关键词: 多示例学习, 性别识别, 期望最大化多样性密度, 示例近邻, k近邻

Abstract: In order to avoid a large number of calculations for getting the threshold of classifier and improve the performance of gender identification, an approach of gender identification based on improved diverse density is presented. The new method puts male and female voice bags in double classes and labels many times. It uses EM-DD algorithm for multi-instance learning and then gets two diverse density points as a double-points model. It puts forward instance-k neighbor classification algorithm for classification by selecting multiple instances. The method considers the influences both of male and female voice samples to the unknown voice bag, doesn’t need to set the threshold and reduces the influences of outlying instances, and it improves the efficiency of the system.

Key words: multi-instance learning, gender identification, Expectation Maximization Diverse Density（EM-DD）, instances-nearest neighbour, k-nearest neighbour