INFLUENCE OF OUTLIERS ON CLUSTER CORRESPONDENCE ANALYSIS

This paper focuses on determining the influence of outliers on a joint dimension reduction and clustering method for categorical data, namely Cluster Cor- respondence Analysis (CCA). Joint methods, such as CCA, solutions consist of both a cluster membership vector and a set of low dimensional scores for observations and attributes. We evaluate the impact of outliers on the identification of the cluster struc- ture. As a benchmark, we use the tandem approach, which is a sequential application of multiple correspondence analysis followed by K-means clustering. The appraisal is based on synthetic data and outliers generated using an evolutionary algorithm that provides data with a user-defined cluster structure.