面向稀疏矩阵偏置的协同过滤推荐算法研究

发布时间：2018-04-20 03:38

本文选题：网络信息 + 个性化推荐系统　；参考：《长安大学》2017年硕士论文

【摘要】：为了让用户从海量信息中高效地获取自己需要的信息,推荐系统可以通过分析用户历史行为来了解用户的偏好,从而主动地为用户推荐其感兴趣的信息,满足用户的个性化需求。目前最近邻思想、相似度思想和加权思想依然是各类推荐算法中最常用的思想,但随着用户数量和系统规模的不断扩大,推荐系统面临着数据稀疏、流行偏置和可扩展性差等问题。传统的推荐算法受数据稀疏的影响,相似度的准确性不足,从而导致最近邻搜索不准确。本论文针对数据稀疏问题,提出了基于奇异值分解的协同过滤改进算法,该算法中用户之间的相似度不再使用通用的评分矩阵计算,而是采用用户特征向量矩阵计算。其具体实现方法是通过奇异值分解获得用户特征向量矩阵和物品特征向量矩阵,并利用用户和物品之间的潜在关系,用奇异值去提取一些本质特征,计算两个用户对应特征向量的相似程度,从而得到用户之间的相似度。此外,传统的推荐算法往往倾向于推荐流行度较高的物品。本论文针对流行偏置问题,在传统的算法中根据物品流行度和用户兴趣信息引入惩罚函数,从而提出了基于惩罚函数的协同过滤改进算法。其算法的思想是从用户行为数据的产生过程对用户模型的影响进行分析,根据物品流行度和用户的兴趣信息构建惩罚度函数,并使用惩罚函数调节不同流行度物品在用户模型中权重。在MovieLens100K数据集上,对面向稀疏矩阵偏置的协同过滤推荐算法性能进行了验证,实验结果表明基于奇异值分解的协同过滤改进算法比基于用户的协同过滤推荐算法的准确率提升了0.68%,改善了最近邻搜索的准确性,其适用于新闻推荐领域。基于惩罚度的协同过滤推荐算法比基于用户的协同过滤推荐算法的覆盖率提高了1.47%,缓解了流行度对用户行为的影响,其适合电子商务推荐领域。
[Abstract]:In order for users to obtain the information they need efficiently from the mass of information, the recommendation system can analyze the historical behavior of users to understand the preferences of users, so as to actively recommend the information of interest to the users. To meet the personalized needs of users. At present, the nearest neighbor thought, the similarity thought and the weighted thought are still the most commonly used ideas in all kinds of recommendation algorithms, but with the continuous expansion of the number of users and the scale of the system, the recommendation system is faced with sparse data. Popular bias and poor scalability and other problems. The traditional recommendation algorithm is influenced by sparse data, and the accuracy of similarity is insufficient, which leads to inaccuracy of nearest neighbor search. In this paper, an improved collaborative filtering algorithm based on singular value decomposition (SVD) is proposed to solve the data sparsity problem. In this algorithm, the similarity between users is calculated not by the common score matrix, but by the user eigenvector matrix. The implementation method is to obtain user eigenvector matrix and item eigenvector matrix by singular value decomposition, and extract some essential features by singular value using the latent relation between user and item. The similarity between the two users is obtained by calculating the similarity of the corresponding feature vectors between the two users. In addition, traditional recommendation algorithms tend to recommend items with high popularity. Aiming at the problem of popular bias, this paper introduces a penalty function according to the information of item popularity and user's interest in the traditional algorithm, and then proposes an improved collaborative filtering algorithm based on penalty function. The idea of the algorithm is to analyze the influence of the user behavior data on the user model, and to construct the penalty degree function according to the item popularity and the user's interest information. And the penalty function is used to adjust the weight of different popular items in the user model. Based on MovieLens100K data set, the performance of collaborative filtering recommendation algorithm for sparse matrix bias is verified. The experimental results show that the improved collaborative filtering algorithm based on singular value decomposition improves the accuracy of collaborative filtering recommendation algorithm by 0.68, improves the accuracy of nearest neighbor search, and is suitable for news recommendation field. Compared with the user-based collaborative filtering recommendation algorithm, the penalty degree based collaborative filtering recommendation algorithm increases the coverage by 1.47%, which alleviates the influence of popularity on user behavior and is suitable for e-commerce recommendation field.
【学位授予单位】：长安大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】