Research And Application Of Music Retrieval On Learning To Rank

With the development of mobile Internet and multimedia technology,multimedia resources are facing explosive growth,and mobile music is widely concerned by researchers in its unique form.In the face of massive mobile music,how to understand users' information needs and provide satisfactory ranking results is the focus of music retrieval.Music retrieval belongs to the field of information retrieval.Although there are mature research theories and achievements in the field of information retrieval,there are few achievements in music ranking research.Music retrieval system is different from other information retrieval systems.It has strong color in the field.For music ranking,there are a lot of relevance features that need to be considered,for example,whether the song is remix and whether it is a cover version are both important ranking features.It is difficult to get satisfactory ranking results through simple retrieval model and artificial parameter adjustment to fit relevance formula.Aiming at the ranking problem of music retrieval system,this paper proposes the research and implementation of music retrieval method based on learning to rank.This paper mainly includes the following research work:1)design a hot model based on user behavior analysis.By considering the characteristics of user behavior,we calculate the hot of songs and singers,and design a strategy of hot caching to solve the problem of log sparsity.2)based on the business requirements of the music products and the user's information requirements,user query request parameters are constructed and the user query intention is expressed by the language that meets the syntax rules of specific query parser.A field centered retrieval is implemented,and a retrieval strategy of artificial rule fusion learning to rank is designed and implemented based on the results returned.3)to preprocess the real user log provided by a company,in view of the position bias and trust bias of user click,a click model based on the click times,click position and click trust of the song is proposed to dig out the implicit feedback information in user click log to generate high quality relevance annotation.4)from the perspective of user information demand and the operation demand of music products,the features of the query itself,the song itself and the query and the song are obtained through feature engineering,and the training set of learning to rank algorithm is generated.5)study and implement five classical learning to rank algorithms of Rank Net,List Net,Rank SVM,Lambda MART and Rank Boost,and verify the performance differences of different learning to rank algorithms on different experimental data sets.Based on the above research,this paper designs and implements a music retrieval system based on learning to rank,and verifies the practical application value of the research results through the online effect of a music product.