基于用户行为的信任感知推荐方法研究

发布时间：2018-09-13 21:33

【摘要】：推荐为用户利用网络资源提供了极大的便利。按照输入信息是否清晰明确,可以将推荐分为两类:(1)基于关键词的推荐,(2)基于用户潜在行为和关系的推荐。最常用的基于关键词的推荐即搜索。用户输入关键词,搜索引擎返回与该关键词最接近的搜索结果。然而传统搜索引擎主要从提高搜索结果的查全率和查准率提升用户对搜索结果的满意程度,忽视了区分背景不同的用户对相同关键词期望的搜索结果也不同。单纯提高搜索引擎的查全率和查准率已经不能为用户提供满意的推荐,这是由于基于关键字的搜索包含的信息有限或者缺乏对用户隐含需求的挖掘,导致搜索结果不能满足用户的个性化需求。因此满足用户的个性化需求是提高搜索质量的关键因素之一。本文认为提高个性化搜索质量主要取决于挖掘用户隐含信息的准确度,以及实时地反映用户偏好变化。本文通过对搜索行为的系统研究,提出从以下几个方面改进搜索质量:(1)提出基于用户行为的个性化搜索结果预测方法,通过分析用户的历史访问行为,建立用户行为和偏好的隐马尔可夫模型(HMM),预测用户的搜索偏好,实现用户个性化搜索。为了提高该方法的效率,通过将相似的用户聚类,减少了估计HMM参数的时间,得到一种效率较高的个性化搜索方法。(2)研究网页排名对搜索质量的影响,针对网页通过彼此链接提高自身排名的现象,本文通过分析现有网页的拓扑结构,提出一种由网页提升系数识别并剔除排名异常提升网页的方法,有效地提高了搜索结果的质量。推荐过程中必需面对的另一个问题是当用户没有明确的需求时,如何为用户提供合理的推荐,帮助用户做出决定。本文通过对用户行为和信任关系的系统研究,提出从以下几个方面改进推荐质量:(1)研究如何为新加入用户提供合理的推荐,即用户的冷启动问题。由于受信任用户给出的推荐更可信,为了既扩展可信任的用户范围又保证扩展后的信任关系是可靠的,提出通过不信任关系约束信任关系的扩展,并基于扩展后的信任关系和用户对商品的评价信息为用户提供推荐,提高了对新加入用户和历史信息稀少用户的推荐效果。(2)探讨时间因素对推荐的影响,研究用户行为和其偏好变化之间的关系,提出一种描述用户偏好随时间而变化的推荐模型。并且将用户间相似度计算过程转化为二分图最优匹配,既保证了推荐算法的准确度,又降低了其时间复杂度。(3)针对部分存在潜在需求的商品很少有机会被关注的现象,提出一种由长尾分布约束的推荐方法,该方法首先基于用户行为确定用户间相似关系,然后合理地扩展用户间相似关系,最后通过长尾分布约束商品的推荐权重,解决了部分商品由于评价数量少而很难被用户关注、发现的问题。
[Abstract]:The recommendation provides great convenience for users to make use of network resources. According to whether the input information is clear or not, the recommendation can be divided into two categories: (1) recommendation based on keywords, (2) recommendation based on user's potential behavior and relationship. The most commonly used keyword-based recommendation is search. The user enters the keyword, and the search engine returns the search results closest to the keyword. However, traditional search engines mainly improve the recall and precision of search results, and ignore that users with different backgrounds expect different search results for the same keywords. Simply improving the recall and precision of search engines can not provide users with satisfactory recommendation, which is due to the limited information contained in keyword-based search or the lack of mining the implicit needs of users. As a result, the search results can not meet the personalized needs of users. Therefore, to meet the personalized needs of users is one of the key factors to improve the search quality. This paper holds that improving the quality of personalized search mainly depends on the accuracy of mining the implicit information of users and reflecting the change of user preferences in real time. Based on the systematic research on the search behavior, this paper proposes the following aspects to improve the search quality: (1) A personalized search result prediction method based on user behavior is proposed, and the historical access behavior of the user is analyzed. A hidden Markov model (HMM),) based on user behavior and preference is established to predict user's search preference and realize personalized search. In order to improve the efficiency of this method, the time of estimating HMM parameters is reduced by clustering similar users, and a more efficient personalized search method is obtained. (2) the influence of web page ranking on search quality is studied. Aiming at the phenomenon that web pages can improve their ranking by linking to each other, this paper analyzes the topology of existing web pages, and puts forward a method to identify and eliminate the abnormal ranking lifting pages by the lifting coefficient of web pages. The quality of search results is improved effectively. Another problem that must be faced in the process of recommendation is how to provide users with reasonable recommendation and help them to make decisions when there is no clear requirement. Based on the systematic study of user behavior and trust relationship, this paper proposes to improve the quality of recommendation from the following aspects: (1) study how to provide reasonable recommendation for new users, that is, the cold start problem of users. Because the recommendation given by trusted users is more credible, in order to extend the scope of trusted users and ensure that the extended trust relationship is reliable, this paper proposes to restrict the extension of trust relationships through distrust relationships. Based on the extended trust relationship and the evaluation information of the products, the recommendation effect of the new users and the users with few historical information is improved. (2) the influence of time factors on the recommendation is discussed. This paper studies the relationship between user behavior and their preferences, and proposes a recommendation model to describe the change of user preferences over time. And the similarity calculation process between users is transformed into bipartite graph optimal matching, which not only ensures the accuracy of the recommendation algorithm, but also reduces its time complexity. (3) aiming at the phenomenon that some products with potential needs have little chance to be paid attention to. A recommendation method constrained by long tail distribution is proposed. The method firstly determines the similarity relationship between users based on user behavior, and then reasonably extends the similarity relationship between users. Finally, the recommended weight of commodities is constrained by long tail distribution. It solves the problem that some commodities are difficult to be noticed and discovered by users because of their small quantity of evaluation.
【学位授予单位】：哈尔滨工程大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP391.3

【相似文献】