面向完全冷启动的深度混合协同过滤推荐算法

Deep Hybrid Collaborative Filtering Recommendation Algorithm for Complete Cold Start

基于矩阵分解的协同过滤算法近年来获得了巨大的成功,但是依然存在冷启动,忽视用户及物品特征等问题,从而导致推荐质量不佳,用户体验度下降。论文提出了一种基于深度学习的混合协同过滤推荐算法,尝试引入堆栈降噪自编码器学习物品的隐含特征,同时结合半监督S4VM和隐含因子模型,综合考虑物品的内容特征及时间因素,以预测未评分的数据,解决冷启动问题。在标准数据集Movielens上进行的测试表明:该算法能有效预测冷启动物品的评分,性能提升显著,较传统推荐性能提升约为12%。

Recently collaborative filtering based on matrix factorization has been very successful,but it still exists some problems such as cold-start and the ignorance of characteristic of users and items,as a result,the efficiency of recommendation is poor and have a negative impact on the user experience. The paper presents a hybrid collaborative filtering recommendation algorithm based on deep learning. The paper intends to predict the unrated items and solve cold-start problem by introducing a stacked denoising auto encoder to learn latent factors of items. Semi-supervised support 4 vector machine and latent factor model are combined and content feature of items and time are considered comprehensively. Evaluations on a standard Movielens dataset indicate that the algorithm can efficiently predict the cold-start items and improve performance significantly. Compared to traditional recommendation,the performance is increased by 12%.