RT-GSOM: Rough tolerance growing self-organizing map

Abstract

The concept of rough tolerance set is introduced within growing self-organizing map (GSOM) to reduce the uncertainty in decision-making by developing a new algorithm, namely rough tolerance GSOM (RT-GSOM). This algorithm aims to address the issues of (i) identifying the suitable size of clusters in SOM, (ii) information loss in rough SOM (RSOM), and (iii) uncertainty arising from the overlapping patterns of decision classes in GSOM. In RT-GSOM, the network is allowed to grow based on indiscernible reducts and tolerance thresholds extracted from data in an unsupervised way. The network is initialized with the samples extracted from these reducts, and the weights are initialized with random category index. For each decision class, one set of indiscernible reducts is obtained. The tolerance threshold for each decision class is defined using the average distance among all the samples present in the reduct set corresponding to the same class. The superiority of RT-GSOM is demonstrated over twelve benchmark datasets (both categorical and continuous) obtained from UCI machine learning repository. Results reveal that RT-GSOM is efficient than some state-of-the-art algorithms in terms of learning rate, and quality of clusters for both categorical, and continuous data.