Adaptive Mutual Supervision Multi‑Task Graph Neural Network for Fine‑Grained Urban Traffic Demand Prediction

Urban traffic situation and group differences are widely studied for city governance and online services, yet most existing work focuses on coarse, grid‑level demand prediction and ignores heterogeneous travel behaviors of distinct population groups.

To address this gap, the authors propose a fine‑grained adaptive mutual‑supervision multi‑task graph neural network (AdaMSTNet). The model consists of two main components: (1) a spatio‑temporal neural network that combines graph convolution (with attention‑based edge weighting) and a GRU‑based temporal encoder, and (2) an adaptive soft‑group multi‑task learning module that treats each group‑region pair as an independent task, learns task grouping via a mask function, and employs mutual‑supervision signals from both population‑wise and region‑wise views to avoid negative transfer.

The architecture further incorporates hierarchical graph pooling to capture long‑range spatial dependencies, allowing distant but semantically similar regions (e.g., business districts) to share information.

Extensive experiments on real‑world Baidu datasets from Beijing and Shanghai (covering 25 demographic groups and 50 task streams) demonstrate that AdaMSTNet consistently outperforms strong baselines such as GRU, STGCN, GBDT, CoST‑Net, and ST‑ResNet, achieving more than 10% lower RMSE/MAE for 1‑3 hour forecasts. Ablation studies confirm the contributions of multi‑view graph construction, attention‑based spatial modeling, and adaptive soft grouping. Sensitivity analyses show the impact of input sequence length, group count, representation dimension, and loss weighting.

Practical case studies illustrate the model’s ability to capture peak travel patterns of specific groups (e.g., teenagers on Friday evenings, seniors at night), and the authors discuss deployment considerations for smart‑city platforms.

In summary, the paper defines a new task—population‑aware traffic demand prediction (CATDP)—and provides a comprehensive solution that integrates graph neural networks, hierarchical pooling, and adaptive multi‑task learning, achieving state‑of‑the‑art performance on large‑scale urban mobility data.