Deep Linearization Mappings for Canonical Correlation Clustering

Kernelized canonical correlations analysis has been utilized as a means to carry out unsupervised data clustering. This work puts forth a method to obtain a non-linear mapping that transforms data to a feature space where the data has the potential to be more ‘linearly’ correlated. An auto encoder network is employed to minimize a reconstruction error and learn a proper non-linear mapping to transform the input data. The non-linear mapping found via the novel scheme goes beyond kernel functions and results in better clustering performance than linear and kernel-based canonical correlations as demonstrated by numerical tests conducted across two different data sets.