Research On Learning To Rank Algorithm Based On Machine Learning

Nowadays,in the era of big data,the Internet is flooded with a large number of data resources.People need to efficiently obtain useful information from complex and redundant resources.In order to achieve this demand,the importance of information retrieval technology is becoming more and more obvious.In addition,machine learning technology has involved almost all fields of human beings.In this context,machine learning algorithms and information retrieval models are combined in the process of information retrieval to obtain more efficient and accurate sorting algorithms.This algorithm is called Learning to Rank(LTR),and the efficiency of the learning to rank algorithm fundamentally affects the performance of information retrieval.The Matthew Effect is a desirable phenomenon in information retrieval systems,however most rank learning algorithms do not take the Matthew effect into account.Lambda MART is a well-known LTR algorithm that can be further optimized based on the Matthew effect.Therefore,the main research work of this thesis is as follows:(1)The Matthew effect refers to making the effective more effective and the ineffective more ineffective.The Matthew effect is a very desirable phenomenon in the ranking model.A large number of user clicks indicates that the information is desired by the user,and a user with few clicks indicates that the information is not what the user wants.Inspired by the Matthew effect,we distinguish queries with different validity and then assign higher weights to queries with higher validity.We improve the gradients in the Lambda MART algorithm to optimize queries efficiently,i.e.to highlight the Matthew effect of the generated ranking model.(2)This thesis also proposes a strategy for evaluating the current ranking model and a strategy for dynamically reducing the learning rate.The purpose of the former is to enhance the Matthew effect of the ranking model,and the purpose of the latter is to speed up the convergence of the loss function and improve its effectiveness.(3)In this thesis,the Gini coefficient,mean,variance and information retrieval evaluation indicators are combined to measure the Matthew effect of the ranking model.The improved Lambda MART algorithm was tested on two datasets respectively.Compared with the original algorithm and other state-of-the-art LTR algorithms,the ranking model generated by the improved Lambda MART algorithm can show stronger Matthew effect and higher efficiency.