面向显式反馈和隐式反馈的协同推荐算法研究

发布时间：2018-07-11 11:42

本文选题：推荐算法 + 协同过滤　；参考：《深圳大学》2016年硕士论文

【摘要】：在当今的电子商务领域,为了挖掘出用户的兴趣爱好并提供个性化推荐,用户的在线行为(如评分、评级、浏览)被认为是最有价值的来源之一。然而,之前大多数的工作主要集中在对单类型的用户行为的研究,如单一的显式反馈(如评分、评级)或单一的隐式反馈(如浏览、点击、加入购物车),极少有工作研究如何将显式反馈和隐式反馈融合起来。在本论文中,我们结合显式反馈和隐式反馈,主要研究并解决一种面向推荐系统中异构反馈的协同推荐问题。该问题目前主要面临两大挑战,如用户反馈数据的异构性和用户隐式反馈的不确定性。为了解决这个问题,本文分别从面向推荐系统中异构反馈的协同过滤和面向推荐系统中异构反馈的协同排序两个角度开展研究。从面向协同过滤角度,本文设计了一种自主迁移学习模型来自主学习和挖掘用户的隐式爱好,提出了一种自主迁移学习算法(Self Transfer Learning),简称sTL;然后,从面向协同排序角度,通过融合面向排序的模型和面向评分的模型,设计了一种分步矩阵分解(Staged Matrix Factorization)的协同排序算法,简称SMF。为了证明新提出的sTL算法和SMF算法推荐的有效性,本文在三个数据集上做了大量实验,实验结果表明我们提出的sTL算法明显优于其它几大经典算法(如SVD、SVD++),能有效提高精度、改善推荐效果,且实验还表明新提出的sTL算法具有很好的稳定性和继承性;同时,实验结果还表明,与其它面向排序的推荐算法(如BPR)相比,我们新提出的SMF算法能充分利用显式反馈和隐式反馈,能有效改善推荐效果。据我们所知,在当今的电子商务领域,很少存在单一的显式反馈,也很少存在单一的隐式反馈,大部分既包含显式反馈,又包含隐式反馈,由此可以看出,我们新提出的sTL算法和SMF算法能很好适应当前电子商务领域的需求,因此在当今电子商务领域具有很好的研究价值和应用价值。
[Abstract]:In the field of electronic commerce, in order to find out the interests of users and provide personalized recommendations, the online behavior of users (such as rating, browsing) is considered to be one of the most valuable sources. However, most of the previous work focused on a single type of user behavior, such as a single explicit feedback (such as rating, rating) or a single implicit feedback (such as browsing, clicking, etc.). There is little work on how to combine explicit and implicit feedback. In this paper, we combine explicit feedback and implicit feedback to study and solve a collaborative recommendation problem for heterogeneous feedback in recommendation systems. There are two main challenges to this problem, such as the heterogeneity of user feedback data and the uncertainty of user implicit feedback. In order to solve this problem, this paper focuses on collaborative filtering for heterogeneous feedback in recommendation system and collaborative sorting for heterogeneous feedback in recommendation system. From the point of view of collaborative filtering, this paper designs an autonomous transfer learning model from the main learning and mining users' implicit hobbies, and proposes a self transfer learning algorithm, called sTL, and then, from the perspective of collaborative sorting, proposes a self transfer learning algorithm. By combining the sorting oriented model and the scoring oriented model, a step Matrix factorization (SMF) cooperative sorting algorithm is designed. In order to prove the validity of the new proposed STL algorithm and SMF algorithm, a large number of experiments have been done on three datasets. The experimental results show that the proposed sTL algorithm is obviously superior to some other classical algorithms (such as SVD- SVD) and can effectively improve the accuracy. The experimental results show that the proposed sTL algorithm has good stability and inheritance. At the same time, the experimental results also show that compared with other ranking oriented recommendation algorithms (such as BPR), the proposed algorithm has good stability and inheritance. Our new SMF algorithm can make full use of explicit feedback and implicit feedback and can effectively improve the recommendation effect. As far as we know, in the field of electronic commerce today, there are few single explicit feedback and few single implicit feedback, most of which contain both explicit feedback and implicit feedback. Our new STL algorithm and SMF algorithm can meet the needs of the current electronic commerce field, so they have good research value and application value in the field of electronic commerce.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.3

【相似文献】