Learning To Rank Algorithm Based On Sparse Representation

The emergence of search engines let people gain relevant information in the Internet, Hence sorting of retrieved results is crucial.The World Wide Web WWW (World Wide Web) is a huge, global distribution of the information service center, developing in a fast ;peed.Compared with the traditional document on the WEB document, there are a lot of new features, they are distributed, heterogeneous, no structure or half structure, it is for i new challenge to the traditional information retrieval technology,nformation Retrieval is a procedure of oganizing information and finding out the information which users need. The key problem of it is how to create a retrieval model that can sort documents (web pages) based on their relevance to the given query. Early retrieval models are very simple and only have a few parameters to tune, but their precision is low which cause the user dissatisfied with the retrieval results. To solve this problem, recently, methods of learning to rank have been applied to the retrieval model construction in order to get promising results. The so called learning to rank is to automatically create a retrieval (ranking) model by using labeled training data and machine learning techniques. For its wide usage in the area of document retrieval and collaborative filtering, the research of learning to rank has received more and more attentions, and become a hot research field in machine learning.The traditional information retrieval model is divided into two main categories:one is related to the query method based on content, such as based on word frequency statistics;the method of probability model and language model;Another kind is with the query independent method based on link analysis, such as PageRank,HITS algorithm.The core of the two methods is ranking the web pages."learning to rank" making it possible to the mix of retrieval model; training the new ranking model can effectively improve the effect of retrieval.Learning to rank belongs to the field of information retrieval and machine learning, according to the inputspace, it can be divided into three categories:single document level (pointwise),document pairs for level (pairwise)and the list of documents (listwise). Loss function a ranking method is used to construct the objective function learning.The quality of the loss function will eventually impact on the performance of ranking unction.The structure of this paper is, first, put forward the background of the thesis, Then introduces respectively the traditional information retrieval models and the learning to rank algorithm.Then analyzed the advantages and disadvantages of the ranking algorithm, put forward the improvement methods of the loss function to improve the pairwise and listwise, through the way of combining different loss functions.On the basis of the previous step to add sparse representation to simplify the calculation.