ClustOfVar-based approach for unsupervised learning: Reading of synthetic variables with sociological data

This paper proposes an original data mining method for unsupervised learning, replacing traditional factor analysis with a system of variable clustering. Clustering of variables aims to group together variables that are strongly related to each other, i.e. containing the same information. We recently proposed the ClustOfVar method, specifically devoted to variable clustering, regardless of whether the variables are numeric or categorical in nature. It simultaneously provides homogeneous clusters of variables and their corresponding synthetic variables that can be read as a kind of gradient. In this algorithm, the homogeneity criterion of a cluster is defined by the squared Pearson correlation for the numeric variables and by the correlation ratio for the categorical variables. This method was tested on categorical data relating to French farmers and their perception of the environment. The use of synthetic variables provided us with an original approach of identifying the way farmers reconfigured the questions put to them.