Frequent Pattern Discovery Without Binarization: Mining Attribute Profiles

Frequent pattern discovery has become a popular solution to many scientific and industrial problems in a range of different datasets. Traditional algorithms, developed for binary (or Boolean) attributes, can be applied to such data with a prerequisite of transforming non-binary (continuous or categorical) attribute domains into binary ones. As a consequence of this binarization, the discovered patterns no longer reflect the associations between attributes but the relations between their binned independent values, and thus, interactions between the original attributes may be lost. In this paper we propose to overcome this limitation by introducing the concept of mining frequent

attribute profiles

that describes the relationships between the original attributes. By this concept, previously hidden interactions can be discovered and redundant patterns that are identified by traditional methods are eliminated. A novel algorithm, called MAP, has been developed for

m

ining

a

ttribute

p

rofiles that can be potentially applied to diverse data domains. The effectiveness of the proposed method is shown by using gene expression or microarray data.