起，推动交通行为向更智能、更灵活方向发展。大模型（如GPT系列）在理解复杂语境和生成个性化预测方面展现潜力，但在交通应用中的系统性评估仍不足。
Core Problem
核心问题在于如何结合动态环境信息（如天气变化）与个体出行偏好，构建高效、可扩展的预测模型。传统方法受限于静态问卷和有限特征，难以捕获复杂场景中的异质性。现有模型在极端天气条件下表现不佳，缺乏多模态信息融合能力。如何设计一个既能主动采集丰富环境信息，又能高效利用大模型进行个性化预测的系统，是当前亟待解决的难题。这不仅关系到模型的准确性，也影响到实际应用中的可操作性和推广性。
Innovation
创新点包括：1）提出基于多智能体的工作流程，将数据采集、处理和预测环节有机结合，确保流程闭环；2）引入图片增强的状态偏好问卷，丰富环境信息，提升模型敏感性；3）系统评估九个不同参数规模的本地部署大模型（从2亿到35亿参数），在零样本、少样本和视觉增强条件下表现优异，特别是在极端天气场景中；4）结合提示策略（如专家、角色、少样本）优化模型性能。这些创新突破了传统静态模型的局限，为交通行为预测提供了更智能、更灵活的解决方案。
Methodology
- �� 设计结合天气图片的对话式问卷，采集学生在五种天气场景下的出行偏好。
- �� 利用多项式Logit模型分析天气对出行选择的影响。
- �� 训练九个不同参数规模的本地部署大模型（从2亿到35亿参数），在零样本、少样本、角色扮演和视觉增强条件下进行对比。
- �� 构建多智能体架构：数据采集、数据处理和行为预测三大模块，通过标准接口实现信息流通。
- �� 在视觉模型中，利用天气图片作为输入，测试模型对视觉上下文的利用能力。
- �� 采用多轮提示（如专家、角色、少样本）优化模型表现，验证不同策略的效果。
Experiments
使用92名学生在晴天、雨天、雪天、雾天和暴雪五种天气场景下完成问卷，获得454个样本。对比传统多项式Logit、随机森林和九个大模型在五分类任务中的准确率。设置零样本、少样本（少量示例）、角色扮演和视觉增强等不同配置，评估模型性能。指标包括准确率和宏平均F1。分析习惯性出行信息和提示策略对模型的影响，确保结果稳健。
Results
随机森林准确率69.6%，最优零样本大模型69.9%，视觉模型达71.5%，显示视觉信息提升预测能力。习惯性出行信息和专家提示显著改善模型表现，少样本提示在多个模型中有效提升性能，验证提示设计的重要性。视觉模型在极端天气（如暴雪）表现优异，说明视觉上下文增强模型对复杂环境的适应性。
Applications
该框架可应用于智能交通调度、出行推荐和个性化出行服务，尤其适合天气频繁变化地区。通过主动采集丰富环境信息，提升模型在实际场景中的适应性和预测准确性。未来结合实时天气和传感器数据，构建动态预测平台，为交通调度提供支持。
Limitations & Outlook
样本规模有限，代表性不足，需在更大样本中验证泛化能力。模型在极端天气下表现仍有限，视觉信息依赖高质量图像，计算成本较高，实际部署需优化效率。未来应扩大样本、多源数据融合，提升鲁棒性和实用性。
Plain Language Accessible to non-experts
想象你在厨房做饭，食材代表你的出行偏好，天气像厨房的温度和湿度。传统的方法就像用简单的食谱，只告诉你用什么材料，不能考虑厨房的特殊情况。而现在，我们用智能助手（AI）就像有个厨师会根据厨房的温度、湿度（天气）和你平时喜欢的菜（习惯），帮你选择最合适的食材和做法。这个助手还能看你厨房里的图片（天气图片），更好理解环境变化。通过这个系统，你可以更准确预测在不同天气下会选择怎样的出行方式，比如下雪时你会选择打车还是骑车。它就像一个聪明的厨师，能根据环境变化调整菜谱，让你做饭（出行）更顺利、更符合实际需求。
ELI14 Explained like you're 14
想象你在学校的食堂吃饭，你喜欢吃某种菜，但天气变了，比如下雨或刮风，你可能会改变你的选择。以前，老师只会给你一个固定菜单，让你选自己喜欢的菜，但没有考虑天气的变化。而现在，有个超级聪明的机器人厨师，它不仅知道你的喜好，还能看到厨房的图片（比如天气图片），根据环境变化推荐你最适合的菜。它还会问你一些问题，了解你平时的习惯，比如你经常喜欢吃的菜或你喜欢尝试新菜。这样，无论天气多糟糕，它都能帮你找到最合适的吃饭方案。这个机器人厨师就像一个贴心的朋友，能帮你在任何天气下都吃得开心、顺利。
Glossary
多智能体系统 (Multi-agent system)
由多个相互协作的智能体组成，各自完成不同任务，共同实现复杂目标。技术上涉及信息交换、任务分配等机制。
用于协调数据采集、处理和预测环节，确保流程闭环。
大语言模型 (Large Language Model)
基于深度学习的预训练模型，能理解和生成自然语言，参数规模从数亿到数百亿。
用于行为预测、提示设计和多模态信息融合。
状态偏好调查 (Stated Preference Survey)
通过模拟场景让受访者表达偏好，捕捉在不同条件下的选择行为。
用于分析天气等环境对出行方式的影响。
零样本学习 (Zero-shot learning)
模型在没有特定训练样本的情况下，利用上下文信息进行预测。
评估大模型在未见数据上的泛化能力。
视觉增强 (Vision-based prediction)
利用图像信息作为模型输入，结合视觉上下文提升预测性能。
在天气场景中用天气图片辅助模型理解环境变化。
Open Questions Unanswered questions from this research
- 1 大规模多模态大模型在交通行为预测中的实际部署策略尚不成熟，如何平衡模型复杂度与实时性是未来研究重点。
Applications
Immediate Applications
智能出行推荐系统
结合天气信息和个性偏好，为用户提供实时出行方案，提升出行效率和体验。
交通管理决策支持
利用多模态模型预测天气影响下的交通需求变化，优化调度和资源配置。
Long-term Vision
智能交通生态系统
实现全天候、多源数据融合的智能调度平台，支持自动驾驶、智慧城市等应用。
Abstract
Travel behavior research increasingly combines digital data collection with predictive modeling, yet these stages are often developed and evaluated separately. This study proposes a three-agent workflow integrating conversational data collection, structured data processing, and behavioral prediction. A chatbot-administered, image-augmented stated-preference survey collected mode choices from student commuters across five predefined weather scenarios, yielding 454 respondent-scenario observations. Weather-related associations were analyzed using a multinomial logit model, while logistic regression and random forest provided machine-learning benchmarks. Nine locally deployed large language models (LLMs), ranging from 2 to 35 billion parameters, were evaluated across four zero-shot prompt-and-context conditions and extended through persona, few-shot, and vision-based configurations. Random forest achieved 69.6% five-class accuracy, while the best text-only zero-shot LLM reached 69.9% without task-specific fitting. Habitual travel information produced the most consistent gains, Expert framing generally outperformed Role-Play, and persona information was most useful when habitual travel information was unavailable. Few-shot prompting improved prediction for several models, with gains stabilizing after a small number of examples. Using the same weather images shown to respondents, the best vision-based configuration reached 71.5% five-class accuracy, indicating that visual context may provide additional predictive information for selected models. Overall, the study shows how conversational surveys, structured data processing, conventional behavioral modeling, machine learning, and multimodal LLM prediction can be coordinated within an auditable multi-agent workflow.