TLDR: A new lightweight federated learning approach improves traffic forecasting by incorporating spatial relationships between traffic sensors. It uses graph-based aggregation methods, GraphFedAvg and MPFedAvg, to weight model updates based on connectivity, outperforming traditional methods while remaining computationally efficient.
Traffic prediction is crucial for managing urban mobility, helping to estimate traffic speed or flow in specific areas using data from sensors. Imagine each traffic sensor or road segment as an individual client, collecting its own local traffic data. This setup makes Federated Learning (FL) a highly suitable approach, as it allows these clients to collaboratively train a model without directly sharing their sensitive raw data, thus preserving privacy.
In a typical FL system, a central server gathers model updates from various clients, aggregates them, and then sends an updated shared model back to each client. However, standard FL methods, like Federated Averaging (FedAvg), often assume that clients operate independently. This assumption can limit their effectiveness in real-world scenarios like traffic prediction, where spatial relationships between clients (e.g., how one road segment connects to another) are incredibly important.
While some advanced Federated Graph Learning methods exist to capture these spatial dependencies, they often come with a significant computational cost. This paper introduces a novel, lightweight graph-aware FL approach that combines the simplicity of FedAvg with key principles from graph learning. Instead of training complex, full graph models, this new method applies basic neighborhood aggregation to guide how model parameters are updated. It intelligently weights client models based on their connectivity within the road network, effectively capturing spatial relationships while remaining computationally efficient.
The researchers propose two main aggregation methods:
Graph Neighbourhood-Aware Averaging (GraphFedAvg)
This method extends traditional FedAvg by incorporating the structure of the communication graph. It essentially means that each client’s model parameters are averaged not just uniformly, but with a weighting that considers its direct neighbors in the network (including itself). This allows information to propagate through immediate connections, making the aggregation spatially aware.
Also Read:
- Geo-ORBIT: Advancing Roadway Digital Twins with Privacy-Preserving Lane Detection
- Optimizing Federated Learning for Real-World Data and Device Variations
Graph Message Passing-Aware Averaging (MPFedAvg)
Inspired by algorithms like Label Propagation, MPFedAvg iteratively refines client parameters. It blends a client’s own model values with those of its neighbors through the network, using a normalized adjacency matrix. A hyperparameter controls the balance between retaining original local parameters and incorporating neighborhood information. This approach enables parameters to spread across the network, allowing clients to benefit from multi-hop neighbors while still maintaining a strong influence from their own initial models.
Both GraphFedAvg and MPFedAvg are designed to be context-sensitive, enhancing accuracy while keeping data local and private. They offer a structured alternative to simple averaging by leveraging local connectivity, which is ideal for networked settings like traffic sensors where client behavior is influenced by their surroundings. The ability for multi-hop message passing further facilitates broader information exchange, leading to better performance and faster convergence.
The effectiveness of these methods was evaluated on two well-known traffic datasets: METR-LA (Los Angeles traffic speed readings) and PEMS-BAY (Bay Area traffic speed readings). The results demonstrated that the proposed graph-based aggregation methods are highly competitive and consistently outperform standard baseline approaches, including traditional FedAvg and more complex graph-based federated learning techniques. Notably, even single-layer variants of their methods often matched or exceeded the performance of two-layer variants, suggesting that a simpler, more computationally attractive design is sufficient to capture essential spatial dependencies.
In conclusion, this research presents a practical and efficient way to improve traffic forecasting in federated learning environments. By integrating graph-aware aggregation into the server-side process, the methods effectively preserve inter-client dependencies without the heavy computational overhead typically associated with training deep Graph Neural Networks. This offers a scalable and computationally attractive alternative for real-world traffic systems. For more details, you can read the full research paper here.
