Using ART1 Neural Networks to Determine Clustering Tendency

Clustering is an unsupervised, data driven learning paradigm that aims at discovering natural groups in data [8, 9]. This type of learning has found many useful applications in domains with large amount of data where labeling of a training set for supervised learning is cost prohibitive or where autonomy is essential [1, 10, 11, 12]. However, clustering algorithms generally rely on some prior knowledge of the structure present in a data set. For instance, one needs to know whether or not clusters actually exist in data prior to applying a clustering procedure. Indeed, clustering applied to a data set with no naturally occurring clusters would merely impose meaningless structure. The procedure that consists in examining a data set to determine if structure is actually present and thus determine if clustering is a worthwhile operation is a poorly investigated problem known as cluster tendency determination [8]. Research in the area of cluster tendency has mainly focussed on the somewhat related problem of establishing the true number of clusters present in the data [6], often as part of cluster validity, the evaluation of clustering output quality [8]. Of course, should it be ascertained that the best clustering contains only one group, then null tendency must be concluded. The main problem with these approaches is that they either rely on yet other optimization procedures and similarity metrics (just as the clustering procedure itself), or depend on some parameter estimation. We show how to avoid these problems by using Adaptive Resonance Theory (ART) neural networks [3, 7] to determine clustering tendency of binary data. The binary version of ART (ART1) is used.