Efficient federated recommender system based on Slimify Module and Feature Sharpening Module

Download PDF
Open Webpage

Yicheng Di, Hongjian Shi, Jiansong Fan, Jiayu Bao, Gaoyuan Huang, Yuan Liu

Published: 01 Jan 2025, Last Modified: 12 Nov 2025Knowl. Inf. Syst. 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: With the development of communication networks and smart devices, research in the field of recommender systems has garnered widespread interest among researchers. The prediction of click-through rate (CTR) holds significant importance inside recommender systems. Many current CTR prediction approaches face the following challenges: first, they typically require centralized storage of user behavioral data to create accurate user models. However, users are highly sensitive to privacy, and privacy concerns and related data protection regulations make it difficult to aggregate behavioral data across platforms. Second, most CTR prediction models adopt large-scale deep models, whose substantial size hinders their broader application. To address these issues, we propose the efficient federated recommender system based on Slimify Module and Feature Sharpening Model (FedMS). The novelty of the proposed approach lies in its ability to effectively combine feature representations from several platforms in a manner that ensures privacy preservation, while also minimizing the communication overhead and storage requirements of the model. FedMS comprises two modules. The Slimify Module reduces the model’s size through inner product substitution and parameter reduction achieved by stacking compression layers. The Feature Sharpening Module enhances the importance of different field embeddings using position-based attention mechanisms. Furthermore, we introduce local differential privacy to further mitigate user privacy leakage. The empirical results obtained from conducting extensive experiments on three publicly available datasets provide evidence that FedMS has exceptional performance in predicting CTR, while also requiring a reduced amount of training time. Using RelaImpr as the evaluation metric, it achieves improvements of 11.04%, 3.38%, and 4.82% on the Criteo, Avazu, and MovieLens datasets, respectively.