用于鲁棒协同推荐的元信息增强变分贝叶斯矩阵分解模型

A Metadata-enhanced Variational Bayesian Matrix Factorization Model for Robust Collaborative Recommendation

托攻击是协同过滤推荐系统面临的重大安全威胁.研究可抵御托攻击的鲁棒协同推荐技术已成为目前的重要课题.本文在引入用户嫌疑性评估策略的基础上,通过将用户嫌疑性及项类属等元信息与贝叶斯概率矩阵分解模型相融合,提出了用于鲁棒协同推荐的元信息增强变分贝叶斯矩阵分解模型(Metadata-enhance dvariationa lBayesian matri xfactorization,MVBMF),并设计了相应的模型增量学习策略.实验表明,与现有推荐模型相比,这种模型具备更强的攻击耐受力,能够有效提高推荐系统的鲁棒性.

Shilling attacks pose a significant threat to the security of collaborative filtering recommender systems. It has come to be an important task to develop the attack-resistant techniques for robust collaborative recommendation. Through evaluating the user suspiciousness, and further integrating Bayesian probabilistic matrix factorization model with the metadata including user suspiciousness as well as item types, this paper proposes the metadata-enhanced variational Bayesian matrix factorization （MVBMF） model for robust collaborative recommendation, and designs the corresponding incremental learning strategy. Experimental results show that comparing with the existed recommendation models, this model has stronger resistibility and can effectively improve the robustness of recommender systems.