ADPFedGNN: Adaptive Decoupling Personalized Federated Graph Neural Network

1 Introduction

Graph data finds widespread applications in various domains, such as social networks [Quan et al., 2023], financial transactions [Pareja et al., 2020], and recommendation systems [Yu et al., 2022]. Representing data as graphs, with entities as nodes and relationships as edges, better reflects real-world scenarios [Zhang et al., 2024a]. Analyzing these graphs allows leveraging node relationships to generate accurate representations, which can significantly enhance subsequent node classification tasks [Khoshraftar and An, 2024]. However, in real-world applications, high-quality data is typically owned by governments, enterprises, or other organizations. Due to privacy concerns, regulatory restrictions, and conflicting interests, data sharing is typically restricted [Wang et al., 2020], thereby limiting the effectiveness of graph-based node classification models.

∗Corresponding author

Federated Graph Neural Network (FGNN) enable the training of effective models using multi-party graph data while keeping data localized [Fu et al., 2022; Li et al., 2024b]. Due to the attribute shifts in data across clients, FGNN face the challenge of Non-IID data [Wan et al., 2024], which disrupts the performance of federated learning. To address this, personalized federated learning approaches allow clients to adopt differentiated aggregation strategies [Dhillon et al., 2020; Long et al., 2023a; Zhang et al., 2024b], primarily through techniques such as gradient weighting [He et al., 2021b; Zhang et al., 2023], regularization [Li et al., 2020; Li et al., 2021b], and client sampling [Fraboni et al., 2021; Long et al., 2023b]. In FGNN, each client holds knowledge. During federated training, this knowledge is encoded into the model parameters, comprising global knowledge applicable across clients and client-specific local knowledge relevant only to the individual client. Existing methods typically couple global knowledge-related parameters with local knowledge-related parameters, which can lead to interference from local knowledge on global knowledge-related parameters during the federated learning process.

For graph data, the Non-IID nature introduces structural shift issues. Some approaches mitigate these shifts by sharing graph data information [Zhang et al., 2021a; Huang et al., 2023]. However, graph node neighbors may carry global or local knowledge, and directly incorporating them into federated training without distinguishing their suitability for training global or local knowledge-related parameters can introduce noise, ultimately interfering with the training process and degrading the performance of FGNN [Tang et al., 2021; Li et al., 2021a]. Determining which neighbors should contribute to training parameters related to global or local knowledge is crucial. Nevertheless, there is a lack of studies that effectively address how to select and assign neighbors to these roles in FGNN.

Based on the above, we identify two key challenges for PFGNN:

• Challenge 1: How to decouple global knowledge from client-specific local knowledge parameters during the training process, ensuring that only parameters related to global knowledge are included in federated aggregation, while effectively handling attribute shifts in feature distributions across clients.
• Challenge 2: How to effectively select neighbors that contribute to the training of either global or local knowledge-related parameters, filtering out irrelevant nodes that do not contribute to model performance, especially under structural shifts where neighbor relationships vary across clients.

To address these two challenges, we propose Adaptive Decoupling Personalized Federated Graph Neural Network (ADPFedGNN), which leverages multi-party graph data to train models capable of classifying nodes in each client's graph data. For Challenge 1, we propose federated maskbased parameter decoupling method to separate local and global knowledge-related parameters, and personalized federated masked parameter aggregation method to prevent interference between these parameters during federated aggregation. For Challenge 2, we propose reinforcement federated adaptive neighbor selection strategy, which adaptively selects node neighbors for federated training. ADPFedGNN effectively prevents interference from local knowledge-related parameters on other clients' models, while efficiently leveraging multi-party data to train personalized federated Graph Neural Network. Extensive experiments conducted on three public datasets validate the effectiveness of the proposed method.

