Hypernetwork-Driven Trustworthy Recommendation

In the context of rapidly advancing electronic technologies, a growing number of consumer electronic products have increasingly incorporated recommendation systems to improve overall user experience. Conventional recommendation approaches primarily rely on deep neural models to estimate user preferences for items, yet such methods often require extensive user data sharing, which may undermine user trust in the system outputs. To address this issue, federated learning has been introduced into recommendation frameworks to enable privacy-preserving and reliable recommendation processes. However, existing federated recommendation models typically depend on repeated access to user–item interaction data across clients to learn shared model parameters, which limits their practicality. To overcome these limitations, this work proposes a Hypernetwork-Driven Trustworthy Recommendation framework (HDTR), aiming to provide reliable recommendations while accommodating personalized user requirements. Specifically, hypernetworks are first employed to efficiently generate and initialize client-side recommendation models, where user preference representations are encoded as inputs to produce personalized parameterizations. Next, within the client model, item attribute embeddings are incorporated as a form of global contextual information, enriching the representation with additional semantic cues. Furthermore, attention residual blocks are introduced to adaptively capture the relative importance of different item attributes, thereby enhancing feature interaction modeling. Extensive experiments conducted on the Movielens1M, Hetrec-movielens, and Douban datasets demonstrate that the proposed approach consistently outperforms existing baselines, achieving improvements of approximately 4.31% in MAE, 4.01% in RMSE, and 3.70% in Accuracy, which verifies its effectiveness in both accuracy and robustness.