基于在线评论挖掘的网络购物混合推荐模型及策略研究

发布时间：2018-06-03 03:46

  本文选题：在线评论 + 网络口碑　； 参考：《江苏大学》2016年博士论文

【摘要】：随着web2.0的快速发展,网络数据几何增长,快速、准确获取用户需求信息成为相关企业和客户的迫切需要,各种各样的产品推荐系统应运而生。传统推荐系统大多是以产品为中心,以用户评分为数据源,系统性能依赖于用户偏好模型的质量,但用户的偏好信息很难以简单的评分来全面表征。因此,推荐系统常出现冷启动、数据稀疏性等问题。为了解决这些难题。本文提出以在线评论为基础数据源,在对传统网络购物推荐系统以及在线评论挖掘相关理论、技术、方法等进行梳理、分析的基础上:首先,作为网络口碑的主要传播形式,在线评论在已有文献的研究中大都先验地被当作外生变量来处理,并将两者间的动态相互关系简单的看成静态单方向作用。本文在动态内生性假说的视角下,引入在线评论各属性变量外的可测量的控制变量与难以观测或度量的哑变量,研究消费者基于在线评论的产品网络口碑感知问题。在动态面板数据模型中,通过控制变量与哑变量控制住内生性的影响后发现:(1)在静态分析框架下,在线评论与网络口碑感知之间是相互影响的,哑变量会对网络口碑感知与在线评论之间同时产生影响;(2)在动态分析框架下,在线评论与网络口碑感知之间存在一定的跨期作用,但滞后期并不确定,并且负面在线评论的比例与网络口碑感知之间并没有本文预期的反馈效应,这说明,二者的跨期动态作用并不是相互的而是单方向的。通过分析网络口碑感知的影响因素,确定在线评论的各属性对消费者网络口碑感知的不同影响,识别关键因素,为在线评论信息的差异化挖掘提供依据。其次,在以上分析的基础上,重点研究了在线评论的挖掘,包括在线评论数据源的挖掘以及在线评论信息的挖掘两部分。不同于以往的在线评论分析数据直接取自网络购物平台或专业点评网站,本研究将整个互联网作为在线评论的数据源,并从中挖掘可靠的数据源。通过将研究分解成三个子任务,对网络数据从运用改进的PageRank剔除作弊网页开始;再运用改进的TC-PageRank提炼与产品主题高度相关并包含大量在线评论数据的网页集;到运用改进的HITS确定在线评论分析数据来源的权威网页集结束。而对于在线评论信息的挖掘研究,在线评论作为潜在消费者网购的重要参考依据,挖掘其有价值的信息是有效利用的关键。针对网购平台的设计原则以及消费者的实际需求,融合社会化标注构建领域本体,基于领域本体的层次结构,将在线评论的特征词映射为本体概念,并利用Jess推理引擎提取评论中的隐性产品属性,再将概念间的层次关系映射到产品属性中,构建层次化产品属性集。基于层叠CRFs模型以及情感词典,从在线评论的极性分析到句子级的情感强度分析再到产品属性级的褒贬强度分析,实现在线评论情感倾向性的层次化分析。最后,针对网络购物推荐系统的数据稀疏性、冷启动问题日益突出以及传统的基于评分信息的用户偏好信息采集的不足,导致推荐算法的推荐效果不能令用户满意的问题,本文在上述网络口碑感知影响因素分析的基础上,提出了基于在线评论信息挖掘的用户偏好模型以及产品特征模型的构建方法。基于层次化的用户偏好信息,构建了基于本体建模方法的动态用户偏好模型,并通过用户偏好的更新本体进行用户偏好的增加、删减以及调整,时刻保持用户偏好本体的动态更新。在构建网络购物混合推荐模型之前,研究如何设计推荐系统才能获得用户的信任,进而实现推荐系统的预期作用。利用管理学、心理学、信息学等相关理论和方法,基于人际信任理论将用户对推荐系统的信任过程划分为从初始信任到交互信任再到推荐信任的三个阶段。探讨了影响各阶段信任的关键影响因素,并构建了多阶段用户信任的综合模型。基于该研究分析了用户信任及采纳推荐系统的关键影响因素,得出用户感知可信并采纳的推荐系统特征。在实证研究的基础上,根据Walls等提出的ISDT框架,分别从元需求与元设计两个方面对用户感知可信并采纳的推荐系统特征进行了详细阐述。根据上述研究结果,构建了基于在线评论挖掘的网络购物混合推荐模型,将推荐细化到产品特征层次并按消费者对产品的各特征评价进行综合排序。该模型以协同过滤算法为框架,结合基于内容推荐算法,通过产品的多属性评分来缓解稀疏性问题,并通过基于用户属性的相似度与基于产品属性的相似度计算算法在一定程度上解决了用户冷启动与产品冷启动问题;结合多种相似度算法构建了基于用户偏好与产品特征的混合推荐算法。仿真实验通过采集淘宝网、亚马逊中国网、京东网这三个国内大型网络购物平台的手机频道的10000条在线评论信息,验证了基于在线评论挖掘的网络购物混合推荐模型良好的推荐精确度以及解决冷启动问题的能力。并基于上述的研究结果,探讨了网络购物推荐系统的推荐策略以及网购平台在产品营销的管理实践中的主要对策建议。结合全文研究,总结归纳本文主要研究内容与贡献,并阐述本文不足之处以及对后续研究的展望。
[Abstract]:With the rapid development of Web2.0, the geometric growth, rapid and accurate acquisition of user demand information has become an urgent need for the related enterprises and customers. All kinds of product recommendation systems emerge as the times require. Most of the traditional recommendation systems are product centered, users are rated as data sources, and the performance of the system depends on the quality of user preference models. However, the user's preference information is difficult to be characterized in a simple way. Therefore, the recommendation system often appears cold start, data sparsity and other problems. In order to solve these problems, this paper proposes the online comment based data source, the traditional network shopping recommendation system and the line comment mining related theories, techniques, methods and so on. On the basis of analysis, first, as the main form of communication of internet word-of-mouth, online reviews are mostly treated as exogenous variables in the study of existing literature, and the dynamic relationship between them is simply regarded as static single direction. This paper introduces online reviews from the perspective of dynamic endogenous hypothesis and introduces online reviews to various genera. The measurable control variables outside the sex variables and the dumb variables that are difficult to observe or measure, study the consumer product network word-of-mouth perception based on online reviews. In the dynamic panel data model, the effects of control variables and dumb variables are found to control the endogenous effects: (1) online comment and the sense of internet word-of-mouth under the static analysis framework There is a mutual influence between knowledge and the interaction between the dumb variables and the online comments. (2) under the framework of dynamic analysis, there is a certain intertemporal effect between the online comment and the internet word-of-mouth perception, but the delay is not definite, and the proportion of the negative online comments and the internet word-of-mouth perception is not predefined. The feedback effect of the period shows that the dynamic role of the two parties is not mutual but single direction. By analyzing the influence factors of the online word-of-mouth perception, we can determine the different effects of the attributes of online comments on the perception of WOM, identify the key factors and provide the basis for the differential mining of online comment information. On the basis of the analysis, it focuses on the mining of online reviews, including the two parts of the online comment data source and the online review information. Different from the previous online comment analysis data, the online reviews are directly derived from the online shopping platform or the professional review site. Excavate reliable data sources. By decomposing