In this paper, we perform the first study of heterogeneity-aware TFP within the FL framework, proposing a unified and effective framework named Fed4TFP. It solves the data heterogeneity challenges of TFP in FL and offers general federated extension capability for existing centralized TFP models. To address the traffic feature heterogeneity, Fed4TFP employs multi-dimensional positive samples contrastive learning for clustering to achieve personalized FL across diverse clients. To overcome the time coverage heterogeneity, Fed4TFP introduces a time window-based federated training, sequentially training client models over different time windows and learning missing information with varying time coverage. To tackle the data quality heterogeneity, Fed4TFP incorporates a dual-driven method, employing global-defense and local-denoising approaches to improve federated client data quality. Experiments on three benchmark datasets demonstrate that Fed4TFP significantly outperforms state-of-the-art baselines, e.g., 5.47%–38.48% improvement.
FederatedScope.version = 0.3.0
Python.version = 3.9.18
Cuda.version = 11.3
Xlrd.version = 2.0.1
All the experiments are conducted in the federated environment on seven nodes, one as a server and the other six as clients, each equipped with two Intel(R) Xeon(R) CPU E5-2650 v4@2.20GHz 12-core processors, 128GB of RAM, and an internet speed of 100MB/s.
The datasets used can be downloaded in the paper.
The running example of Fed4TFP is as follows:
python main.py --cfg federatedscope/contrib/configs/pems04.yaml