Attributes for Understanding Groups of Binary Data

In this paper, we are interested in determining relevant attributes for multi-class discrimination of binary data. Given a set of observations described by the presence or absence of a set of attributes and divided into groups, we seek to determine a subset of attributes to explain and describe these groups. A pattern is set of Boolean values that are shared by many observations in a given group. Thanks to a new pattern computation algorithm, we present an approach to optimize the choice of the important attributes. Using real biological instances, we compare our results with two other different approaches and discuss the difference in information obtained by each.