TelTrans: Applying Multi-Type Telecom Data to Transportation Evaluation and Prediction via Multifaceted Graph Modeling

TL;DR

This paper addresses the limitations of sensor-based traffic prediction by introducing Geographical Cellular Traffic (GCT) as a scalable mobility proxy, with a focus on vehicle-related flows (V-GCT). It proposes MultiFaceted Graph Modeling (MFGM), a three-facet architecture—Multivariate, Temporal, and Spatial—built on Channel-Specific Graph Attention Layers (CGATL) to integrate multi-type GCT flows and capture inter-type, temporal, and bidirectional spatial dependencies. Through extensive experiments on a 21-road-segment Taiwan dataset, MFGM achieves superior accuracy for V-GCT forecasting, particularly for long horizons, and ablation studies confirm the importance of each facet for capturing regional functionality and mobility dynamics. The work demonstrates practical potential for real-time transportation management and supports the deployment of telecom data into city-scale traffic systems.

Abstract

To address the limitations of traffic prediction from location-bound detectors, we present Geographical Cellular Traffic (GCT) flow, a novel data source that leverages the extensive coverage of cellular traffic to capture mobility patterns. Our extensive analysis validates its potential for transportation. Focusing on vehicle-related GCT flow prediction, we propose a graph neural network that integrates multivariate, temporal, and spatial facets for improved accuracy. Experiments reveal our model's superiority over baselines, especially in long-term predictions. We also highlight the potential for GCT flow integration into transportation systems.

Figures (7)

• Figure 1: An overview of GCT flow. Cellular traffic, generated by mobile users on various road segments, corresponds to entries in the GCT database. The accumulated count of GCT over a given time interval is defined as GCT flow. Over time, the flow series reflects road traffic patterns.
• Figure 2: Selected road segments in Hsinchu, situated near functional regions or congested routes (e.g., the commuting route from segments 62 to 56), capture representative GCT flow patterns. Segment color indicates the average V-GCT flow from August 28 to September 28, 2022.
• Figure 3: Pearson Correlation Coefficients were calculated between pairs of road segments based on 1-hour V-GCT flow at 2022/9/25. As mobility patterns change over time, areas with high coefficients (i.e., blue boxes) shift among different segments, unveiling the evolving spatial correlations.
• Figure 4: Despite distinct GCT flow patterns in commuting routes and residential zones, Pearson correlation coefficients with 1-day data between V-GCT (V) and P-GCT (P) or S-GCT (S) are similar across both segments: 0.78 (P) and 0.75 (S) vs. 0.77 (P) and 0.74 (S). However, subtracted flows (V-P and V-S) yield more noticeable differences: 0.98(V-P) and 0.99(V-S) versus 0.89(V-P) and 0.86(V-S). This suggests that using subtracted flows can reveal more nuanced insights than directly exploring correlations among GCT flows.
• Figure 5: System integration of GCT flow enables real-time crowd monitoring and threshold alerts. Exceeding city-defined limits triggers traffic optimization strategies.