Learning to rank is a fast growing research problem in Machine Learning and Information Retrieval. Ranking Support Vector Machine (RSVM) is a widely adopted ranking method in various fields due to its good generalization performance. RSVM transforms the learning to rank problem into a classification problem, and employs a single hyperplane to separate the instances. Recently several ranking methods have been proposed based on RSVM. These methods employ multiple hyperplanes so that a local ranking is produced from each hyperplane. Rank aggregation is then conducted to combine the local rankings. However, these methods do not fully utilize the information from the individual hyperplanes. In this thesis, we address the problem of aggregating the rankings using the SVM output values and pro...[ Read more ]

Learning to rank is a fast growing research problem in Machine Learning and Information Retrieval. Ranking Support Vector Machine (RSVM) is a widely adopted ranking method in various fields due to its good generalization performance. RSVM transforms the learning to rank problem into a classification problem, and employs a single hyperplane to separate the instances. Recently several ranking methods have been proposed based on RSVM. These methods employ multiple hyperplanes so that a local ranking is produced from each hyperplane. Rank aggregation is then conducted to combine the local rankings. However, these methods do not fully utilize the information from the individual hyperplanes. In this thesis, we address the problem of aggregating the rankings using the SVM output values and propose a novel rank aggregation framework based on a probabilistic view. In this framework we define two rank aggregation methods and conduct experiments to show the improvement achieved by utilizing the SVM output values.