Our main contributions include:

• We propose a federated mask based parameter decoupling method that utilizes an automatically updated mask mechanism and mutual information minimization to decouple model parameters into global and local components. The mask is updated based on the client's local training gradients and their similarity to the global model, ensuring effective decoupling.
• We propose a personalized federated masked parameter aggregation method that aggregates the masked model parameters from each client to form the global model. The global parameters are then updated using the inverted local masks, preventing the global model from disrupting the local models' adaptability to clientspecific data.
• We propose a reinforcement federated adaptive neighbor selection strategy that uses reinforcement learning to select suitable node neighbors for training global or local knowledge related parameters while filtering out irrelevant nodes. This approach enhances the performance of PFGNN.

2 RelatedWork

2.1 Federated Learning

Federated learning is a distributed paradigm that addresses data silos [Liu et al., 2024; Li et al., 2024a]. For graph data, FedGraphNN [He et al., 2021a] is a federated learning benchmark system for GNN, supporting multi-domain datasets. SpreadGNN [He et al., 2021b] enables federated multi-task learning, and FGGP [Wan et al., 2024] enhances generalization and classification by introducing cluster prototypes and global knowledge contrast.

However, data across clients often exhibits non-IID characteristics, making it difficult to train a single global model. Personalized federated learning methods, such as FedProx [Dhillon et al., 2020], introduce regularization to address data heterogeneity, while FedSEM [Long et al., 2023b] clusters clients based on model parameters to improve personalization. MOON [Li et al., 2021b] improves performance through model-level contrastive learning. FedSage+[Zhang et al., 2021a] improves federated node classification performance by generating missing neighbors. FedALA [Zhang et al., 2023] adapts the aggregation process to local data. Despite these advancements, existing methods still couple global and local knowledge parameters, limiting their effectiveness in preventing interference from local knowledge.

2.2 Graph Sampling

As graph data scales, training on the entire graph becomes computationally and memory-intensive, making minibatch mechanisms essential for large datasets [Hamilton et al., 2017]. Traditional methods, such as random sampling, ignore noisy or irrelevant links between nodes [Zhao et al., 2023; Wang et al., 2021]. Recent work has focused on optimizing neighbor sampling. BanditSampler [Liu et al., 2020] and Thanos [Zhang et al., 2021b] reduce sampling variance through multi-armed bandit formulations and novel reward functions. However, these approaches require per-node updates, making them difficult to adapt to federated learning due to the lack of shared strategies across devices.

Parameterized neighbor selection strategies have recently gained attention. Bayesian GNN [Hasanzadeh et al., 2020], DSKReG [Wang et al., 2021], and Learnable Sampling [Zhao et al., 2023] dynamically adjust sampling probabilities, but they rely on complete computation flows, limiting their applicability in federated learning where global gradient computation is infeasible. Reinforcement learning offers a promising solution by learning parameterized strategies without relying on a complete computation flow [Lai et al., 2020; Sun et al., 2021; Yang et al., 2020], making it ideal for federated settings.

3 Methodology

We propose ADPFedGNN to address challenges in adaptive neighbor selection and parameter coupling for personalized federated graph node classification. ADPFedGNN enables collaborative training on multi-party graph data to classify nodes in each client's local graph while ensuring data privacy.

The method consists of three components: (1) Reinforcement federated adaptive neighbor selection strategy using reinforcement learning to select suitable neighbors for training global or local knowledge-related parameters, while filtering irrelevant nodes; (2) Federated mask-based parameter decoupling method that uses trainable masks and mutual information minimization to decouple global and local knowledgerelated parameters; and (3) Personalized federated masked parameter aggregation method that shares only global parameters, preserving local model personalization. An overview of the method is shown in Figure 1.

