Ensemble-Based Semi-Supervised Learning with Optimal Feature Weighting

• Activity Number: 521
• Type: Contributed
• Date/Time: Wednesday, August 5, 2009 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Joseph Retzer*+
• Companies: Maritz Research
• Keywords: COSA ; Random Forests ; Cluster ensembles ; semi-supervised ; pairwise consensus

Abstract

Market segmentation via unsupervised learning (aka cluster analysis) is a commonly used tool in market research. The goal of a segmentation study should ultimately be to inform marketing decisions. In order for this to happen, clusters must be produced that are both of high quality (i.e., high intra-cluster similarity and low inter-cluster similarity) and effective for driving marketing tactics. This analysis will address both the inclusion of noisy variables as well as the problem of profiling on tactical measures. Noisy variables will be identified and appropriately incorporated using COSA (Clustering Objects on Subsets of Attributes), see Friedman and Meulman 2004. Cluster profiling on tactical measures will be enhanced via incorporation of supervised learning results from Random Forest Analysis, see Breiman 2001. Cluster ensemble analysis will be used to combine results (i.e. SSL).