Individualized Federated Learning for Traffic Prediction with Error Driven Aggregation

Low-latency traffic prediction is vital for smart city traffic management. Federated Learning has emerged as a promising technique for Traffic Prediction (FLTP), offering several advantages such as privacy preservation, reduced communication overhead, improved prediction accuracy, and enhanced adaptability to changing traffic conditions. However, majority of the current FLTP frameworks lack a real-time model updating scheme, which hinders their ability to continuously incorporate new incoming traffic data and adapt effectively to the changing dynamics of traffic trends. Another concern with the existing FLTP frameworks is their reliance on the conventional FL model aggregation method, which involves assigning an identical model (i.e., the global model) to all traffic monitoring devices to predict their individual local traffic trends, thereby neglecting the non-IID characteristics of traffic data collected in different locations. Building upon these findings and harnessing insights from reinforcement learning, we propose NeighborFL, an individualized real-time federated learning scheme that introduces a haversine distance-based and error-driven, personalized local models grouping heuristic from the perspective of each individual traffic node. This approach allows NeighborFL to create location-aware and tailored prediction models for each client while fostering collaborative learning. Simulations demonstrate the effectiveness of NeighborFL, offering improved real-time prediction accuracy over three baseline models, with one experimental setting showing a 16.9% reduction in MSE value compared to a naive FL setting.

💡 Research Summary

The paper introduces NeighborFL, a novel federated learning framework designed for real‑time traffic prediction in smart‑city environments. Existing federated traffic‑prediction (FLTP) approaches suffer from two major drawbacks: they typically train models offline and then broadcast a single global model to all devices, which fails to incorporate newly arriving traffic data and cannot handle the non‑IID nature of traffic streams collected at geographically dispersed sensors.
NeighborFL addresses these issues by (1) enabling continuous model updates at each communication round and (2) constructing personalized model aggregations for every client based on spatial proximity and recent prediction error. Each device maintains a “Favorite Neighbors” (FN) set, initially empty, that contains nearby devices (selected within a predefined haversine‑distance radius). At the start of a round, a device aggregates the parameters of its own local model together with those of the devices in its FN using FedAvg, producing a personalized model θ_i. Simultaneously, it creates an evaluation model θ_i⁺ by additionally incorporating the most recent local model of the nearest candidate neighbor that is not yet in FN. Both models generate predictions for the current time step; their mean‑squared errors (MSE) are compared against ground truth. If θ_i⁺ yields a lower error, the candidate neighbor is added to FN and θ_i⁺ becomes the new base for local learning; otherwise, the candidate is either discarded or placed on a retry list. Devices may also prune FN members when the set stabilizes or when prediction errors consistently rise. This error‑driven, distance‑aware grouping implements a reinforcement‑learning‑style exploration‑exploitation trade‑off within the federated setting.
The authors implement NeighborFL with LSTM networks as the client model and evaluate it on the public PEMS‑Bay dataset, which contains speed measurements from 26 traffic detectors. Two initialization scenarios are examined: (a) random weight initialization and (b) pre‑training each client on the first week of data. Over a two‑week test period, NeighborFL consistently outperforms three baselines, including a naïve FL that shares a single global model. In the random‑init case, the average MSE reduction is 12.3 %; with pre‑trained initial models, the best improvement reaches 16.9 % relative to the naïve FL baseline. The radius‑based candidate selection limits communication overhead, while overlapping FN sets enable knowledge propagation across the network, allowing distant nodes to benefit indirectly from updates made by intermediate devices.
Key contributions are:

1. A real‑time, client‑centric aggregation scheme that dynamically adapts to traffic drift.
2. A hybrid grouping heuristic that combines haversine distance with online error feedback to form personalized neighbor sets.
3. Demonstrated communication‑efficient operation through radius constraints and selective model exchange.
4. Empirical evidence that personalized aggregation yields significant MSE reductions on realistic traffic data.
   The paper also discusses limitations: the additional download and evaluation of candidate models increase computational and bandwidth costs; hyper‑parameters such as radius size and error thresholds may need careful tuning for different urban layouts; and the current study is limited to LSTM architectures, leaving open the question of applicability to newer models like Transformers or Graph Neural Networks. Future work is suggested in automatic hyper‑parameter adaptation, lightweight evaluation mechanisms, graph‑based neighbor relationship learning, multi‑modal data integration (e.g., weather, events), and field trials on actual smart‑city infrastructure.
   Overall, NeighborFL offers a compelling solution that merges privacy‑preserving federated learning with the adaptability required for real‑time traffic forecasting, paving the way for more responsive and efficient urban traffic management systems.