3.1 ProblemDefinition

In federated graph node classification, each client Cm holds private graph data Gm = (Vm, Em, Xm) and node labels Ym, where Vm, Em, and Xm denote the node set, edge set, and node feature matrix. Due to privacy constraints and non-IID characteristics, clients cannot share raw data, leading to attribute and structural shifts that complicate federated learning [Wan et al., 2024]. Attribute shift arises when the distribution of node features conditioned on labels varies across clients, i.e., Pm(h | y) ̸= Pn(h | y), even if the marginal label distribution P(y) remains similar. Structural shift occurs when the relationship between edge and label distributions differs across clients, i.e., Pm(E, Y ) ̸= Pn(E, Y ), despite consistent label distributions Pm(y) = Pn(y). These shifts hinder the training of a globally generalized model by introducing inconsistencies across clients, which degrade model performance. To address these challenges, we propose a personalized federated learning approach that effectively handles data heterogeneity and distribution shifts to achieve accurate node classification while preserving data privacy.

3.2 Reinforcement Federated Adaptive Neighbor Selection Strategy

We propose a reinforcement Federated Adaptive Neighbor Selection Strategy, which formulates the neighbor sampling as a Markov Decision Process (MDP) and uses a Deep QNetwork (DQN) to learn an adaptive policy.

At each training round, we adopt a minibatch mechanism to sample a two-layer block of neighbors. For each node in the batch, k neighbors are selected in each layer, forming an initial two-layer neighbor set. The DQN then estimates the expected reward for each neighbor, refining the selection by identifying suitable neighbors for updating global or local knowledge-related parameters and filtering out irrelevant or redundant ones.

State Design: The state sijb is formed by concatenating the feature vector hi of the target node vi, the feature vector hj of the candidate neighbor vj, and a binary indicator b. The binary indicator b = 0 denotes the selection for updating global parameters, while b = 1 denotes the selection for updating local parameters. The state representation enables the agent to distinguish between tasks for global and local parameter updates. The state is represented as:

Action Design: The action aijb ∈ {0, 1} determines whether the candidate neighbor vj is selected for the target node vi. Actions are generated by the DQN network, which outputs the state-action value Q(sijb, aijb; θq). Action selection is based on the Q-value distribution:

By selecting actions with higher Q-values, the DQN network identifies suitable node neighbors for the task while filtering out irrelevant nodes.

The reward r(sijb, aijb) is defined as the difference between the federated task loss Lrtask at round r and the average task loss from previous rounds:

Experience Replay Mechanism: To stabilize training, we use an experience replay mechanism, storing state-actionreward-next state tuples in a buffer D. Minibatches are randomly sampled from D to update the DQN, helping the agent learn adaptive neighbor selection for better federated graph node classification.

3.3 FederatedMask-Based Parameter Decoupling Method

We propose a federated mask-based parameter decoupling method. This approach uses a GNN with automatic update masks to construct a global model suitable for federated aggregation and a local model retained only on the client side.

Global and local features are extracted using GNN with separate masks to ensure feature-level decoupling. The masks are updated after each federated training round, with global masks, Mfeat global and Mcls global , preserve global knowledgerelated parameters, applicable across clients, while the local masks, Mfeat local and Mcls local , retain local knowledge-related parameters. The global mask assigns a value of 1 to the top q% of parameters most similar to the local training gradients and global non-zero gradients. Similarly, the local mask assigns a value of 1 to the top q% of parameters with the highest nonzero gradients from local training. This mechanism ensures that during the update, both the global and local masks focus on the most influential parameters for each respective task.

The global feature extraction is defined as:

where N(i, ai,t,0) denotes the set of neighbors selected for global parameter updates when b = 0. When b = 1, the set of selected local neighbors for updating local parameters is used for local feature extraction, defined as:

where N(i, ai,t,1) refers to the set of neighbors selected for local parameter updates.

To decouple global and local features, we minimize their mutual information using the Contrastive Log-ratio Upper Bound (CLUB) [Cheng et al., 2020], which estimates the upper bound of mutual information Lmi and reduces it to achieve feature independence:

where µi and log σ2 i are the mean and log-variance estimated from global features hglobal i using neural networks:

where θµ and θlog σ2 represent the network parameters. These parameters are optimized through maximum likelihood estimation with the log-likelihood loss:

To classify nodes, the global and local features are passed through masked classifiers, focusing on their respective information types. The global classifier output is defined as:

where zglobal i is the global classifier output for node i, fcls represents the classifier function, θcls denotes the classifier parameters, andMcls global is the global classifier mask. Similarly, the local classifier output is defined as:

where zlocal i is the local classifier output for node i.

