Reconstructing Causal Networks From Temporal Data - A Genetic Programming Based Approach

Abstract

This work details a new systematic approach based on Genetic Programming for finding out relationships and causality among different variables in a multivariate system and present them in a network form. The main focus is on analyzing temporal output of biological phenomenon. The developed GP based Variable Interaction Methodology (GPVIM) can be used to analyze multivariate temporal data such that the inherent interactions could be represented in the form of Multivariate Vector Autoregressive Model-guided relationship network. The methodology is further improved by use of quicker analysis methods such as Correlation, Granger Causality, and Dynamic Bayesian Network, as mode of providing `pre-cooked? data for GPVIM. This helped in resolving problems associated with large number of variables and in improving the accuracy of the final networks. The methodology has been found promising compare to other available methods, for practical network reconstruction problems, with higher accuracy and specificity.