Supervised sampling for clustering large data sets

The problem of clustering large data sets has attracted a lot of current research. The approaches taken are mainly based either on the more efficient implementation or modification of existing methods or/and on the construction of clusters from a small sub-sample of the data and then the assignment of all observations in those clusters. The current paper focuses on the latter direction. An alternative supervised procedure to create the clusters is proposed. For learning the clusters, the procedure is using subsets of the data which are still constructed via sub-sampling but within partitions of the observation space. The general applicability of the approach is discussed together with tuning the parameters that it depends on to increase its ability. The procedure is applied to clustering the navigation patterns in the msnbc.com database. K eywords: sub-sample; partition; hard clustering; clustering click-stream data