Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CrossTrafficLLM: A Human-Centric Framework for Interpretable Traffic Intelligence via Large Language Model

Zeming Du, Qitan Shao, Hongfei Liu, Yong Zhang

TL;DR

CrossTrafficLLM presents a unified Generative AI framework that jointly forecasts spatiotemporal traffic states and generates conditional abnormal event descriptions by aligning semantic textual context with the traffic network. It introduces a text-guided adaptive graph convolutional network and a TextCrossformer-based predictor, coupled with road-aware text generation mechanisms including road importance, cross-attention, and a road-text memory module. The approach achieves state-of-the-art performance on the BjTT dataset for both forecasting accuracy and text description quality, demonstrating the value of deep semantic alignment between unstructured text and structured traffic data for human-centric ITS decision support. The work highlights substantial improvements in interpretability and actionable insight, enabling more robust and explainable GenAI-augmented traffic intelligence with potential for real-world deployment and extensions to additional data modalities and interactive analysis.

Abstract

While accurate traffic forecasting is vital for Intelligent Transportation Systems (ITS), effectively communicating predicted conditions via natural language for human-centric decision support remains a challenge and is often handled separately. To address this, we propose CrossTrafficLLM, a novel GenAI-driven framework that simultaneously predicts future spatiotemporal traffic states and generates corresponding natural language descriptions, specifically targeting conditional abnormal event summaries. We tackle the core challenge of aligning quantitative traffic data with qualitative textual semantics by leveraging Large Language Models (LLMs) within a unified architecture. This design allows generative textual context to improve prediction accuracy while ensuring generated reports are directly informed by the forecast. Technically, a text-guided adaptive graph convolutional network is employed to effectively merge high-level semantic information with the traffic network structure. Evaluated on the BJTT dataset, CrossTrafficLLM demonstrably surpasses state-of-the-art methods in both traffic forecasting performance and text generation quality. By unifying prediction and description generation, CrossTrafficLLM delivers a more interpretable, and actionable approach to generative traffic intelligence, offering significant advantages for modern ITS applications.

CrossTrafficLLM: A Human-Centric Framework for Interpretable Traffic Intelligence via Large Language Model

TL;DR

CrossTrafficLLM presents a unified Generative AI framework that jointly forecasts spatiotemporal traffic states and generates conditional abnormal event descriptions by aligning semantic textual context with the traffic network. It introduces a text-guided adaptive graph convolutional network and a TextCrossformer-based predictor, coupled with road-aware text generation mechanisms including road importance, cross-attention, and a road-text memory module. The approach achieves state-of-the-art performance on the BjTT dataset for both forecasting accuracy and text description quality, demonstrating the value of deep semantic alignment between unstructured text and structured traffic data for human-centric ITS decision support. The work highlights substantial improvements in interpretability and actionable insight, enabling more robust and explainable GenAI-augmented traffic intelligence with potential for real-world deployment and extensions to additional data modalities and interactive analysis.

Abstract

While accurate traffic forecasting is vital for Intelligent Transportation Systems (ITS), effectively communicating predicted conditions via natural language for human-centric decision support remains a challenge and is often handled separately. To address this, we propose CrossTrafficLLM, a novel GenAI-driven framework that simultaneously predicts future spatiotemporal traffic states and generates corresponding natural language descriptions, specifically targeting conditional abnormal event summaries. We tackle the core challenge of aligning quantitative traffic data with qualitative textual semantics by leveraging Large Language Models (LLMs) within a unified architecture. This design allows generative textual context to improve prediction accuracy while ensuring generated reports are directly informed by the forecast. Technically, a text-guided adaptive graph convolutional network is employed to effectively merge high-level semantic information with the traffic network structure. Evaluated on the BJTT dataset, CrossTrafficLLM demonstrably surpasses state-of-the-art methods in both traffic forecasting performance and text generation quality. By unifying prediction and description generation, CrossTrafficLLM delivers a more interpretable, and actionable approach to generative traffic intelligence, offering significant advantages for modern ITS applications.
Paper Structure (39 sections, 9 equations, 3 figures, 3 tables)

This paper contains 39 sections, 9 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Traffic Forecasting: From Statistical Models to Multi-modal Integration
  4. Large Language Models for Traffic Prediction and Description
  5. Methodology
  6. Problem Formulation
  7. Traffic Prediction
  8. TextCrossformer Architecture
  9. Text Encoding and Feature Integration
  10. Text-Guided Adaptive Graph Convolutional Network
  11. Location-Aware Text Encoding
  12. Sparse Cross-Modal Alignment
  13. Text-Conditional Feature Modulation
  14. Adaptive Graph Convolution
  15. Variable Definitions
  16. ...and 24 more sections

Figures (3)

  • Figure 1: An illustration of the benefits provided by this approach includes: 1) The model can learn from abnormal event texts to better recognize and anticipate non-routine traffic situations, thereby improving the accuracy of traffic forecasting under abnormal conditions. 2) The system is able to generate corresponding abnormal event descriptions, offering clear and intuitive explanations for complex or unexpected traffic states.
  • Figure 2: Method overview. The proposed model comprises two key components: Traffic Prediction and Text Generation. The Traffic Prediction component integrates a Crossformer architecture, Large Language Model (LLM), and adaptive graph convolution. Textual features extracted by the LLM guide the adaptive matrix in processing traffic data before feeding into the Crossformer for prediction. The Text Generation component processes traffic data through Cross-attention and an Abnormal Road Detector layer before fine-tuning a large language model to generate descriptive text output.
  • Figure 3: An example of text generated by DDCap, DiffCap, BLIP, Transformer and CLIPCap. The red portions highlight content that does not match the ground truth. There is a high consistency between the text generated by our CrossTrafficLLM and the ground truth text, with only a a few event in this example deviating.