基于用户分解和社交融合的推荐算法研究

发布时间：2017-12-27 01:04

本文关键词：基于用户分解和社交融合的推荐算法研究　出处：《华东师范大学》2016年博士论文　论文类型：学位论文

【摘要】：在互联网时代,人们不受时空限制地享受着互联网提供的信息和服务的同时,也不得不面对海量且规模不断增长的数据以及大量的无效信息。推荐系统作为信息过滤的重要工具,同时为用户提供个性化的信息服务,越来越受到人们的欢迎,并且必不可少。典型的推荐技术为协同过滤算法,该算法利用其他购买过共同物品的相似用户来预测当前用户的偏好,由于其重要的应用价值和学术价值,已经被工业界和学术界广泛的研究。然而,在多个用户的历史行为混合且无法显式区分时,或用户行为受到系统中的其他用户影响时,很多算法都不能较好地准确表示用户的真实兴趣或偏好,进而影响到推荐服务的性能。因此,本文尝试用虚拟用户重新表示每个用户的真实偏好,并提高推荐的品质。虚拟用户是表示真实用户兴趣并用于帮助产生推荐的向量或配置文件。针对以上两种情况,本文分别从用户分解和社交融合两个方面研究虚拟用户对推荐性能的影响。本文首先提出两种用户分解算法来研究用户历史行为混合的情况下,如何表示虚拟用户,利用虚拟用户识别真实用户,并为识别的用户做个性化推荐。在网络协议电视服务中,家庭成员无差别地使用该服务,为研究混合的用户行为提供良好的研究对象,这里的两种方法都是在该服务上进行。第一种方法是基于时间片分割的用户识别推荐算法。该方法定义一个时间片内的活动为一个虚拟用户的活动,并利用隐式评分捕获虚拟用的偏好,合并偏好相似的虚拟用户作为真实用户,并为这些用户做个性化的推荐；第二种方法是基于子空间聚类的用户识别推荐算法。该算法基于账户-项目-时间的张量分解和子空间聚类来发现用户的时序行为,利用聚类的时间段来表示虚拟用户,合并偏好相似的虚拟用户作为真是用户,并为这些用户做个性化的推荐。实验表明两种方法都比之前的方法性能更优,并且第二种方法比第一种方法更自动化且性能更好。在社交网络中,用户之间进行各种各样的交流,为研究用户间的兴趣相互影响提供了理想的研究对象。因此,基于用户(或用户对物品的偏好)受到其社交邻居的影响这一假设,本文进一步提出融合社交影响力的两种新颖的推荐算法。这两种方法在用户-物品矩阵分解框架的基础上融合用户-用户社交链接矩阵。具体而言,利用用户评分和该用户社交邻居的评分构建虚拟用户来表示用户的真实偏好。第一种融合个性化因素和加权社交影响力的方法,利用用户和社交邻居之间的隐式偏好构建社交影响力,在预测时同时结合用户个人因素和社交影响力带来的偏好影响,在很大程度上可以提高对评分较少的用户的预测性能。然而,每个用户受到其社交邻居的影响程度不同：不仅个数不同,而且相互的影响程度也不同。因此,第二种方法利用非对称社交影响力来重新表示用户间产生的相互影响。同时,实验结果证明本文提出的方法比目前的其他方法更加准确而高效。
[Abstract]:In the Internet era, people enjoy the information and services provided by the Internet without being limited by time and space. At the same time, they also have to face massive and growing data and lots of invalid information. As an important tool for information filtering and providing personalized information services for users, the recommendation system is becoming more and more popular and indispensable. The typical recommendation technology is collaborative filtering algorithm, which uses other similar users who purchase common goods to predict the current user preferences. Because of its important application value and academic value, it has been widely studied by industry and academia. However, when many users' historical behaviors are mixed and cannot be distinguished distinctions, or users' behaviors are influenced by other users in the system, many algorithms can not accurately represent the users' real interests or preferences, and further affect the performance of recommendation services. Therefore, this article attempts to reexpress the true preferences of each user with a virtual user and improve the quality of the recommendation. A virtual user is a vector or configuration file that represents a real user's interest and is used to help produce a recommendation. In view of the above two situations, this paper studies the effect of virtual users on the performance of recommendation from two aspects of user decomposition and social integration. In this paper, two user decomposition algorithms are first proposed to study how to represent virtual users under the condition of mixed user behavior, to identify the real users by virtual users, and to make personalized recommendation for the identified users. In the network protocol TV service, family members use this service without distinction, providing good research objects for studying the mixed user behavior. The two methods here are carried out on the service. The first method is a user recognition recommendation algorithm based on time slice segmentation. This method defines a time slice of the activity as a virtual user activity, and the implicit score for virtual capture preferences, preferences with similar virtual users as real users, and these users make personalized recommendation; the second method is user identification subspace clustering recommendation algorithm based on. Based on account item time tensor decomposition and subspace clustering, the algorithm finds user's temporal behavior, uses clustering time to express virtual users, and combines virtual users with similar preferences as real users, and makes personalized recommendation for these users. Experiments show that the two methods are better than the previous ones, and the second methods are more automated and better than the first one. In social networks, a variety of communication between users provides an ideal research object for the study of the interaction between users. Therefore, based on the hypothesis that users (or users' preferences for goods) are influenced by their social neighbors, this paper further proposes two novel recommendation algorithms that integrate social influence. These two methods integrate the user - user social link matrix on the basis of the user - object matrix decomposition framework. In particular, the user's real preference is expressed by building a virtual user by using the user score and the score of the user's social neighbor. The first method and the weighted fusion individual factors influence social construction, social influence the implicit preference between the user and the social neighbors, when combined with the impact of the user's personal factors and social influence preferences, to a large extent can improve the prediction performance of the score less user. However, each user has a different degree of influence by its social neighbors: not only a different number, but also a different degree of influence. Therefore, the second methods use asymmetric social influence to rerepresent the interaction between users. At the same time, the experimental results show that the proposed method is more accurate and efficient than the other methods.
【学位授予单位】：华东师范大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.3

【相似文献】