加权高效用因子的两阶段混合推荐算法

传统协同过滤算法大多是围绕如何降低评分误差展开研究,未涉及用户评分过程。本文考虑到用户评分动机和用户本身评分倾向的情况,将用户评分过程分为用户评分和物品选择两个阶段,从预测用户兴趣概率和用户效用角度出发,采用潜在狄利克雷分布模型(LDA)挖掘出用户潜在高效用因子和物品被靶向概率因子,进而将两种因子加权融合作为第一阶段;第二阶段采用奇异值分解预测用户评分值并根据该评分值选择物品。综上,本文提出一种加权高效用因子的两阶段混合推荐算法(hybrid recommendation algorithm based on two-phase weighted high utility factor,Htp_Uf)。在 MovieLens数据集上,实验结果表明,该算法在归一化累计折损增益(NDCG)和 1-Call两种评价标准下优于其他4种推荐算法,能够有效提高推荐质量。

Two-phase weighted high-utility factor-based hybrid recommendation algorithm

The traditional collaborative filtering algorithm focuses on methods for reducing the rating error, but it does not involve a user rating process. In this paper, we divide the user rating process into two phases, i.e., user rating and item selection, and take into account user-rating motivation and user-rating tendency. From the perspectives of predicting a user’s interest probability and user utility, we adopt a latent dirichlet allocation (LDA) to mine a user latent high-utility factor and an article-targeted probabilistic factor, and then weight and fuse these two factors in the first phase. In the second phase, we use singular value decomposition to predict a user rating value, which we then use in user recommendation lists. Our main contribution is our proposal of a two-phase weighted high-utility factor-based hybrid recommendation algorithm, which we abbreviate as Htp_Uf. We compare the effectiveness of the Htp_Uf algorithm with that of others using the MovieLens dataset. The results show that the performance of the proposed algorithm is superior to that of four other recommendation algorithms in terms of two evaluation criteria, namely, the normalized discounted cumulative gain and 1-call, which can effectively improve the quality of the recommendation.