Communication-Efficient Personalized Federal Graph Learning via Low-Rank Decomposition

TL;DR

The proposed CEFGL method (with GIN as the base model) improves accuracy by 5.64\% on cross-datasets setting CHEM, reduces communication bits by a factor of 18.58, and reduces the communication time by a factor of 1.65.

Abstract

Federated graph learning (FGL) has gained significant attention for enabling heterogeneous clients to process their private graph data locally while interacting with a centralized server, thus maintaining privacy. However, graph data on clients are typically non-IID, posing a challenge for a single model to perform well across all clients. Another major bottleneck of FGL is the high cost of communication. To address these challenges, we propose a communication-efficient personalized federated graph learning algorithm, CEFGL. Our method decomposes the model parameters into low-rank generic and sparse private models. We employ a dual-channel encoder to learn sparse local knowledge in a personalized manner and low-rank global knowledge in a shared manner. Additionally, we perform multiple local stochastic gradient descent iterations between communication phases and integrate efficient compression techniques into the algorithm. The advantage of CEFGL lies in its ability to capture common and individual knowledge more precisely. By utilizing low-rank and sparse parameters along with compression techniques, CEFGL significantly reduces communication complexity. Extensive experiments demonstrate that our method achieves optimal classification accuracy in a variety of heterogeneous environments across sixteen datasets. Specifically, compared to the state-of-the-art method FedStar, the proposed method (with GIN as the base model) improves accuracy by 5.64\% on cross-datasets setting CHEM, reduces communication bits by a factor of 18.58, and reduces the communication time by a factor of 1.65.

Figures (7)

• Figure 1: Illustration of the CEFGL framework. The process begins with the client transforming the input feature data into vectors of consistent dimensions using a multilayer perceptron (MLP). The client then performs local training to obtain the initial model. During this phase, the client trains the model with its local data while keeping the model parameters $\bm{W}_i^t$ frozen. Simultaneously, the sparsification components, which aim to reduce the complexity of the model, are fine-tuned. After local training, the client compresses the trained model using low-rank approximation techniques to reduce its size and upload it to the server. On the server side, the local gradients and the global model are updated based on the local model information received from the client. Finally, the server compresses the updated global model using low-rank methods and sends it back to the client. This process ensures efficient communication and model updates, promoting collaboration across distributed clients while minimizing communication overhead.
• Figure 2: Variation of test accuracy and loss with the number of communication rounds for different sparse ratios $\beta$.
• Figure 3: Variation of test accuracy and loss with the number of communication rounds for different communication probabilities $p$.
• Figure 4: Test accuracy and loss of CEFGL after employing quantization compression $Q_r(\cdot)$. The number of quantization bits $r$ is set to $r \in \{4,8,16,32\}$.
• Figure 5: (a) Accuracy on CHEM when clients accidentally drop out with different probabilities $\rho$. (b) The distributions' corresponding probability density functions (PDF).