个性化推荐的可解释性研究

发布时间：2018-01-26 05:29

  本文关键词： 个性化推荐 协同过滤 情感分析 可解释性 计算经济学 人工智能　出处：《清华大学》2016年博士论文　论文类型：学位论文

【摘要】：随着互联网的迅速发展,个性化推荐系统已经逐渐成为各种网络应用中不可缺少的核心功能,并以各种各样的方式影响着人们日常生活的方方面面:电子商务网站中的购物推荐引擎为用户提供可能感兴趣的商品推荐;社交网络中的好友推荐为用户寻找潜在的好友关注;视频网站中的视频推荐为用户提供最可能点击的视频推荐;新闻门户网站中的内容推荐为用户提供最有信息量的新闻——个性化推荐技术已经是支撑互联网智能的基础技术之一。个性化推荐系统已经经过了长达十几年的研究和发展,然而隐变量方法的大量使用使得个性化推荐算法及其推荐结果的可解释性仍然是困扰学术界重要问题之一,并且至今仍然没有在产业应用中得到很好的体现。举例而言,在很多实际推荐系统中,算法只为用户提供一份个性化的推荐列表作为结果,而难以向用户解释为什么要给出这样的推荐。缺乏可解释性的推荐降低了推荐结果的可信度,进而影响推荐系统的实际应用效果。考虑到推荐系统的应用范围之广和影响之大,可解释性推荐的研究具有其重要性和紧迫性。在本文中,我们从数据、模型和经济意义三个方面对推荐系统的可解释性进行研究,主要有贡献如下:1.数据的可解释性:数据输入是个性化推荐系统的第一步,而用户物品评分矩阵是个性化推荐算法,尤其是基于矩阵分解的个性化推荐算法最主要的数据输入形式。本文提出了基于双边块对角矩阵的局部化矩阵分解框架,并将其应用于矩阵分解的并行化。传统的矩阵分解算法将原始矩阵看做一个整体进行分解和预测,而缺乏对矩阵内在结构的理解。在本工作中,我们提出矩阵的双边块对角结构,并在理论上证明该结构与二部图上社区发现算法的数学等价性,从而解释矩阵内在的社区结构和社区关系。在社区结构的基础上,我们进一步提出了局部化的矩阵分解框架,并理论证明了它与传统矩阵分解算法的兼容性,从而为常用的矩阵分解算法提供了一个统一的并行化框架,在提高预测精度的同时大幅提高计算效率。2.模型的可解释性:在用户物品评分矩阵的数据基础上,个性化推荐模型对用户进行偏好建模并给出个性化推荐。本文提出了基于短语级情感分析的显式变量分解模型及其基于时间序列分析的动态化建模。基于矩阵分解的隐变量模型由于其较好的评分预测效果和可扩展性,逐渐成为了个性化推荐的基础算法并在实际系统中得到广泛的应用。然而由于变量本质上的未知性,隐变量模型难以对推荐算法和推荐结果给出直观可理解的解释,进而降低了推荐系统对用户的可信度。在本工作中,我们利用短语级情感分析技术从大规模的用户评论中抽取产品属性词及用户在不同属性上表达的情感,进而引入显式变量并提出基于显式变量分解模型的个性化推荐算法,一方面使得模型的优化过程具备了直观意义,另一方面给出在模型层面可解释的推荐结果和个性化推荐理由。由于用户在不同属性上的偏好具有间周期性,我们利用时间序列分析对用户偏好进行动态建模和预测,从而实现动态时间意义的可解释性推荐。3.推荐的经济学解释。推荐系统在用户行为数据和个性化偏好建模的基础上,以个性化推荐的方式隐式地调节商品在用户中的匹配和购买,从而在最终层面上影响所属系统的经济效益。本文提出基于互联网系统总福利最大化的个性化推荐框架并给出典型应用场景中的具体实现。随着人类传统线下活动的不断线上化,常见的互联网应用均可以形式化为“生产者—服务—消费者”模型,例如在电子商务网站中,网络商家(生产者)提供在线商品(服务),而网络用户(消费者)则在众多的商品中进行选择和购买。基于传统经济学的基本定义,本文首先给出了互联网环境下效用、成本和福利的基本概念与统一形式,并进一步给出了互联网应用中总社会福利的通用计算方法。在此基础上,我们以互联网服务分配为基本问题,提出基于网络福利最大化的个性化推荐框架。进一步,本文在典型的网络应用(电子商务、P2P借贷、在线众包平台)中对该框架进行具体化,并进行个性化的网络服务推荐与评测。实验结果表明,该方法可以在为用户提供高质量服务推荐的同时提升社会总福利,即在提升用户体验的同时又增强了社会效益。
[Abstract]:With the rapid development of Internet, personalized recommendation system has gradually become an indispensable core function of network application, and in a variety of ways affects all aspects of people's daily life: the e-commerce website in the shopping recommendation engine to provide users may be interested in recommendation; network of friends recommended for users to find the potential friends attention; video website video recommendation to provide users with the most likely to click on the video recommendation; in news portal content recommendation to provide users with the most informative news - personalized recommendation technology is the support of the Internet based intelligent technology. Personalized recommendation system has been the research and development of up to ten year, however the extensive use of latent variable methods make the personalized recommendation algorithm and recommendation results interpretation is still One of the most important problems in the academic circles, and are still not well reflected in the industrial application. For example, in many practical recommendation system. The algorithm only for the user to provide a list of recommended personalized to the user as a result, to explain why to give this recommendation. The lack of explanation can be recommended to reduce the recommendation of the credibility of the results, and then affect the recommendation system actual effect. Considering the application of recommendation system and wide influence, can explain the research recommendation has its importance and urgency. In this paper, from three aspects of our data model and the economic significance of recommender system research to explain, the main contributions are as follows: 1. the interpretability of data: data input is the first step of personalized recommendation system, and the user item rating matrix is a personalized recommendation algorithm, especially Is the data input form of personalized recommendation algorithm based on the matrix decomposition. This paper proposed a decomposition framework of localization matrix block diagonal matrix based on bilateral, parallel and its application in matrix decomposition. The traditional matrix decomposition algorithm for the original matrix as a whole decomposition and