Towards Intelligent Information Retrieval:Ensemble Ranking SVM,Constraint Adaptive Propagation,and Interactive Image Retreval

This thesis deals with three main research topics in machine learning based information retrieval:learning to rank, semi supervised kernel matrix learning, and user interactive image retrieval.Learning to rank have been successfully apply into document retrieval. Ranking SVM is a powerful technology which has been commonly used for learning to rank. However, in Ranking SVM, the training time of generating a train model grows exponentially when the size of the training set is large. To solve this problem and improve the ranking accuracy, this thesis proposed to introduce the ensemble learning method into learning to rank, Ensemble Ranking SVM not only greatly improves the efficiency of the model training, but also achieves high ranking accuracy.Semi-supervised kernel matrix learning (SS-KML) aims to learn a kernel matrix from the given samples which contain just a little supervised information such as class labels or pairwise constraints. This thesis provides a novel Constraint Adaptive Propagation (CAP) method, in which two adaptive fidelity terms are designed to overcome the information deficiency problem that occurs occasionally in two representatives SS-KML methods PCP and KP. CAP outperforms state-of-the-art SS-KML methods such as PCP and KP in terms of effectiveness and efficiency.Traditional region based image retrieval technologies are hard to focus on the user actual objective. In order to solve the problem, this thesis presents a novel user interactive image retrieval system. Our key idea is to extract the query objective from query image through a little help from user by a new foreground extraction method, then retrieval the image database, and CAP is introduced to improve the performance of retrieval system through active learning the user relevance feedback information. The experimental results show that, the retrieval images more relevance to the query image region.