Résumé
In this paper we introduce a new learning approach, which provides automated topological co-clustering based on Self-Organizing Map. The proposed approach (wd-TCoC) is computationally simple, learns a different feature’s weights vector for each prototype (relevance vector) and estimate the data density distribution on the map to produce an automatic clustering. The feature’s weights are computed for each observation during the learning process in order to attempt a clustering of both features and observations simultaneously. After the learning phase, a selection method with weight vectors is used to characterize features for each cluster. Compared to classical co-clustering algorithm, the features clustering in the proposed method is a fuzzy clustering, which allows a feature to belongs to different clusters and in this way to be able to eliminate the noisy features. The originality of the proposed method consists of 2 points: i) the co-clustering is made automatically, i.e. there is no parameter to fix to cluster the data and the features and ii) the features are fuzzy clustered. We illustrate the performance of the proposed approach using different data sets.