An Improvement Of Ranknet Learning Sorting Algorithm

With the rapid development of information technology,the use of Search Engin access to network resources is the way of life.At the same time,the massive web page information on the search,engine to bring great challenges,such as how quickly and accurately from the ocean of information to find the user want information,how to first show the most useful information in the user search results.The key factor in measuring SE performance is the sorting algorithm.Early sorting algorithm to consider the sorting factor is relatively simple,while retrieving the accuracy of the results is difficult to guarantee.With the continuous development of artificial intelligence,in recent years,the study of machine learning and sorting learning has also been widely regarded by experts at home and abroad,sorting learning algorithm IR,collaborative filtering,NLP,emotional analysis,online advertising,system recommendation and other fields Important role,and more and more artificial intelligence scholars as a hot research direction.This paper aims to study the RankNet neural network learning sorting algorithm,which is mainly proposed by Chris Burges et al.,And is widely used in related search engines.Through the RankNet neural network research to improve the user experience of web search results.To sum up,the focus of the paper include the following three points:(1)This paper studies the evolution of the sorting algorithm and the current research situation,and makes a general description of the Learn to Rank algorithm.The evaluation criteria and optimization direction of the web search sorting algorithm are studied in order to guide the RankNet The algorithm does two improvements and optimizations.(2)the first point of improvement:focus on the cross entropy and the variance of the linear combination of the RankNet algorithm to improve the loss of the original algorithm to solve a sample on the two documents and inquiries related to the size of the problem is ignored;Two improvements:by increasing the weight of the query to solve the different query corresponding to the number of documents vary greatly when the learning process misleading,making the algorithm trained model more accurate.(3)Finally,the BP neural network model is used to verify and compare the RankNet and the improved algorithm in the data set of Microsoft.The algorithm of different sorting algorithm is used to analyze the algorithm before and after the improvement.And the principle of gradient descent is analyzed,and the correctness of the loss function of RankNet algorithm is proved.