The final classifier output is obtained by fusing the global and local logits:

where zi is the final output for node i, and β ∈ [0, 1] balances the contributions of global and local classifiers. This weighted fusion ensures that the model can effectively leverage both globally shared knowledge and locally specialized patterns, enabling accurate node classification across diverse client data distributions.

The task is graph node classification, and the task loss is computed using cross-entropy:

where ˆ yi,c denotes the predicted probability of node i belonging to class c, yi,c is the one-hot encoded ground truth label, and V represents the set of nodes.

The total loss integrates the classification loss, mutual information loss, and an L2 regularization loss on the model parameters to prevent overfitting:

where λmi and λreg are hyperparameters balancing the classification, mutual information, and regularization terms.

3.4 Personalized FederatedMasked Parameter Aggregation Method

To achieve efficient aggregation of the global model while preserving the personalization of local models, we propose a personalized federated masked parameter aggregation method. In each federated training round, clients use two automatically updated masks, global masks Mglobal and local masks Mlocal, to decouple their model parameters into global and local components. These masks are applied uniformly to both the feature extraction layers and the classifier parameters, ensuring that the global model focuses on generalizable knowledge while the local model preserves clientspecific characteristics.

The aggregation goal is to compute a weighted average of global model parameters based on local data sizes and the global masks from selected clients S. Let θi denote the local model parameters of client i, andMglobal i represent its global mask. The server updates the global model parameters θglobal using:

After aggregation, the global model is distributed to clients for local updates. For client i, the local model parameters θi are updated using:

where θglobal [p] is the global model parameter at position p, θi[p] is the local model parameter of client i at position p, and Mlocal i [p] is the local mask value. The complementary mask 1 - Mlocal i [p] ensures that global parameters are applied only at positions where the local mask value is zero. The training process is detailed in Algorithm 1.

Algorithm 1 Training process of ADPFedGNN for a single epoch

• 1: Input: Local data G, batch size B.
• 2: Output: θµ, θlogσ2, θgnn, θcls, θq.
• 3: Initialize parameters θgnn, θcls, θµ, θlogσ2,Mglobal,Mlocal.
• 4: Client-Side Training:
• 5: for batch b ← 0 to B -1 do
• 6: Compute actions with Equation (2) to filter random neighbors and obtain N(i, ai,t,0) and N(i, ai,t,1);
• 7: Extract global and local features using Equations (4) and (5);
• 8: Estimate µ and logσ2 using Equation (7);
• 9: Compute classification task loss using Equation (12);
• 10: Update parameters θgnn, θcls;
• 11: Update parameters θµ, θlogσ2 using the log-likelihood loss in Equation (8);
• 12: Calculate reward using Equation (3);
• 13: Save experience (st, at, s′t , a′t ,R(st, at)) in replay buffer;
• 14: Train Q-network θq on sampled mini-batches from the replay buffer;
• 15: end for
• 16: Server-Side Aggregation:
• 17: Perform federated aggregation for θµ, θlog σ2 , θgnn, θcls, θq through Equations (14) and (15).

4 Experimental Analysis

4.1 Datasets

We conduct experiments on three public graph datasets: Cora [Sen et al., 2008], CiteSeer [Sen et al., 2008], and PubMed [Namata et al., 2012]. To simulate federated learning with varying levels of data heterogeneity, we partition the datasets using two methods: Louvain community partitioning [Peng et al., 2022], which assigns nodes to clients based on community structures, and Dirichlet label partitioning [Zhang et al., 2021a], which simulates non-IID client distributions using the Dirichlet distribution, where the parameter α controls the heterogeneity-smaller values result in more imbalanced client data distributions.

