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

Travel demand forecasting is crucial for city governance and online services, yet most existing studies focus on grid‑level predictions and ignore heterogeneous group differences. This paper addresses this gap by introducing a novel adaptive mutual‑supervision multi‑task graph neural network (AdaMSTNet) that can capture the relationships of different population groups across various spatio‑temporal scenarios.

1. Urban traffic situation and group differences – Traffic situation includes real‑time road conditions, flow volume, and resident travel demand. It exhibits strong periodicity, randomness, and spatio‑temporal dependency, making accurate prediction challenging. Different demographic groups (age, education, income) show distinct temporal and spatial travel patterns, motivating a fine‑grained, group‑aware forecasting approach.

2. Model architecture – The model consists of two components: (a) a spatio‑temporal neural network built on graph neural networks (GNN) and recurrent units (GRU) to capture spatial and temporal dependencies, and (b) an adaptive soft‑group multi‑task learning module that treats each group‑region pair as an independent task and learns task relationships through mutual supervision.

2.1 Graph Neural Network basics – GNN generalizes convolution to irregular graph structures, handling variable neighbor counts and non‑Euclidean distances. It enables semi‑supervised learning by propagating information along edges, with variants such as GraphSAGE, attention‑based GAT, and hierarchical pooling to improve expressiveness.

2.2 Spatial and population graphs – Two graph views are constructed: a spatial graph connecting city regions based on geographic distance, and a population graph linking demographic groups based on Pearson correlation of their time‑series flows. Edge selection balances computational cost and information richness.

2.3 Hierarchical spatio‑temporal GNN – A hierarchical GNN captures long‑range spatial correlations by pooling region nodes into abstract super‑nodes, allowing distant but similar areas (e.g., business districts) to share information while preserving local detail.

3. Multi‑task learning – Each group‑region pair is a task; tasks are softly grouped into k clusters learned by the model. Mutual supervision provides two self‑supervised signals (group‑level and region‑level) to guide the grouping and improve knowledge sharing while suppressing noise from unrelated groups.

4. Experiments – Real‑world datasets from Beijing and Shanghai (25 demographic groups × 2 flow types = 50 tasks) are used. AdaMSTNet achieves >10% improvement on 1‑3 hour forecasts over baselines such as GRU, STGCN, GBDT, CoST‑Net, and ST‑ResNet. Ablation studies show the contribution of multi‑view graphs, attention‑based spatial modeling, and adaptive soft‑grouping.

5. Sensitivity analysis – Performance improves with longer input sequences but saturates; both too coarse and too fine grouping granularity hurt accuracy; larger representation dimensions increase capacity but raise computation; higher loss coefficients aid grouping.

6. Application and conclusion – The model can be deployed in smart‑city platforms for fine‑grained travel demand profiling, resource allocation, and targeted services. The paper defines a new task (group‑aware travel demand prediction), introduces AdaMSTNet, hierarchical spatial modeling, and adaptive soft‑group multi‑task learning, and validates its effectiveness on large‑scale urban data.