Community-Aware Multi-Task Transportation Demand Prediction

Hao Liu, Qiyu Wu, Fuzhen Zhuang, Xinjiang Lu, Dejing Dou, Hui Xiong

Introduction

Traditional urban transportation demand prediction models typically forecast the total number of people entering or leaving a specific grid cell in a city. While useful, this "aggregate" view ignores the fact that different groups of people have vastly different mobility patterns. For instance, high school students and vocational students might have highly synchronized schedules, while university students or retirees follow different periodicities.

This challenge is identified as Community-Aware Transportation Demand Prediction (CATDP). The core difficulty lies in the "diversified yet correlated" nature of the data. Predicting demand for a specific community in a specific region is a "sparse" problem compared to aggregate demand. Furthermore, these communities influence each other across different spatial and temporal contexts, making it difficult to decide which "tasks" (region-community pairs) should share information without introducing noise.

Methodology: Ada-MSTNet

To address these challenges, the authors propose the Multi-task Spatio-Temporal Network with Mutually-supervised Adaptive task grouping (Ada-MSTNet). The architecture, illustrated in Figure 1, consists of three main pillars:

  1. Multi-View Graph Construction: The model builds two types of graphs. A Region View connects areas based on physical adjacency, while a Community View connects different demographic groups based on the similarity of their historical demand patterns (calculated via Pearson Correlation Coefficient).
  2. Spatio-Temporal Backbone: To capture the "where" and "when," the model uses Graph Neural Networks (GNNs) to model spatial/structural dependencies and Gated Recurrent Units (GRUs) to handle temporal sequences.
  3. Mutually-supervised Adaptive Task Grouping: This is the core innovation. Instead of treating every region-community pair as an independent task or forcing them all to share the same parameters, the model "softly" clusters tasks into groups. Crucially, it uses a mutual supervision strategy: the hidden states from the region view help guide the task grouping in the community view, and vice versa. This prevents "negative transfer"—where unrelated tasks interfere with each other's learning.
The architecture of Ada-MSTNet, featuring multi-view graph construction and adaptive task grouping.

Experimental Setup

The model was evaluated on two massive, real-world datasets from Beijing and Shanghai, collected in 2019. Each dataset contains demand for 25 communities (defined by attributes like age, gender, education, and income) across 784 regions (1km x 1km grids). The Beijing dataset alone contains over 125 million queries from nearly 10 million users, resulting in 50 distinct prediction tasks (inflow/outflow) per region.

Key Results and Discussion

Ada-MSTNet was compared against seven baselines, including classical methods (Historical Average, XGBoost) and deep learning models (ST-ResNet, STGCN). Ada-MSTNet consistently outperformed all baselines, achieving a 15.77% to 22.78% improvement in RMSE (Root Mean Square Error) on the Beijing dataset.

Performance and Ablation

As shown in Figure 2, the performance gains are robust across different time steps (τ). The ablation study highlights several key insights:

  • Multi-View Importance: Combining both region and community views (rcview) significantly reduces error compared to using either alone.
  • Adaptive Grouping: The ada-group strategy consistently outperforms no-group and pre-group (static grouping) approaches, validating the effectiveness of the mutually-supervised clustering.
  • Long-term Robustness: Performance gains were even more pronounced for longer-term predictions (3 hours ahead vs. 1 hour ahead), suggesting the model captures underlying periodicities better than its peers.
Ablation study results showing the impact of different views, model components, and grouping strategies on RMSE.

Conclusion

Ada-MSTNet demonstrates that incorporating community-level identity and behavioral similarity is crucial for fine-grained transportation demand forecasting. By leveraging multi-view graphs and adaptive task grouping, the model effectively manages the sparsity and complex correlations inherent in community-aware data. Future work may explore latent community discovery to reduce reliance on explicit demographic labels, further enhancing the model's applicability in privacy-sensitive environments.

Summary updated

Abstract

Transportation demand prediction is of great importance to urban governance and has become an essential function in many online applications. While many efforts have been made for regional transportation demand prediction, predicting the diversified transportation demand for different communities (e.g., the aged, the juveniles) remains an unexplored problem. However, this task is challenging because of the joint influence of spatio-temporal correlation among regions and implicit correlation among different communities. To this end, in this paper, we propose the Multi-task Spatio-Temporal Network with Mutually-supervised Adaptive task grouping (Ada-MSTNet) for community-aware transportation demand prediction. Specifically, we first construct a sequence of multi-view graphs from both spatial and community perspectives, and devise a spatio-temporal neural network to simultaneously capture the sophisticated correlations between regions and communities, respectively. Then, we propose an adaptively clustered multi-task learning module, where the prediction of each region-community specific transportation demand is regarded as distinct task. Moreover, a mutually supervised adaptive task grouping strategy is introduced to softly cluster each task into different task groups, by leveraging the supervision signal from one another graph view. In such a way, Ada-MSTNet is not only able to share common knowledge among highly related communities and regions, but also shield the noise from unrelated tasks in an end-to-end fashion. Finally, extensive experiments on two real-world datasets demonstrate the effectiveness of our approach compared with seven baselines.
BibTeX
@inproceedings{liu2021community,
  author = {Hao Liu and Qiyu Wu and Fuzhen Zhuang and Xinjiang Lu and Dejing Dou and Hui Xiong},
  title = {{Community-Aware Multi-Task Transportation Demand Prediction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i1.16107},
}
Figure