TLDR: A new federated learning framework, GFed-PP, allows recommendation systems to offer personalized suggestions while respecting diverse user privacy preferences. It distinguishes between public and private users, leveraging shared data from public users to enhance recommendations through a user relationship graph, while keeping private user data strictly local. Experiments show it significantly improves accuracy over existing methods and can be enhanced with differential privacy for stronger protection.
The field of recommendation systems, which helps us discover new movies, music, or products, has long faced a significant challenge: how to provide personalized suggestions without compromising user privacy. Traditional recommendation models often rely on collecting and centralizing vast amounts of user data, which raises concerns about personal data security. In response to these concerns and regulations like GDPR, a new approach called Federated Learning has emerged.
Federated Learning allows recommendation models to be trained directly on users’ devices (clients) while a central server coordinates the training process by aggregating model parameters, rather than raw data. This means sensitive user data never leaves the device, significantly enhancing privacy.
However, existing federated recommendation systems (FedRecs) often make a simplifying assumption: that all users have the same privacy preferences, meaning no data is uploaded to the server. This overlooks a crucial opportunity to improve recommendation quality by utilizing data from users who are willing to share some information. In reality, users often have different privacy needs; some may prefer to keep all their data private, while others might be comfortable sharing certain interaction data.
Addressing this, a new framework called Graph Federated Learning for Personalized Privacy Recommendation (GFed-PP) has been proposed. This innovative system is designed to adapt to these varying privacy requirements while simultaneously boosting recommendation performance. GFed-PP introduces a distinction between “private users,” whose data remains entirely local, and “public users,” who permit their interaction data to be shared.
The core idea behind GFed-PP is to leverage the interaction data from public users to build a “user relationship graph” on the central server. This graph helps understand similarities between users based on their shared interests. A lightweight Graph Convolutional Network (GCN) is then used to learn personalized item embeddings for each user. These embeddings are essentially numerical representations of items that capture individual user preferences.
To maintain privacy, user embeddings and the scoring function (which predicts a user’s rating for an item) are kept strictly local on each client’s device. Only the item embeddings, which are considered less privacy-sensitive, are communicated to the server. The framework further refines this by providing public users with personalized item embeddings, while private users receive a global item embedding that reflects popular preferences. This dual approach ensures that both personalization and privacy are effectively managed.
The GFed-PP framework operates in several steps during each training round. Public users upload their interaction graphs to the server. All clients, both public and private, train their local recommendation models. Public users initialize their local item embeddings with personalized embeddings, while private users use global embeddings. After local training, clients upload their updated item embeddings to the server. The server then uses the public user data to construct and aggregate the user relationship graph, generating new user-specific and global item embeddings. Finally, these updated embeddings are distributed back to the respective public and private users for the next round of local training.
Extensive experiments conducted on five different datasets, including MovieLens and Lastfm, have shown that GFed-PP significantly outperforms existing centralized and federated recommendation methods. For instance, on the Amazon-Video dataset, it achieved substantial improvements in recommendation accuracy. An ablation study confirmed that each component of GFed-PP – Client Item Embedding Initialization, User Graph Construction, and User-Personalized Item Embedding – plays a crucial role in its overall effectiveness.
The research also explored the impact of various factors, such as the size of embedding, the number of GCN layers, and the ratio of public users. It was found that an embedding size of 32 and a single GCN layer generally yielded optimal performance. Furthermore, increasing the proportion of public users generally improved overall performance, demonstrating the benefits of shared information.
To further strengthen privacy, the researchers incorporated a Local Differential Privacy (LDP) mechanism, which adds a small amount of noise to the item embeddings before they are uploaded. While this slightly reduces accuracy, it provides a stronger privacy guarantee, allowing a balance between performance and privacy protection.
Also Read:
- Advancing Private AI: A New Framework for Neural Fields on Edge Devices
- New Research Shows AI Models Vulnerable to Privacy Attacks Even with Partial Data
In conclusion, GFed-PP offers a practical and effective solution for federated recommendation systems that can cater to diverse user privacy preferences. By intelligently leveraging publicly available data while strictly protecting private information, it enhances recommendation accuracy without compromising user trust. This framework represents a significant step forward in building more privacy-aware and personalized recommendation services. You can find more details about this research in the full paper available at arXiv.org.