the research into three subtasks, network data begins with an improved PageRank culling web page; the improved TC-PageRank is used to extract a web set which is highly related to the product theme and contains a large number of online commentary data; to use the improved HITS to determine the online review analysis data source In the study of online review information, online review is an important reference for potential consumer online shopping. Mining its valuable information is the key to effective use. In view of the design principles of online shopping platform and the actual needs of consumers, social annotation is used to construct domain ontology based on the domain. The hierarchical structure of ontology maps the feature words of online reviews into ontology concepts, and uses Jess reasoning engines to extract the hidden product attributes in comments, and then maps the hierarchical relationship between concepts to product attributes and constructs hierarchical product attributes. Based on cascading CRFs models and emotional dictionaries, the polarity analysis of online reviews to sentences is made. The level of emotional intensity analysis and the analysis of the appreciation intensity of the product attribute level to realize the hierarchical analysis of the emotional tendencies of online reviews. Finally, the recommendation algorithm is recommended in view of the data sparsity of the online shopping recommendation system, the increasingly prominent cold start problem and the insufficient information collection of the traditional users based on the score information. In this paper, the user preference model based on online comment information mining and the construction method of product feature model are put forward on the basis of the analysis of the factors affecting the network word of mouth perception. Based on the hierarchical user preference information, a dynamic user preference model based on the ontology modeling method is constructed. And the user preferences are added, deleted and adjusted to maintain the dynamic update of the user's preference ontology. Before constructing the mixed recommendation model of the network shopping, it is studied how to design the recommendation system to obtain the trust of the user, and then realize the expected function of the recommendation system. Based on the theory and methods of Informatics, based on the interpersonal trust theory, the trust process of the user to the recommendation system is divided into three stages from the initial trust to the interactive trust to the recommended trust. The key influencing factors that affect the trust in various stages are discussed, and a comprehensive model of multistage user trust is constructed. The users trust and adopt the key influencing factors of the recommendation system, and obtain the characteristics of the recommendation system that the user perceiving and adopting. On the basis of the empirical research, according to the ISDT framework proposed by Walls and so on, the characteristics of the recommendation system of the users' perceived trust and adoption are described in detail from the two aspects of the meta demand and the meta design. As a result, a mixed recommendation model of online shopping based on online review mining is constructed, and the recommendation is refined to the product feature level and comprehensive ranking according to the consumer's evaluation of the product features. The model is based on the collaborative filtering algorithm and combines the content recommendation algorithm to alleviate the sparsity problem through the multi attribute score of the product. With the similarity degree based on user attributes and similarity calculation based on product attributes, the problem of cold start and cold start of product is solved to a certain extent, and a hybrid recommendation algorithm based on user preferences and product characteristics is constructed in combination with a variety of similarity algorithms. The simulation experiment is carried out by collecting Taobao network, Amazon China network, and Beijing. The 10000 online comment information on the mobile channel of the three large network shopping platforms of the East Network validates the good recommendation accuracy and the ability to solve the cold start problem based on the online shopping mixed recommendation model based on online review mining. Based on the above research results, the recommendation strategy of the network shopping recommendation system is discussed. And the main countermeasures and suggestions of the online shopping platform in the management practice of product marketing. Combined with the full text research, the main research content and contribution of this article are summed up, and the shortcomings of this paper and the prospect of the follow-up research are expounded.
【学位授予单位】：江苏大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：F724.6
，

本文编号：1971342