Data and cluster weighting in target selection based on fuzzy clustering

We study the construction of probabilistic fuzzy target selection models for direct marketing by using an approach based on fuzzy clustering. Since fuzzy clustering is an unsupervised algorithm, the class labels are not used during the clustering step. However, this may lead to inefficient partitioning of the data space, as the clusters identified need not be well-separated, i.e. homogeneous, in terms of the class labels. Furthermore, the regions of the data space with more high-prospect customers should be explored in larger detail. In this paper, we propose to use a weighting approach to deal with these problems. Additional parameters, i.e. weight factors, associated with data points and the clusters are introduced into the fuzzy clustering algorithm. We derive the optimal update equations for the weighted fuzzy c-means algorithm. A heuristic method for estimating suitable values of the weight factors is also proposed. The benefits of our approach for the target selection models are illustrated by using data from the target selection campaigns of a large charity organization.