Toward Uniformity of Clusters Across Methods by Reducing Errors in Data

Abstract

Iterative Intercolumnar Correlation Classification (IICC) computes the correlation coefficients for the entries of every column of a matrix with those of every other column of the matrix. The procedure is iterated until a matrix is generally realized of only two clusters. Every object of each cluster has an index of plus-one with every other object of its cluster. Indices of minus one only mediate between objects across clusters.

IICC probably goes to an extreme. It disallows that any object of a cluster can be less than a perfect respresentative of that cluster. This extreme derives from the definition of a cluster by IICC. The stringency of the definition is reduced. This justifies discontinuing iteration earlier.

Both initial indices and those of the middle iteration were clustered by four methods.

Iteration increases the size and validity of indices, reduces errors in the indices, and increases homogeneity amongst them. Increased similarity in procedures ensues as at least some methods are adapted to middle-iteration indices. Both the validity of clusters and their similarity across methods are increased.

Other sets of data, other methods, and other definitions need to be studied in the above fashion.