Efficient Error Setting for Subspace Miners

A typical mining problem is the extraction of patterns from subspaces of multidimensional data. Such patterns, known as a biclusters, comprise subsets of objects that behave similarly across subsets of attributes, and may overlap each other, i.e., objects/attributes may belong to several patterns, or to none. For many miners, a key input parameter is the error used which greatly affects the quality, quantity and coherency of the mined clusters. As the error is dataset dependent, setting it demands either domain knowledge or some trial-and-error. The paper presents a new method for automatically setting the error to the value that maximizes the number of clusters mined. This error value is strongly correlated to the value for which performance scores are maximized. The correlation is extensively evaluated using six datasets, two mining algorithms, seven prevailing performance measures, and compared with five prior literature methods, demonstrating a substantial improvement in the mining score.