Learning Correlations and Causalities Through an Inductive Bootstrapping Process
Xiwen Yang, Seng-Beng Ho · 2018
Both temporal correlation and causality between events are important for an intelligent system for prediction and problem solving purposes. However, learning temporal correlation or causality from real world data consisting of intervening noise is a challenging inductive learning problem. We propose an inductive learning method that capitalizes on a bootstrapping process to recover correlational or causal relations between events that could otherwise be obscured by noise. The method relies primarily on “explaining away” to skip over noisy events or other unrelated correlations and causalities to identify and enhance the strengths of putative correlations and causalities between events of concern. We use artificially generated event data as well as data obtained from real world videos containing events in putative correlational and causal relations to test the formula, and it is found that the formula is able to successfully uncover the putative correlations and causalities involved.