Departing from traditional clustering algorithms based on distance in a high dimensional feature space, a new clustering method that combines single feature clusterings is developed. The new scheme separates samples into an unknown number of classes considering only one feature at a time. Then the quality of each single feature clustering is measured and used to estimate the feature's discriminant power. These featurewise clusterings are combined into a global clustering, each contributing by the relative amount of its quality.
Due to the subdivision of classification task into simpler single feature clusterings, the algorithm is simpler without normalizing different feature measurements. For the same reason, the clustering procedure and the criterion function are more domain-independent. Moreover the clustering quality of each feature can be used as an estimate of its discriminant power for further tasks such as the classifier design.
