摘要 |
<P>PROBLEM TO BE SOLVED: To enable a reduction in the number of support vectors by preventing lowering of ranking accuracy. <P>SOLUTION: A support vector selection device comprises: a ranking function generation section 22 to learn weight for multiple support vectors in a ranking function and for each pair of the support vectors, on the basis of training data stored in a training data database 21; a retrieval evaluation section 25 to rank a retrieval result for each retrieval query, with respect to each pair of the support vectors, on the basis of a retrieval score to be calculated by the ranking function not using the support vector pair, and to calculate an influence degree score that shows an influence on a decrease in an evaluation index when not using the support vector pair, on the basis of the retrieval results ranked with respect to each retrieval query; and a support vector selection section 27 to set the weight for the support vector pair with a small influence degree score at 0. <P>COPYRIGHT: (C)2013,JPO&INPIT |