Selecting Interesting Articles Using Their Similarity Based Only on Positive Examples

Abstract

The task of automated searching for interesting text documents frequently suffers from a very poor balance among documents representing both positive and negative examples or from one completely missing class. This paper suggests the ranking approach based on the k-NN algorithm adapted for determining the similarity degree of new documents just to the representative positive collection. From the viewpoint of the precision-recall relation, a user can decide in advance how many and how similar articles should be released through a filter.