TLDR: FedGDVE is a new federated learning framework for e-commerce recommender systems that addresses the limitations of sparse local data by selectively incorporating relevant global data. It uses a Graph Data Value Estimator (GDVE) with a graph encoder, valid predictor, and a reinforcement learning-based probability estimator to filter and sample the most beneficial global interactions. This approach significantly improves recommendation performance (up to 34.86% gain) and scalability while preserving data privacy, outperforming existing federated learning methods across various datasets.
In the rapidly evolving world of e-commerce, personalized recommendations are crucial for enhancing user experience and driving sales. However, the increasing concerns around data privacy have made it challenging for online stores to leverage vast amounts of user interaction data. Federated Learning (FL) has emerged as a promising solution, allowing recommendation models to be trained across multiple online stores without centralizing sensitive user data. Each store trains a model locally and shares only model updates, preserving privacy.
Despite its advantages, existing FL-based recommender systems often face a significant hurdle: the limited size and inherent biases of each client’s (individual online store’s) private dataset. This limitation can severely impact the quality of recommendations, as models struggle to learn comprehensive user preferences from sparse local data. While publicly available global datasets could potentially enrich local training, simply combining all global interactions can introduce noise, irrelevant patterns, and increase computational costs, often leading to worse personalization.
Addressing these challenges, a new framework called FedGDVE (Federated Graph Data Value Estimator) has been proposed. This innovative approach allows a large shopping platform to collaborate with numerous small online stores to build a global recommender system. The platform holds global data, such as shareable user and item lists, while each store maintains its private interaction data. FedGDVE’s core idea is to selectively augment each client’s local data with semantically aligned samples from the global dataset, rather than indiscriminately merging all available global information.
The FedGDVE framework employs a sophisticated mechanism to achieve this selective data utilization. It consists of three main modules within each client: a graph encoder, a valid predictor, and a probability estimator. The graph encoder is pre-trained on the global dataset to extract structural features from global user-item interactions. The valid predictor, on the other hand, is pre-trained on the client’s local dataset to assess the client-specific relevance of global interactions. Finally, the probability estimator combines the insights from both the graph encoder and the valid predictor to compute a selection probability for each global interaction. Based on these probabilities, only the most pertinent global interactions are sampled and used to enhance the client’s local dataset.
A key innovation in FedGDVE is its use of a reinforcement learning scheme to optimize the probability estimator. Since the sampling process is non-differentiable, a task predictor is introduced. This predictor is trained on the selected subset of global interactions, and its performance on the client’s local data serves as a reward for the probability estimator. This iterative process ensures that the probability estimator learns to prioritize global interactions that genuinely improve the local model’s performance, making the data selection process highly effective and adaptive.
The benefits of FedGDVE are substantial. By carefully filtering global data, it mitigates the limitations of sparse and biased local datasets without overwhelming the local models with irrelevant information. This selective augmentation leads to more accurate and personalized recommendations. Experiments conducted on recognized benchmarks, including MovieLens-100k, Gowalla, and Yelp2018 datasets, demonstrate that FedGDVE significantly outperforms recent FL recommender systems. It achieves performance gains of up to 34.86% in precision, 14.25% in recall, and 34% in NDCG on the MovieLens-100k dataset compared to other federated learning methods. These improvements highlight FedGDVE’s robustness and scalability across various e-commerce scenarios, even when dealing with diverse and heterogeneous datasets.
Furthermore, FedGDVE addresses the computational challenges associated with large global datasets. By operating on a compact, relevance-filtered subset of global data, it remains cost-effective and scalable, regardless of the global dataset’s size. This makes it a practical solution for real-world e-commerce platforms where resources might be constrained.
Also Read:
- Enhancing Private In-Context Learning Through Public Information
- Advancing User Behavior Simulation with Persona-Driven Small Language Model Fine-Tuning
In conclusion, FedGDVE represents a significant advancement in federated recommender systems. By intelligently leveraging global data while strictly preserving local data privacy, it offers a powerful framework for e-commerce platforms to deliver highly personalized and accurate recommendations. For more in-depth technical details, you can refer to the full research paper here.
