Database Classification by Hybrid Method combining Supervised and Unsupervised Learnings

This paper presents a new hybrid learning algorithm for unsupervised classification tasks. We combined Fuzzy c-means learning and the supervised version of Minimerror to develop a hybrid incremental strategy allowing unsupervised classifications. We applied this new approach to a real-world database in order to know if the information contained in unlabeled signals of a Geographic Information System (GIS), allow to well classify it. Finally, we compared our results to a classical classification obtained by a multilayer perceptron. keywords : Minimerror, hybrid method, classification, unsupervised learning, Geographic Information System. I. SUPERVISED AND UNSUPERVISED LEARNINGS For a classification task, the learning is supervised if the labels of the classes of the input patterns are given a priori by a professor. A cost function calculates the difference between desired and real outputs produced by a network, then, this difference is minimized modifying the network’s weights by a learning rule. A supervised learning set L is constitued by P couples (~ ξ, τ), μ = 1, ..., P , where ~ ξ is the input pattern μ and τ = ±1 its class. ~ ξ is a N -dimension vector, with numeric or categoric values. If labels τ are not present in L, it may be used as unsupervised learning. Learning is unsupervised when the object’s class is not known in advance. This learning is performed by extraction of intrinsic regularities of patterns presented to the network. The number of neurons of the output layer corresponds to the desired number of categories. Therefore, the network develops its own representation of input patterns, retaining the statistically redundant traits. II. SUPERVISED MINIMERROR Minimerror algorithm [1] performs correctly in binary problems of high dimensionality [3], [4], [10]. The supervised version of Minimerror performs a binary classification using the minimization of the cost function: