A0905
Title: Variable weighting for mixed-type data clustering
Authors: Alexander Foss - Sandia National Laboratories (United States)
Marianthi Markatou - University at Buffalo (United States) [presenting]
Abstract: Conceptual and technical challenges are discussed in variable weighting and variable selection in cluster analysis of mixed-type data. The specific objectives of existing weighting procedures are quite different and are often selected without a thorough discussion of competing objectives and consideration of the implications of these selection choices. The different objectives of existing weighting and variable selection techniques are described and investigated, and then the variable weighting technique MEDEA (multivariate eigenvalue decomposition error analysis) is developed. It is shown that MEDEA weighting fills an important and previously unfilled niche in cluster analysis of mixed-type data, as demonstrated by its superior performance in numerous experiments with both synthetic and actual datasets.
