ASCLU : Alternative Subspace Clustering

Finding groups of similar objects in databases is one of the most important data mining tasks. Recently, traditional clustering approaches have been extended to generate alternative clustering solutions. The basic observation is that for each database object multiple meaningful groupings might exist: the data allows to be clustered through different perspectives. It is thus reasonable to search for deviating clusters compared to a given clustering result, that the user is not satisfied with. The existing methods focus on full space clustering. However, for today’s applications, where many attributes per object are recorded, traditional clustering is known to generate no meaningful results. Instead, the analysis of subspace projections of the data with subspace or projected clustering techniques is more suitable. In this paper, we develop the first method that detects alternative subspace clusters based on an already known subspace clustering. Considering subspace projections, we can identify alternative clusters also based on deviating dimension sets besides just deviating object sets. Thus, we realize different views on the data by using different attributes. Besides the challenge of detecting alternative subspace clusters our model avoids redundant clusters in the overall result, i.e. the generated clusters are dissimilar among each other. In experiments we analyze the effectiveness of our model and show that meaningful alternative subspace clustering solutions are generated.