Abstract: Most previous collaborative filtering approaches employ the matrix factorization techniques to learn latent user feature profiles and item feature profiles. However, the matrix factorization based methods may not make full use of the rating information, leading to unsatisfying performance. In recent years, deep learning has been approved to be able to find good representations in natural language processing, speech recognition, image classification, and so on. Moreover, the rating matrix not only contains the rating information, but also the ranking information. Furthermore, the local structure of rating matrix is considered to learn latent users' and items' representations. To this end, this paper proposes an ensemble local representation learning based recommendation algorithm. In the proposed algorithm, the randomly selected anchor is used to obtain local matrix, and then AutoEncoder is applied to learn latent representations and obtain submodel. Finally, the ranking information is incorporated by defining (user, item) pairs. The experimental results on two data sets validate the effectiveness of the proposed algorithm, compared with matrix factorization algorithms. Also, the algorithm outperforms the state-of-the-art model LCR (local collaborative ranking), which indicates the superiority of applying deep learning techniques to recommendation.

Key words: recommender systems, deep learning, AutoEncoder, ranking learning, local structure

摘要： 以往的协同过滤方法大部分采用基于矩阵分解的方法来学习用户和商品的隐性特征表示，但是基于矩阵分解的方法没有完全利用评分信息，导致不好的效果。近年来，深度学习已经在自然语言处理、语音识别以及图像分类等领域被证明可以很好地进行表示学习。而且在用户对商品的评分矩阵中，不仅只有评分信息，还有隐含的倾向性排序信息。更进一步，针对整个评分矩阵进行特征表示学习的时候，不能满足用户聚类以及商品类别的局部结构特性。因此，提出了一种基于集成局部性特征学习的推荐算法。在该算法中，利用随机选择的锚点得到局部矩阵，然后在局部矩阵上利用自动编码机进行学习得到子模型，同时定义一种（用户，商品）二元组来考虑评分信息的排序关系。在两组数据上进行了实验，结果表明该算法显著优于经典的基于矩阵分解的推荐算法，并且该算法将深度学习用于推荐系统中，效果比LCR（local collaborative ranking）优越。

关键词: 推荐系统, 深度学习, 自动编码机, 排序学习, 局部结构