机器学习排序系统的设计与实现

发布时间：2018-07-09 17:50

本文选题：搜索 + 排序　；参考：《北京邮电大学》2013年硕士论文

【摘要】：本文首先对搜索引擎中的排序系统作了简要的分析,介绍了近期新出现的机器学习排序的概念,对其中样本标注所采用的众包标注技术的国内外研究情况进行了介绍。之后本文针对众包标注准确率低的问题提出了一套解决方法,包括了计算对标注数据投票的一致程度、引入用户点击数据来判别标注的正确性等。其中利用点击数据来判别标注正确性的方法包括基于单点点击特征的判别方法和基于配对点击特征的对比判别方法。最后本文设计并实现了一套将用户行为作为排序特征加入到排序模型中的系统。这些用户行为包括了用户在搜索引擎结果页上的点击行为和用户在浏览着陆页的行为等。同时本文还介绍了如何将用户行为描述为用户对相关性的感知及如何预处理用户行为数据。最后为了验证提高标注数据准确率的方法和加入用户行为特征的有效性,本文从目前主流的搜索引擎评估算法中选取了针对本文的评估策略,并且通过相应的实验证明了本文所述方法的有效性。
[Abstract]:In this paper, the sorting system in search engine is analyzed briefly, the concept of machine learning ranking is introduced, and the research situation of crowdsourcing tagging technology used in sample tagging is introduced at home and abroad. Then this paper puts forward a set of methods to solve the problem of low accuracy of crowdsourcing tagging, including calculating the consistent degree of voting for annotated data, introducing user click data to judge the correctness of tagging and so on. Among them, the method of using click data to judge the correctness of tagging includes single click feature based discrimination method and pairing click feature based comparison discrimination method. Finally, this paper designs and implements a system that adds user behavior to the sorting model as a sort feature. These user behaviors include the user's click behavior on the search engine results page and the user's behavior in browsing the landing page. This paper also introduces how to describe user behavior as user perception of correlation and how to preprocess user behavior data. Finally, in order to verify the effectiveness of the methods to improve the accuracy of tagging data and to add user behavior features, this paper selects the evaluation strategy for this paper from the current mainstream search engine evaluation algorithm. The effectiveness of the proposed method is proved by corresponding experiments.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.3;TP181

【相似文献】