| 摘要 |
<P>PROBLEM TO BE SOLVED: To allow detecting a misclassified sample in a general classification problem with a plurality of categories by restraining adverse effects due to a misclassified sample on learning in a statistical classifier used for the detection. <P>SOLUTION: A misclassification detecting device comprises: a probability model generator 22 that calculates a parameter value Θ of a joint probability model p(x<SB POS="POST">n</SB>, y<SB POS="POST">n</SB>) for contents x<SB POS="POST">n</SB>and category y<SB POS="POST">n</SB>of each sample n by using a weight w<SB POS="POST">n</SB>set for each sample n, in such a manner as to maximize predictive likelihood of the joint probability model p(x<SB POS="POST">n</SB>, y<SB POS="POST">n</SB>) based on Leave-One-Out Cross-Validation; and a misclassified sample detector 24 that calculates a predicted class posterior probability P(y<SB POS="POST">n</SB>¾x<SB POS="POST">n</SB>) of the category y<SB POS="POST">n</SB>into which the contents x<SB POS="POST">n</SB>is classified on the basis of the parameter value Θ of the joint probability model p(x<SB POS="POST">n</SB>, y<SB POS="POST">n</SB>) calculated by the probability model generator 22, for each sample n, and detects a misclassified sample on the basis of the predicted class posterior probability P(y<SB POS="POST">n</SB>¾x<SB POS="POST">n</SB>) of each sample n. <P>COPYRIGHT: (C)2013,JPO&INPIT |
