Dealing with different distributions in learning from positive and unlabeled web data

Abstract

In the problem of learning with positive and unlabeled examples, existing research all assumes that positive examples P and the hidden positive examples in the unlabeled set U are generated from the same distribution. This assumption may be violated in practice. In such cases, existing methods perform poorly. This paper proposes a novel technique A-EM to deal with the problem. Experimental results with product page classification demonstrate the effectiveness of the proposed technique.