融合多源信息的推荐算法研究

发布时间：2018-08-27 09:15

【摘要】：随着互联网相关技术的不断发展,从海量数据中找到有价值的信息变得越来越困难,即用户面临严重的信息过载问题。推荐算法通过分析用户的历史活动数据,挖掘用户的隐藏偏好,为用户提供个性化的推荐服务,成为解决信息过载问题的有效手段,近年来受到学术界和工业界的广泛关注。在实际应用中,推荐算法面临各种挑战,如数据稀疏、可扩展性、冷启动、准确性、可解释性等。针对这些挑战,国内外的研究人员提出了大量的解决方案。然而,仅仅利用用户的活动记录信息不能从本质解决推荐系统中存在的固有问题。近年来,多种类型的多源信息越来越丰富,如项目属性信息、社交网络信息、地理位置信息和用户评论信息等。可用的多源信息是用户历史活动记录的有益补充,为解决推荐系统中信息缺乏问题带来契机。同时,如何在推荐系统中融合多源信息,提升推荐算法的性能,解决推荐系统存在的问题,成为推荐系统领域重要的研究问题。在本文中,我们主要针对推荐系统中数据稀疏、可扩展性、冷启动和准确性等问题,在现有工作的基础上,结合协同过滤推荐算法、基于社交网络推荐算法和兴趣点推荐算法等领域的现有成果,研究融合多源信息的推荐算法。本文的主要工作和贡献如下：1.基于项目属性耦合的矩阵分解推荐算法现有一些基于矩阵分解的推荐算法仅关注用户端的冷启动问题,而忽视项目端的冷启动问题。并且,缺乏有效的度量方式计算由类别型数据所描述的项目之间的相似度。为了解决以上问题,本文提出基于属性耦合的矩阵分解推荐算法。在矩阵分解模型中,集成项目的属性信息来改进推荐算法的性能,减轻项目端的冷启动问题。利用属性信息构建正则化项,约束矩阵分解学习隐特征向量,使得属性信息相似的项目,它们的隐特征向量尽可能相似。在构建包含属性信息的正则化项时,利用耦合对象相似度计算项目之间的相似度。实验结果表明,基于属性耦合的矩阵分解推荐算法性能优于目前主流的推荐算法,能有效减轻项目端的冷启动问题。2.融合用户社会地位和矩阵分解的推荐算法随着社交网络的出现,越来越多的推荐系统利用社交网络中用户之间的信任关系来改进推荐算法的性能。然而,现有基于社交网络推荐算法忽略了以下两个问题：(1)在不同的领域中,用户通常信任不同的朋友；(2)由于用户在不同的领域内具有不同的社会地位,因此,用户在不同的领域内受朋友的影响程度是不同的。为了解决以上问题,本文首先利用整体的社交网络结构信息、和用户的评分信息推导特定领域社交网络结构,然后利用PageRank算法计算用户在特定领域的社会地位,最后提出了一种融合用户社会地位信息的矩阵分解推荐算法。实验结果表明,本文提出融合用户地位信息的矩阵分解推荐算法的性能优于传统的基于社交网络推荐算法。3.基于地点重要性和用户权威性增强的兴趣点推荐算法智能移动设备的普及、GPS和WEB2.0等技术的发展促使基于位置的社交网络平台不断涌现。兴趣点推荐从基于位置的社交网络应用提供的多源信息源中挖掘用户兴趣偏好,为用户推荐用户可能感兴趣的、未访问过的地理位置,已经成为基于位置的社交网络应用不可或缺的组成部分。一些研究人员将兴趣点看作传统推荐领域中的项目,提出了一些兴趣点推荐算法。然而,不同于传统领域的推荐,兴趣点推荐具有一些独特的属性,已有的兴趣点推荐算法存在如下的问题：(1)多数已有的兴趣点推荐算法简化用户签到频率数据,仅使用二进制值来表示用户是否访问一个兴趣点；(2)基于矩阵分解的兴趣点推荐算法把签到频率数据和传统推荐系统中的评分数据等同看待,使用高斯分布模型建模用户的签到行为；(3)较少研究工作考虑地点重要性和用户权威性对用户签到行为的影响。为了解决上述问题,本文集成概率因子模型和地点重要性来建模用户的签到行为,提出了地点重要性和用户权威性增强兴趣点推荐算法。具体地,同时考虑用户经验和兴趣点之间的相互影响,以及兴趣点之间的相互影响,采用HITS和PageRank混合模型计算地点重要性和用户权威性。而且,将用户权威性作为个性化因子衡量用户的隐式反馈。实验结果表明,地点重要性和用户权威性增强的兴趣点推荐算法性能优于基准兴趣点推荐算法。4.基于Ranking的泊松矩阵分解兴趣点推荐算法除了简化用户签到数据,仅使用二进制值表示用户是否访问兴趣点,以及将签到频率数据和传统推荐系统中的评分数据等同看待外,已有的多数兴趣点推荐算法忽视用户签到数据的隐式反馈属性,即,仅采用逐点拟合可观测签到数据的方法学习用户和兴趣点的隐特征向量,忽视了用户签到数据之间的偏序关系。为解决以上问题,本文提出一个基于Ranking的泊松矩阵分解兴趣点推荐算法。首先,根据基于位置社交网络中用户的签到行为特点,利用泊松分布模型替代高斯分布模型建模用户在兴趣点上签到行为,然后采用BPR标准优化泊松矩降分解的损失函数,拟合用户在兴趣点对上的偏序关系。最后,利用包含地域影响力的正则化因子约束泊松矩阵分解的过程。实验结果表明,基于Ranking的泊松矩阵分解兴趣点推荐算法的性能优于传统的兴趣点推荐算法。
[Abstract]:With the development of Internet related technology, it is becoming more and more difficult to find valuable information from massive data, that is, users are facing serious information overload problem. Recommendation algorithm can analyze users'historical activity data, mine users' hidden preferences, and provide personalized recommendation service for users, which becomes a solution to information overload problem. In recent years, the effective means of the problem have attracted wide attention from academia and industry. In practical applications, recommendation algorithms face various challenges, such as sparse data, scalability, cold start, accuracy, EXPLANABILITY and so on. In recent years, many kinds of multi-source information, such as item attribute information, social network information, geographic location information and user comment information, have become more and more abundant. The available multi-source information is a useful supplement to the user's historical activity record, and it can solve the information in recommendation system. At the same time, how to integrate multi-source information into the recommendation system, improve the performance of recommendation algorithm and solve the problems existing in the recommendation system has become an important research problem in the field of recommendation system. The main work and contributions of this paper are as follows: 1. Matrix decomposition recommendation algorithm based on item attribute coupling has some recommendations based on matrix decomposition. In order to solve the above problems, a recommendation algorithm based on attribute-coupled matrix decomposition is proposed. In the matrix decomposition model, the items are integrated. Attribute information is used to improve the performance of recommendation algorithm and alleviate the cold start problem at the project side. Regularization items are constructed by attribute information, and implicit eigenvectors are decomposed and learnt by constraint matrix decomposition to make the items with similar attribute information as similar as possible. The experimental results show that the performance of the proposed algorithm based on attribute coupling is better than that of the current mainstream recommendation algorithm and can effectively alleviate the cold start problem of the project. 