A hybrid heuristic algorithm for evolving models in simultaneous scenarios of classification and clustering

摘要

Machine Learning is currently an important research field that attracts interest due to its importance for discovering hidden knowledge or patterns from big datasets. In this paper, we propose a heuristic algorithm which can solve problems related to only classification, only clustering, or classification with clustering by creating models with the ability to evolve to another class/cluster configuration without a retraining process for new incoming data. This algorithm combines supervised and unsupervised learning principles for the incremental construction of both classes and clusters, by using the main guidelines from two classical methods of classification based on distance and clustering based on prototypes, such as KNN and K-means. The algorithm is able to deal with labeled and unlabeled samples as inputs in order to create new groups (classes or clusters), merge or reconfigure existing ones. Basically, the creation of new groups follows three sequential steps: (i) locate the provisional group for an input sample using K-means, (ii) using 1NN, locate the nearest sample to the input sample, only considering the samples in the provisional group, and (iii) merge or reconfigure existing groups following specific guidelines. Several benchmarks, related to classification and clustering problems, were evaluated by our proposal; the results were compared with classical algorithms. On the other hand, artificial datasets with labeled and unlabeled samples have been created to show the ability of our algorithm in the hybrid context to solve classification and clustering combined. As a result, the algorithm is able to create clusters and classes, simultaneously, when required. Finally, a real case study of fault diagnosis in rotating machinery is presented for discovering new groups that might represent patterns from unknown data.