Improve Web Search Ranking by Co-ranking SVM

Learning to rank technique is attacking much attention in search engine optimization. However, it cost a lot to collect labeled data for rank learning. In addition, the features of learning better ranking functions, such as click-through features and social annotation features, proposed from different viewpoints. In this paper, we propose to consider the learning to rank problem in the co-training framework, which can gather information from different types of features and incorporate unlabeled data into training. A bottleneck of considering rank learning in co-training framework is that the single view Web features always fail to give accurate ranking results due to their limitations in real tasks. For instances, sparseness of social annotation and the bias of user click-through may make them fail. To solve this bottleneck, we propose a feature fusion algorithm that enables the features from different views to enhance each other before co-training. Though the independencies assumption might be violated due to feature fusion, we introduce a sample selection strategy which guarantees co-training to work effectively. Experimental results on real search log and social annotations show that our proposed method can effectively improve the ranking performance by utilizing our feature fusion and sample selection strategies.