GPFedRec: Graph-Guided Personalization for Federated Recommendation

📝 Paper Summary

Federated Recommendation Graph Learning Personalization

GPFedRec constructs a user-relation graph on the server using shared item embeddings to capture user correlations and guide personalized aggregation without accessing private interaction data.

Core Problem

Existing federated recommendation methods treat users as isolated individuals to preserve privacy, missing valuable user-user correlations that improve personalization, while graph-based methods typically require centralized data access.

Why it matters:

Users with similar interaction histories often share future preferences, a signal lost in standard federated averaging
Constructing user graphs usually violates the data locality principle of federated learning by requiring access to interaction histories or social networks
Standard FedRec models learn a shared global model that neglects diverse, personalized user preferences

Concrete Example: In a movie recommendation system, User A and User B both love niche Sci-Fi films. A standard FedRec approach averages their models with User C (who likes Romance), diluting the Sci-Fi signal. GPFedRec detects A and B have similar item embeddings, links them in a server-side graph, and aggregates their parameters together to reinforce the Sci-Fi preference.

Key Novelty

Graph-Guided Personalization for Federated Recommendation (GPFedRec)

Constructs a user relationship graph on the server by calculating similarity between clients' uploaded item embeddings, which serve as a privacy-preserving proxy for user preference
Uses a graph-guided aggregation mechanism (GCN-based) to generate user-specific item embeddings by aggregating information only from similar 'neighbor' clients
Combines user-specific embeddings with a global shared embedding to balance personalized and popular preference modeling

Architecture

The overall framework of GPFedRec, illustrating the interaction between clients and the server.

Breakthrough Assessment

7/10

Proposes a clever, privacy-compliant way to reintroduce user graphs into federated learning. The reliance on item embeddings as a proxy for user similarity is intuitive and theoretically sound.

⚙️ Technical Details

Problem Definition

Setting: Federated Recommendation with implicit feedback

Inputs: User set U, Item set I, User-Item interaction matrix Y (stored locally on clients)

Outputs: Predicted user preference scores for items

Pipeline Flow

Client: Local Training & Update → Upload Item Embeddings
Server: Similarity Calculation → Graph Construction → Graph-Guided Aggregation
Client: Receive User-Specific & Global Embeddings → Local Regularization

System Modules

Local Recommendation Model

Learn user and item representations from local interaction data

Model or implementation: Neural Collaborative Filtering style model (User Emb + Item Emb + Score Function)

Graph Constructor (Server-side Processing)

Build user-relation graph based on similarity of uploaded embeddings

Model or implementation: Cosine Similarity + Adaptive Thresholding

Graph-Guided Aggregator (Server-side Processing)

Generate user-specific item embeddings via neighbor aggregation

Model or implementation: Light GCN (1 layer)

Novel Architectural Elements

Server-side construction of user-relation graph solely from item embedding parameters (without interaction data)
Dual aggregation path: generating both a personalized user-specific embedding (via graph) and a global shared embedding (via weighted average)

Modeling

Base Model: General Neural Recommendation Model (User/Item Embeddings + Score Function)

Training Method: Federated Optimization with Server-side Graph Aggregation

Objective Functions:

Purpose: Optimize local recommendation accuracy on client device.

Formally: Binary Cross-Entropy Loss on implicit feedback interactions.
Purpose: Regularize local item embeddings to stay close to the server-aggregated personalized embeddings.

Formally: MSE loss between local item embedding q_i and server-provided user-specific embedding r_i.

Key Hyperparameters:

GCN layers: 1
Similarity metric: Cosine Similarity

Compute: Not reported in the provided text

Comparison to Prior Work

vs. PFedRec: GPFedRec explicitly models user-user correlations via a graph, whereas PFedRec treats users individually
vs. FedPerGNN: GPFedRec constructs the graph on the central server using only embeddings, avoiding the need for a third-party server to align historical interactions
vs. FedAvg [not cited in paper]: GPFedRec uses graph-guided weighted aggregation based on similarity, whereas FedAvg uses indiscriminate averaging

Limitations

Relies on the assumption that item embeddings alone are sufficient proxies for user preference similarity
Graph construction on the server has O(N^2) complexity with respect to the number of clients (though subgraph updates are proposed for scalability)
Potential privacy leakage from item embeddings, though mitigated by differential privacy (Laplacian noise)

Reproducibility

Code: https://github.com/Zhangcx19/GPFedRec

Code is publicly available at https://github.com/Zhangcx19/GPFedRec. The provided text is a truncated excerpt and does not contain the full experimental results tables.

📊 Experiments & Results

Evaluation Setup

Federated recommendation on benchmark datasets with implicit feedback

Benchmarks:

MovieLens-100K (Movie Recommendation)
MovieLens-1M (Movie Recommendation)
Hetrec-Moive (Movie Recommendation)
LastFM (Music Recommendation)
Hetrec-LastFM (Music Recommendation)

Metrics:

Not reported in the provided text
Statistical methodology: Not explicitly reported in the paper

Experiment Figures

Preliminary experiment comparing PFedRec performance with different graph types: Random Graph vs. Interaction-based Graph.

Main Takeaways

Qualitative result: Random graphs hurt performance compared to baselines, whereas informative graphs (like the one constructed by GPFedRec) improve performance.
Qualitative result: GPFedRec achieves comparable performance to an 'upper bound' graph constructed using actual user interaction data (which violates privacy), and significantly outperforms random graphs.
Qualitative result: The graph-guided aggregation mechanism can be used as a plugin to enhance other existing Federated Recommendation methods.
Robustness: Integrating differential privacy (Laplacian noise) allows GPFedRec to maintain steady performance while enhancing privacy.

📚 Prerequisite Knowledge

Prerequisites

Federated Learning (FedAvg)
Graph Convolutional Networks (GCN)
Implicit Feedback Recommendation
Differential Privacy

Key Terms

FedRec: Federated Recommendation—recommendation systems trained across decentralized devices to preserve privacy

Item Embedding: A vector representation of an item (e.g., a movie) learned by the model; in this paper, it is used as a proxy for user preference when shared

User-Relation Graph: A graph where nodes are users and edges represent similarity in preferences or behavior

Graph-Guided Aggregation: A mechanism where the server updates a client's model parameters by aggregating parameters from correlated 'neighbor' clients rather than the global average

GCN: Graph Convolution Network—a neural network architecture that aggregates information from a node's neighbors in a graph

Differential Privacy: A technique to maximize data utility while limiting the impact of any single individual's data on the output, often achieved by adding noise

BCE: Binary Cross-Entropy—a loss function used for binary classification tasks (e.g., interacted vs. not interacted)