Optimizing a Cost Matrix to Solve Rare-Class Biological Problems

In a binary dataset, a rare-class problem occurs when one class of data (typically the class of interest) is far outweighed by the other. Such a problem is typically difficult to learn and classify and is quite common, especially among biological problems such as the identification of gene conversions. A multitude of solutions for this problem exist with varying levels of success. In this paper we present our solution, which involves using the MetaCost algorithm, a cost-sensitive “meta-classifier” that requires a cost matrix to adjust the learning of an underlying classifier. Our method finds this cost matrix for a given dataset and classification algorithm, creating a final classification model. Through a detailed description, a basic evaluation, and the application to the problem of identifying gene conversions, we show the effectiveness of this approach. Our novel approach to generating a cost matrix has proven to be quite effective in the identification of gene conversions and represents a robust way to tackle the rare-class data problem.