4.2 Baseline Methods

We compare ADPFedGNN with the following baselines: FedAvg [McMahan et al., 2017], Clustered Sampling [Fraboni et al., 2021], FedProx [Dhillon et al., 2020], MOON [Li et al., 2021b], FedALA [Zhang et al., 2023], FedSEM [Long et al., 2023a], FGGP [Wan et al., 2024], and FedSage+ [Zhang et al., 2021a].

4.3 Experimental Setup

We use GraphSage [Hamilton et al., 2017] and GAT [Veliˇckovi´c et al., 2017] as the backbone networks. GAT experiments are limited to ADPFedGNN, FedAvg, and FedProx, as GraphSage delivers the best overall performance, while GAT's sensitivity to neighbor information makes it more suitable for evaluating neighbor selection effectiveness. The DQN component for reinforcement learning is implemented with two hidden layers, each of size 128. A fixed client selection ratio of 0.25 is applied throughout the experiments. Model performance is evaluated using accuracy (ACC), macro-F1 score (F1), and macro-recall (Recall). These metrics offer a comprehensive evaluation of both overall and per-class performance in node classification. The reported results for ACC, F1, and Recall are presented as percentages, i.e., all values are multiplied by 100 for clarity and better comparability. For detailed experimental settings, please refer to the Appendix.

4.4 Comparative Experiment Analysis

We conduct extensive experiments on the Cora, CiteSeer, and PubMed datasets, partitioning clients into 5 groups using the Louvain community method. The results, shown in Table 1, indicate that ADPFedGNN surpasses the second-best method, FedALA, with average improvements of 5.66 percent, 5.83 percent, and 12.45 percent in ACC, F1, and Recall, respectively, for the GraphSage backbone. Significant improvements are also observed with the GAT backbone.

These experimental results demonstrate the effectiveness of ADPFedGNN. The reinforcement federated adaptive neighbor selection strategy enhances robustness and stability by adaptively selecting relevant neighbors for training global or local knowledge-related parameters, addressing Challenge 1. The federated mask-based parameter decoupling method effectively separates global and local parameters, and the personalized federated masked parameter aggregation ensures the efficient use of global knowledge while minimizing local interference, addressing Challenge 2. Together, these innovations enable ADPFedGNN to consistently outperform baseline methods across diverse datasets and configurations.

4.5 Ablation Study

We evaluate the contributions of key components in ADPFedGNN through ablation studies on the Cora dataset with 5 clients, using GraphSage as backbone models. The results, shown in Figure 2, compare the full ADPFedGNN model with three ablated variants: ADPFedGNN-d, which integrates the federated mask-based parameter decoupling and personalized federated masked parameter aggregation methods-designed to function together as the latter depends on the former; ADPFedGNN-r, which includes the reinforcement federated adaptive neighbor selection strategy; and ADPFedGNN-n, a baseline without these components.

The results demonstrate that the full ADPFedGNN model achieves the best performance across all metrics, confirming the necessity of integrating all proposed components. Specifically, ADPFedGNN-r enhances performance by adaptively selecting neighbors for training relevant parameters, while ADPFedGNN-d effectively decouples global and local parameters through adaptive masking and mutual information minimization, ensuring the efficient utilization of global knowledge and the preservation of client-specific characteristics.

4.6 The Impact of α onModel Performance

We evaluate the impact of data non-IID characteristics on the Cora dataset with 5 clients using Dirichlet label partitioning, varying the α parameter from 0.1 to 0.9 to simulate different levels of data heterogeneity. As shown in Figure 3, as α increases, the data distribution shifts from highly imbalanced to more balanced, leading to improved performance across most methods. ADPFedGNN consistently achieves superior results under all α settings and demonstrates remarkable effectiveness in scenarios with extreme imbalance at low α values, highlighting its robustness and adaptability to non-IID data.

