Causality : exploratory data analysis and knowledge discovery

D.Phil. (Electrical and Electronic Engineering)Abstract: The phenomenon of cause and effect which rules the natural behaviour of the universe is simple in observation but complicated in interdependency. While all action and reaction states observed in time space are easier to work on, still the difficulty lies in the factor relations. Only knowing the facts/features without the time frame as they occurred/observed heightens the complexity of information retrieval. The relation of cause and effect is vital for knowing the past information which constructs the present state, although feature links remain debatable in this case. The study of Causality deals with these exploratory data analysis problems to inform all possible vital facts which can be extracted from the feature sets. Many researchers also consider the Causal Analysis as the golden standard in data mining and analysis. As is frequently the case, this causal analysis is represented by directed acyclic graphs for simplification of complexity. The directed edges with weighted values inform the flow of information from source/parent to the receiver/child nodes in the graph. The definition of the causal structure for inference analysis is insufficient for many reasons, and the works concluded in this field are inadequate. Most of the techniques proposed provide limited structural analysis, while many others are not able to validate the required criteria for causal analysis. The background study of all the proposed articles with definite contributions towards causality have been studied and are thoroughly analyzed in the literature review. All the methods proposed yet, use the bivariate model for causal analysis. In this scenario, the model, Linear Non-Gaussian Acyclic Model (LiNGAM) is the first to provide estimation for the most number of features. However, it is not completely effective in analyzing the causal models for datasets of mixed distribution types and also constructing a complete causal model from the estimated results is not possible. While using the fundamental structure of LiNGAM, the estimation process for causal detection is newly introduced by the method Altered-LiNGAM (ALiNGAM) in this work. ALiNGAM uses least square estimation on dseparable sets to find the probable causal directions in the observed feature set. The proposed..