Research On Algorithm Of Learning To Rank With Ties

With the rapid development of internet technology,computers and networks not only have been integrated into our daily learning,work and life,but also have become an important way for people to acquire knowledge and understand the world.Hence it is important to use search engines to efficiently retrieve relevant information from large amounts of data.The ranking is essential issue in information retrieval,which plays an important role in accurate recommendation of retrieving information.In current learning to rank problems,the learned ranking function sorts objects according to their predicted scores.It is unreasonable to obtain a full-ordering object list only according to preference relationships if two or more objects have almost identical degrees of relevance(or called objects with ties).Therefore,learning to rank with ties is developed in this thesis,and the learned ranking function can judge both the preference and ties relationships among objects to be indexed.The main works are summarized as follows:(1)Learning to rank with ties algorithm based on paiewise is studied.Based on the softmax loss function,the Euclidean loss function and the Bradley-Terry model,three ranking learning algorithms are designed,which contains ties in both training and testing process.The learned ranking function can be used to rank objects containing ties.(2)Learning to rank with ties algorithm based on subsequences is stedied.The subsequences ranking with ties algorithm is proposed based on Bradley-Terry model,which contain rich structural ranking information by increasing the length of subsequences.The algorithms mentioned above are applied to image re-ranking.By combining the proposed algorithms with the deep learning framework,two tasks are implemented over two publicly available image datasets.The experimental results demonstrate that the proposed approaches can rank images containing ties effectively.