针对稀疏性的协同过滤优化算法研究

发布时间：2018-11-09 18:49

【摘要】：协同过滤是推荐系统中最流行且最成功的推荐算法。它基于群体智慧的思想,为目标用户选取行为最接近的邻居用户并根据邻居用户的喜好为目标用户进行推荐。协同过滤算法在帮助用户解决“信息过载”问题的同时还受到诸多问题的困扰。稀疏性问题是其面临的最主要问题,由于评分数据的过度稀疏导致相似度计算和邻居选择不准确,最终影响推荐结果的精度,同时推荐结果的可信度也大打折扣。本文针对协同过滤中的稀疏性问题,进行了较为深入地研究,提出了以下三种优化的协同过滤算法。(1)提出了基于预测值和用户/项目平均绝对误差(Mean Absolute Error,MAE)值填充的优化算法。计算每个用户/项目的MAE值,按照填充规则使用预测值与MAE值对评分矩阵中的空缺值进行填充,使用填充后的评分矩阵执行协同过滤算法。该算法确保填充值贴近各用户的评分均值,既保障了预测结果能够满足不同用户的个性化评分习惯,又提高了推荐的准确度。(2)提出了基于项目聚类的优化算法。对原始矩阵的项目列进行聚类,根据聚类结果构建两种更加稠密的“用户-类别”矩阵,据此计算出用户的两种相似度。将得到的相似度线性加权并结合修正系数作为最终的相似度进行推荐。该算法在更为稠密的矩阵上进行相似度计算,使推荐结果更为可信;同时由于矩阵规模减小,算法的效率也有所提升。(3)提出了基于信任网络的优化算法。将信任关系引入协同过滤中。使用用户间的共同评分以及传递规则构建信任网络。将信任网络的信任度与传统用户相似度进行线性加权,对用户未评分项目进行预测。该算法根据目标用户所信任用户的喜好进行商品推荐,其结果不仅比传统的基于邻居的推荐更加可信,同时推荐精度也获得了一定程度的提升。通过大量对比实验,证明了本文算法的有效性以及可行性。
[Abstract]:Collaborative filtering is the most popular and successful recommendation algorithm in recommendation systems. Based on the idea of group intelligence, it selects the nearest neighbor users for the target users and recommends them for the target users according to their preferences. Collaborative filtering algorithm not only helps users solve the problem of "information overload", but also suffers from many problems. The problem of sparsity is the most important problem. Due to the inaccuracy of similarity calculation and neighbor selection due to the excessive sparsity of score data, the accuracy of recommendation results is affected, and the credibility of recommendation results is also greatly reduced. In this paper, the sparsity of collaborative filtering is studied in depth, and three optimized collaborative filtering algorithms are proposed. (1) based on the prediction value and the user / item average absolute error (Mean Absolute Error,), the following three optimized collaborative filtering algorithms are proposed. MAE) value filling optimization algorithm. The MAE value of each user / item is calculated, and the vacant value in the rating matrix is filled with the predicted value and the MAE value according to the fill rule. The cooperative filtering algorithm is implemented using the filled rating matrix. This algorithm ensures that the filling value is close to the average value of each user, which not only ensures that the prediction results can meet the individual scoring habits of different users, but also improves the accuracy of recommendation. (2) an optimization algorithm based on item clustering is proposed. Based on the clustering of the item columns of the original matrix, two more dense "user-class" matrices are constructed, and the two kinds of similarity of users are calculated according to the clustering results. The obtained similarity is linearly weighted and the modified coefficient is used as the final similarity to recommend. The algorithm computes the similarity on a denser matrix, which makes the recommendation result more reliable. At the same time, the efficiency of the algorithm is also improved because of the reduction of matrix size. (3) an optimization algorithm based on trust network is proposed. Trust relationship is introduced into collaborative filtering. A trust network is built using common ratings and delivery rules among users. The trust degree of the trust network is linearly weighted to the traditional user similarity, and the ungraded items are predicted. The algorithm is based on the preferences of the users trusted by the target users. The results are not only more reliable than the traditional recommendation based on neighbors, but also improve the recommendation accuracy to a certain extent. The effectiveness and feasibility of the proposed algorithm are proved by a large number of comparative experiments.
【学位授予单位】：西北大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【相似文献】