2. With the emergence of social networks, more and more recommendation systems are using the proposed algorithm based on user social status and matrix decomposition. However, existing recommendation algorithms based on social networks ignore the following two issues: (1) users usually trust different friends in different domains; (2) users have different social status in different domains, so users have different collars. In order to solve the above problems, this paper firstly deduces the social network structure of a specific domain by using the overall social network structure information and user rating information, then calculates the social status of users in a specific domain by using PageRank algorithm, and finally proposes a social status information fusion of users. Experimental results show that the performance of the proposed algorithm is better than that of the traditional recommendation algorithm based on social network. 3. Point of interest recommendation algorithm based on location importance and user authority enhancement is popular in smart mobile devices, and the development of GPS and WEB2.0 technologies. Interest Point Recommendation (IPR) has become an indispensable part of location-based social networking applications to mine users'interest preferences from multiple sources of information provided by location-based social networking applications and to recommend geographic locations that users may be interested in and have not visited. However, unlike traditional recommendation, interest point recommendation has some unique attributes. The existing interest point recommendation algorithms have the following problems: (1) Most of the existing interest point recommendation algorithms simplify user check-in frequency. Data, using only binary values to indicate whether a user has access to a point of interest; (2) Matrix decomposition-based interest point recommendation algorithm treats the checkin frequency data as the score data in the traditional recommendation system, and uses Gaussian distribution model to model the user's checkin behavior; (3) less research considers the importance of location and user rights. In order to solve the above problems, this paper integrates probability factor model and place importance to model user's check-in behavior, and proposes a place importance and user authority enhanced interest point recommendation algorithm. The results show that the performance of the proposed algorithm is better than that of the benchmark interest point recommendation algorithm. 4. The Ran-based algorithm is better than the benchmark interest point recommendation algorithm. Poisson Matrix Decomposition Interest Point Recommendation (POP) algorithm based on King not only simplifies user check-in data, but also uses binary values to indicate whether users have access to interest points, and treats checkin frequency data as well as score data in traditional recommendation systems. Most of the existing POP recommendation algorithms ignore implicit feedback properties of user check-in data, that is, implicit feedback properties of user check-in data. In order to solve the above problems, a Poisson Matrix Decomposition Interest Point Recommendation algorithm based on Ranking is proposed. Firstly, the user's check-in line in location-based social networks is used. For this purpose, Poisson distribution model is used to replace Gauss distribution model to model user's checkin behavior at interest points, then BPR standard is used to optimize the loss function of Poisson moment decomposition and fit the partial ordering relation of user's interest points. The results show that the Ranking-based Poisson Matrix Decomposition Interest Point Recommendation algorithm outperforms the traditional Interest Point Recommendation algorithm.
【学位授予单位】：南京大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.3

【相似文献】