Scalable and Hybrid Ensemble-Based Causality Discovery

원문 초록

Causality discovery mines cause-effect relationships among different variables of a system and has been widely used in many disciplines including climatology and neuroscience. To discover causal relationships, many data-driven causality discovery methods, e.g., Granger causality, PCMCI and Dynamic Bayesian Network, have been proposed. Many of these causality discovery approaches mine time series data and generate a directed causality graph where each graph edge denotes a causeeffect relationship between the two connected graph nodes. Our benchmarking of different causality discovery approaches with real-world climate data shows these approaches often generate quite different causality results with the same input dataset due to their internal learning mechanism differences. Meanwhile, there are ever-increasing available data in virtually every discipline, which makes it more and more difficult to use existing causality discovery algorithms to produce causality results within reasonable time. To address these two challenges, this paper utilizes data partitioning and ensemble techniques, and proposes a two-phase hybrid causality ensemble framework. The framework first conducts phase 1 data ensemble for partitioned data and then conducts phase 2 algorithm ensemble from data ensemble results. To achieve scalability, we further parallelize the ensemble approaches via the Spark big data analytics engine. Our experiments show that our proposed approaches achieve good accuracy through ensemble and high scalability through data-parallelization in distributed computing environments.