电子商务个性化推荐系统研究

【作者】 蒋翀；

【导师】 费洪晓；

【作者基本信息】 中南大学 ， 计算机应用技术， 2009， 硕士

【摘要】 随着电子商务相关技术的日益完善,越来越多的人们接受了网络购物这一新兴消费方式,但个人相对稳定的购买需求与网站所提供纷繁庞杂商品之间存在矛盾,如何解决这一矛盾成为各国研究人员和电子商务参与者关注的问题。本文主要研究应用个性化技术的推荐系统,着力于探讨如何更好的利用用户有意识或无意识反映出的偏好信息,为个性化推荐原型系统服务。分析了目前普遍采用的显式和隐式两种用户偏好获取方式的优缺点,提出了一种混合用户偏好获取模式。在客户端利用嵌入浏览器的脚本语言获取用户浏览行为信息,并通过评分转化规则得到大量反映用户偏好的隐式评分,弥补显式评分数据稀疏的缺陷。将隐式评分和相对准确的显式评分作为用户兴趣模型更新的数据来源。同时,针对电子商务参与者兴趣变化异常频繁的特点,提出了基于线性衰减的用户兴趣模型。根据注册信息和浏览行为建立初始兴趣模型,考虑到用户兴趣项目经常变化的特点,构建了链式向量空间模型表示的兴趣模型。兴趣模型中的用户评分在固定时间间隔t进行自然衰减直至变为0而被淘汰,若在衰减过程中产生了新的访问记录和显式反馈,评分将更新为新的评分并继续参与到衰减过程,构成了用户兴趣模型的主要更新过程。基于协作过滤算法的推荐系统原型对相似度计算、邻居集体积大小和协作过滤推荐算法进行了评估测试,验证了系统的可行性。利用嵌入了特殊脚本的浏览器对用户页面停留时间、鼠标点击次数和页面滚动时间三种可能反映用户偏好的行为进行了相关性测试,结果表明页面停留时间和滚动时间与用户偏好度关联紧密,鼠标点击次数并未表现出明显关联。更多还原

【Abstract】 With the popularity of e-commerce and related technologies improving, more and more people accept the consumption patterns of net purchases, but the contradiction between the relative stability of personal needs and the numerous and complex goods becomes increasingly sharp, and how to resolve this problem become a hot spots in research. The personalized recommender systems are regarded as the study object. The focus is the ways to better use of preference that the user has consciously or unconsciously reflect, for the personalized recommendation service prototype system.The advantages and disadvantages of the explicit and implicit user preferences access are analyzed and a mixed-mode preference access pattern is proposed. Users’ browsing behavior is collected by the client browser and is converted to the ratings by transformation rules, which can make up for sparse explicit data. The mixed-model combines the implicit and explicit ratings to express user profile. For such extremely frequent changes in user interest, a linear-attenuation based user profile is proposed. The registered information and browsing behaviors are used to set up the initial user profile. Because of the frequent change of user interests, use profile is represented by a chain vector space model. User ratings decrease at a fixed time interval t until the rating has been eliminated to zero. User profile updates rating by a new record of visits and explicit rating, and then continues to be involved in the attenuation process. These constitute a model of user profile in the main update process.Collaborative filtering based recommendation prototype system test the similarity algorithms, the neighbor set size and collaborative filtering algorithms, which verify the feasibility of the system. The special script embedded browser test the user residence time, mouse clicks and page scrolling time that may reflect user preferences. The test results show that the residence time and scroll the page of time associated with the user preference closer and mouse clicks don’t show an obvious correlation.更多还原