4.7 Effect of Client Select Rate onModel Effectiveness

We evaluate the impact of client participation rates on the Cora dataset with 5 clients using Louvain community partitioning under select rates of 0.2, 0.5, 0.7, and 0.9. As shown in Table 2, ADPFedGNN consistently outperforms all baseline methods across select rates. With the GAT backbone, ADPFedGNN demonstrates significant improvements at lower select rates, highlighting the effectiveness of its reinforcement adaptive neighbor selection strategy. Similarly, with the GraphSage backbone, it achieves superior accuracy, macro-F1, and recall, demonstrating that the federated maskbased parameter decoupling method ensures effective utilization of global knowledge even with limited client participation. These results confirm the adaptability and robustness of ADPFedGNN across varying participation levels.

4.8 Impact of Neighbor Sampling Size onModel Performance

We evaluate the effect of neighbor sampling size k on model performance using the Cora dataset with 5 clients. As shown in Figure 4, increasing k enhances accuracy, macro-F1, and recall by providing richer contextual information. However, performance improvements slow when k exceeds 10, suggesting that a moderate sampling size is sufficient for effective learning. Larger k also increases memory consumption, posing challenges in resource-constrained environments. Notably, GAT benefits more from larger k values, as the expanded neighbor set allows for better attention mechanism learning, leading to faster performance gains compared to GraphSage. The reinforcement adaptive neighbor selection strategy effectively identifies suitable neighbors for updating global or local knowledge-related parameters while filtering out noisy neighbors.

4.9 Effect of β onModel Performance

We assess the impact of the hyperparameter β on model performance using the Cora dataset with 5 clients. As shown in Figure 5, performance peaks around β = 0.5, indicating that a balanced integration of global and local features effectively leverages both knowledge sources. Performance declines when β approaches extremes; an overemphasis on global features (β ≈ 0.9) limits local knowledge utilization, while excessive focus on local features (β ≈ 0.1) hinders federated knowledge sharing. However, within a reasonable range, performance remains stable, demonstrating the robustness of ADPFedGNN across various settings.

4.10 Sensitivity Analysis of λmi and λreg

We conduct sensitivity analysis experiments on the Cora dataset with 5 clients to evaluate the impact of the mutual information loss weight λmi and the regularization loss weight λreg. The experiments are performed with λmi values ranging from 0.1 to 0.9 and λreg values from 0.001 to 0.009. The results indicate that the optimal performance is achieved when λmi = 0.3 and λreg = 0.003. Figure 6 presents the results of the sensitivity analysis. The analysis shows that λmi results in some fluctuations across the three evaluation metrics; however, the median results remain competitive and within an acceptable range, demonstrating the effectiveness of the model. In contrast, λreg exhibits relatively smaller fluctuations across the three metrics, indicating that the model remains stable under different values of this parameter.

5 Conclusion

In this paper, we propose Adaptive Decoupling Personalized Federated Graph Neural Networks (ADPFedGNN) to enhance personalized federated graph node classification. ADPFedGNN effectively addresses the challenges of the neighbor selection and parameter coupling in non-IID federated environments. By employing a reinforcement adaptive neighbor selection strategy, it selects suitable node neighbors for training global or local knowledge-related parameters while filtering out irrelevant nodes, thereby enhancing model performance. Additionally, the federated mask-based parameter decoupling method separates global and local parameters, while the personalized federated masked parameter aggregation method enables effective parameter sharing without interfering with local feature learning. Experimental results on public datasets demonstrate that ADPFedGNN surpasses existing methods, achieving average improvements of 5.66 percent, 5.83 percent, and 12.45 percent in ACC, F1, and Recall, respectively.

Acknowledgements

This work was supported by the 8th Young Elite Scientists Sponsorship Program by CAST (2022QNRC001), National Natural Science Foundation of China (62172056, 62192784, U22B2038, U23A20319).