prediction, and the lack of the internal structure of the matrix understand. In this work, we propose a bilateral block diagonal structure of the matrix, and prove theoretically that mathematical equivalence of community discovery algorithm of the structure and the two plans, thus explaining the matrix inner community structure and community relations. Based on community structure, we further put forward the localization of the matrix decomposition framework. And it is proved in theory and the traditional matrix decomposition algorithm compatibility, thus provides a unified framework for parallelizing common matrix decomposition algorithm, to improve prediction precision At the same time of increasing the computing efficiency of the.2. model can explain: Based on user item rating matrix, personalized recommendation model of user preference modeling and personalized recommendation are put forward in this paper. The explicit variable analysis phrase level emotion model based on decomposition and dynamic modeling based on time series analysis of the latent variable model. Based on matrix decomposition due to its better prediction effect score and scalability, has gradually become the basis of personalized recommendation algorithm and is widely used in the actual system. However, due to the unknown variables in essence, latent variable model to the recommendation algorithm and recommendation results intuitively understandable explanation, thus reducing the recommendation the system reliability of the user. In this work, we use the phrase level sentiment analysis technology of extraction from large-scale user reviews of product attributes The expression of words and users in different attributes of emotion, and then introduce the explicit variable and proposed recommendation algorithm explicit variable decomposition model based on personalized, on the one hand makes the optimization process model has the intuitive meaning, recommendation results are given on the other hand can be explained in the model level and personalized recommendation reasons. Due to its periodic user preference in different attributes, we use time series analysis for dynamic modeling and prediction of user preferences, so as to realize the significance of dynamic time interpretability of the recommended recommended.3.. Economics recommendation system based on user behavior data and personal preference modeling, the personalized recommendation way implicitly to regulate commodity users the matching and purchase, thus affecting the economic efficiency of the system in the final level. In this paper, based on the total welfare maximization of the Internet system The specific implementation of the recommended framework is given in a typical scenario. With the traditional line activities constantly online, Internet applications are common can be formalized as "model of producer service consumers, for example in the electronic commerce website, online merchants (producers) to provide online commodity (service), and network the user (consumer) is to select and purchase in many commodities. Based on the basic definition of traditional economics, this paper first gives the Internet environment utility, the basic concept and the unified form of costs and benefits, and further gives a general method for calculating the total social welfare in Internet application. On this basis, we use the Internet service the distribution is the basic problem, put forward the recommended network welfare maximization based personalized framework. Further, based on the typical network application (e-commerce, online P2P lending. Crowdsourcing platform) in specific to the framework, and the network service recommendation and personalized evaluation. Experimental results show that this method can improve the total social welfare at the same time in providing high quality services for users to recommend, at the same time that enhance the user experience and enhance the social benefits.

【学位授予单位】：清华大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.3
，

本文编号：1464807