Block-Based Approaches to Learning Ranking Functions with Application to Protein Homology Prediction

In many information retrieval systems such as Web search engines and biological-sequence search engines, the ranking functions that list the search results in order of their relevances to the query are one of the most important components. In the machine learning approaches to constructing ranking-functions, the feature vectors of database items are computed based on queries and thus they are grouped into blocks by queries. However, few existing algorithms take into account the block structure of data when learning a ranking function. This paper describes a series of approaches for more accurate learning of ranking functions by exploiting the block structure of data and applies these approaches to the protein homology prediction problem, a key step of protein structure prediction in bioinformatics. These approaches range from data normalization and reduction to learner training. The data reduction methods, including block selection and support vector under-sampling, contributed to our original winning of the protein homology prediction task in the ACM KDDCUP-2004 competition. By extending the block-selection method to a query-adaptive version and using an ensemble-learning approach, a novel ranking-function learning algorithm named K-Nearest-Block (KNB) Ensemble Ranking is proposed. In this algorithm, given the data block derived from a new query, only those most similar data blocks in training data are used to learn a ranking function. Experiments with the support vector machine (SVM) used as the benchmark learner demonstrate that all the proposed block-based approaches can significantly improve the ranking performance of SVMs. Especially, the KNB SVM ensemble performs so far most accurately overall on the blinded test data set of the KDDCUP-2004 protein homology prediction problem.