Feed-Backwards Construction of Correlation Chains With Embedded Regressors

Correlation analysis focuses on the identification of the direction and strength of the relationship between a pair of attributes in a common dataset, which can be used as a basis for regression analysis models. Since in some cases there is no usable direct relationship between initial and targeted attributes of such a predictive task, approaches like correlation chains create sequences of attributes which maximize the overall correlation between the two attributes of interest. In this way, the pseudo-transitivity of correlation coefficient values is utilized to produce stronger - yet indirect - relationships between attributes and therefore minimize regression error of models built in these scenarios. However, there is a significant lack of information on the inner workings of such chains and especially on the propagation of error values in them. This motivated the main objective of the work, which focuses on the design and implementation of a novel approach to constructing, visualizing, and evaluating correlation chains using an analytical sheet containing several metrics of interest - correlation values, embedded regressor error values, and regression error progression in the studied chain. The proposed model is subsequently evaluated from a qualitative and comparative point of view on two benchmarking datasets.