STABILITY OF JOINT DIMENSION REDUCTION AND CLUSTERING

Several methods for joint dimension reduction and cluster analysis of categorical, continuous or mixed-type data have been proposed over time. These methods combine dimension reduction (PCA/MCA/PCAmix) with partitioning clus- tering (K-means) by optimizing a single objective function. Cluster stability assess- ment is a critical and inadequately discussed topic in the context of joint dimension reduction and clustering. We introduce a resampling scheme that combines boot- strapping and a measure of cluster agreement to assess global cluster stability of joint dimension reduction and clustering solutions and a Jaccard similarity approach for empirical evaluation of the stability of individual clusters. Both approaches are imple- mented in the R package clustrd.