基于高维显性特征的矩阵分解推荐

发布时间：2018-01-04 14:48

本文关键词：基于高维显性特征的矩阵分解推荐　出处：《大连理工大学》2015年硕士论文　论文类型：学位论文

【摘要】：隐含特征模型是协同过滤推荐算法中的重要部分。由于它只需利用用户-条目的打分信息就能达到非常好的预测效果而广为使用。通过特征分解,算法从简单的评分矩阵获得用户与条目的特征向量。特征向量包含用户和条目的隐含属性,能够对未知的缺失值进行预测。由于该过程是将一个高维二维矩阵分解得到两个低维特征矩阵,因此这种推荐算法也称为矩阵分解推荐。矩阵分解推荐需要的信息类型比较单一,只需要用户-条目的稀疏矩阵,因此适合特征比较少的情况。然而,矩阵分解整个特征化的过程完全是隐含的,用户或条目特征向量的哪个维度具体代表哪个特征是完全不得而知的。因此,一些已观测到的显性特征很难加入到模型中。基于矩阵分解模型,在本文中我们提出了一个能够将已观测的显性特征整合到模型中的算法。该算法通过升维的方式整合显性特征,从而提高了预测准确性。本文的主要贡献如下：(1)能够捕捉到用户对具有某些特征的条目偏好信息,根据偏好信息对用户的特征向量做调整。(2)模型也适用于条目特征值具有多值的情况。(3)基本矩阵分解算法通过提升特征向量的维度来包含更多特征,这种方式不仅无法控制具体特征,而且会带来过拟合问题。本文将特征的维度提升，且仅关注条目具有的特征而非全部特征,能够较好解决过拟合问题。实验结果证明,该算法优于现有的基于矩阵分解的算法。
[Abstract]:Implicit feature model is an important part of collaborative filtering recommendation algorithm. It is widely used through feature decomposition because it can only use the user-item scoring information to achieve a very good prediction effect. The algorithm obtains the feature vectors of users and items from a simple scoring matrix, which contains the implicit attributes of users and items. The unknown missing values can be predicted because the process is to decompose a high-dimensional two-dimensional matrix into two low-dimensional characteristic matrices. Therefore, this recommendation algorithm is also called matrix factorization recommendation. Matrix decomposition recommendation requires a single type of information, only needs the sparse matrix of user-item, so it is suitable for the case of fewer features. The process of matrix decomposition is completely implicit, which dimension of the user or item feature vector is completely unknown. It is difficult to add some observed dominant features to the model based on matrix decomposition model. In this paper, we propose an algorithm to integrate observed dominant features into the model. The main contribution of this paper is as follows: 1) to capture user preferences for items with certain characteristics. The model is also suitable for the case where the item eigenvalue has multiple values. The basic matrix factorization algorithm includes more features by lifting the dimension of the feature vector. This approach can not only control the specific features, but also bring about the problem of over-fitting. This paper will enhance the dimension of features, and only focus on the features of the items, not all of the features. The experimental results show that the proposed algorithm is superior to the existing one based on matrix decomposition.
【学位授予单位】：大连理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.3;O151.21

【相似文献】