以資料選擇技術幫助大規模支持向量機排序

Abstract

Learning to rank has become a popular research topic in several areas including information retrieval and machine learning. Pair-wise ranking, which learns all the order preferences between every two examples, is one typical method for solving the ranking problem. In pair-wise ranking, RankSVM is a widely-used machine learning algorithm and has been successfully applied to the ranking problem in the previous work. However, RankSVM suffers a critical problem which is the long training time because of the huge number of pairs. In this thesis, we propose a data selection technique, Pruned RankSVM, that selects the most informative pairs before training. If we use partial pairs instead of total ones, we can train a large-scale data set efficiently. In the experiment, we show the performance of Pruned RankSVM is overall comparable with RankSVM while using significantly fewer pairs. To show the efficiency of Pruned RankSVM, we also compare with one point-wise ranking algorithm : support vector regression. Experimental results demonstrate that Pruned RankSVM outperforms support vector regression on most data sets.