Design And Implementation Of A Learning To Rank System

This paper firstly analysis the ranking system of modern search engine, then present an introduction of learning to rank concept. Then give an introduction of the crowdsourcing label concept which used to produce the label data for a learning to rank system. Secondly, this paper will mark some solutions to the problem of low accuracy of crowdsourcing labels, which include voted unanimously to the extent of the data and use user click-through data to improve label correctness. Method that use user click-through data based on a single point click features discriminant and a pair wise comparison features discriminant. Further more, this paper propos some user behavior as ranking factors and add to the learning to rank system. Those user behaviors include click behaviors on the result page and behaviors on the landing page. Besides, this paper will also introduce how to describe user’s behavior as a web page’s quality, and how to process user data. At last, to demonstrate the effectiveness of the accurary control algorithm and the user behavior ranking facotrs, this paper will choice a ranking evaluate algorithm from modern search engine evaluate algorithms. We take several experiments to demostrate the effectiveness of the algorithm which be proposed in this paper.