Subgroup Discovery on Multiple Instance Data

Abstract

To date, the subgroup discovery (SD) task has been considered in problems where a target variable is unequivocally described by a set of features, also known as instance. Nowadays, however, with the increasing interest in data storage, new data structures are being provided such as the multiple instance data in which a target variable value is ambiguously defined by a set of instances. Most of the proposals related to multiple instance data are based on predictive tasks and no supervised descriptive analysis can be provided when data is organized in this way. At this point, the aim of this work is to extend the SD task to cope with this type of data. SD is a really interesting task that aims at discovering interesting relationships between different features with respect to a specific target variable that is of interest for the user or the problem under study. In this regard, this paper presents three different approaches for mining interesting subgroups in multiple instance problems. The proposed models represent three different ways of tackling the problem and they are based on three well-known algorithms in the SD field: SD-Map (exhaustive search approach), CGBA-SD (Comprehensible Grammar-Based Algorithm for Subgroup Discovery) and NMEEF-SD (multiobjective evolutionary fuzzy system). The proposals have been tested on a wide set of datasets, including 10 real-world and 20 synthetic datasets, aiming at describing how the three methodologies behave on different scenarios. Any comparison is unfair since they are completely different methodologies.