Abstract: In the traditional recommendation systems, the approach of matrix factorization collaborative filtering only just considers the single information of rating, as a shallow model, it can hardly learn deeper feature information. This paper proposes a multi-interaction deep matrix factorization model based on auxiliary information, firstly through deep learning model and merge more auxiliary information as input, effectively alleviates the problem of data sparsity. Then, the structure of multi-interactive non-linear network is leveraged to learn the deep feature representation of more abstract and dense; through inner product interactions on the latent features of users and items repeatedly, it obtains the different layers of feature representation results; finally it aggregates all the interaction results to predict. The experiment results on the Movielens latest 100K dataset show that the proposed model is over the state-of-the-art methods in RMSE.

Key words: collaborative filtering, deep learning, auxiliary information, multi-interaction, neural network, recommendation system

摘要： 传统的推荐系统中，基于矩阵分解的协同过滤方法只考虑单一的评分信息，而且作为浅层模型无法学习到更深层次的特征信息。提出一种基于深度学习的多交互混合推荐模型，通过深度学习模型融合更多的辅助信息作为输入，能够缓解数据的稀疏性问题；利用多层交互的非线性网络结构去学习更抽象、稠密的深层次特征表示；通过对用户和项目的隐表示进行多次内积交互获得不同层次的特征表示结果；聚合所有的交互结果进行预测。在Movieles latest 100K数据集上进行实验，采用[RMSE]指标进行评估，结果表明所提模型在推荐效果上有所提升。

关键词: 协同过滤, 深度学习, 辅助信息, 多层交互, 神经网络, 推荐系统