Automatic design of interpretable fuzzy predicate systems for clustering using self-organizing maps

• We present a method for interpretable data clustering defined by means of fuzzy predicates.
• Fuzzy predicates are self-discovered by using data information.
• Self-organizing maps are used to obtain knowledge for grouping criteria, analyzing characteristics of the variables in each data group.
• In a posterior analysis, predicates allow linguistic interpretation for the discovered clustering.
• The methodology is validated with labeled databases to show the accuracy obtained, and some linguistic interpretation examples are given.

In the area of pattern recognition, clustering algorithms are a family of unsupervised classifiers designed with the aim to discover unrevealed structures in the data. While this is a never ending research topic, many methods have been developed with good theoretical and practical properties. One of such methods is based on self organizing maps (SOM), which have been successfully used for data clustering, using a two levels clustering approach. Newer on the field, clustering systems based on fuzzy logic improve the performance of traditional approaches. In this paper we combine both approaches. Most of the previous works on fuzzy clustering are based on fuzzy inference systems, but we propose the design of a new clustering system in which we use predicate fuzzy logic to perform the clustering task, being automatically designed based on data. Given a datum, degrees of truth of fuzzy predicates associated with each cluster are computed using continuous membership functions defined over data features. The predicate with the maximum degree of truth determines the cluster to be assigned. Knowledge is discovered from data, obtained using the SOM generalization aptitude and taking advantage of the well-known SOM abilities to discover natural data grouping when compared with direct clustering. In addition, the proposed approach adds linguistic interpretability when membership functions are analyzed by a field expert. We also present how this approach can be used to deal with partitioned data. Results show that clustering accuracy obtained is high and it outperforms other methods in the majority of datasets tested.