Abstract:

The recommendation system is one of the most important technologies to help users quickly discover the content they are interested in from massive amounts of data. Sparsity and cold start are the main problems facing recommendation systems. For the sparsity problem, many recommendation algorithms have considered the use of additional auxiliary information, such as reviews, summary, or summaries, to improve prediction accuracy. These algorithms have indeed improved prediction accuracy to some extent. However, most of the existing algorithms are based on the bag-of-words model, and the understanding and utilization of this auxiliary information lacks depth and remains on the surface. This paper proposes a new type of recommendation system algorithm, the Deep Collaborative Filtering（DCF） algorithm. DCF integrates Long Short-Term Memory（LSTM） networks and Probability Matrix Factorization（PMF）. This algorithm can not only learn user characteristics based on user ratings, but also use natural language processing technology to deeply mine auxiliary information and learn very accurate item characteristics. After verification on the real datasets Movielens 100K and 1M, the standard error of the DCF algorithm is reduced by 2.54% and 3.96% respectively compared with the existing algorithms.

Key words: Deep Collaborative Filtering（DCF）, Long Short-Term Memory（LSTM）, Probability Matrix Factorization（PMF）, recommendation system

摘要：

推荐系统是帮助用户在海量的数据中快速发掘出他们感兴趣内容的最重要的技术之一。稀疏性和冷启动是推荐系统面临的主要问题。针对稀疏性问题，已有多种推荐算法考虑利用额外的辅助信息，如评论、摘要或概要等来提高预测准确性。这些算法确实已经在一定程度上提高了预测准确性，但是，已有的算法大都是基于词袋模型，对这些辅助信息的理解和利用缺乏深度，留于表面。提出了一种新型的推荐系统算法：深度协同过滤算法（DCF）。DCF集成了长短期记忆网络（LSTM）和概率矩阵分解（PMF）。该算法不仅能够基于用户评分学习用户特征，而且能深度挖掘辅助信息，学习到更精确的物品特征。经过在真实数据集MovieLens100K和1M上的验证，结果表明DCF算法的根均方误差比现有算法分别降低了2.54%和3.96%。

关键词: 深度协同过滤, 长短期记忆网络, 概率矩阵分解, 推荐系统