Implementación de algoritmos de inferencia causal utilizando computación paralela: PC y PC-stable

Abstract

There are complex systems whose dynamics are difficult to understand and of which only information of certain variables is available. The amount of information collected from those variables makes the problem an ideal scenario for using knowledge discovery techniques such as PC and PC-stable algorithms. In this work we create an efficient implementation of PC and PC-stable in the C programming language, adding background knowledge and applying parallelization techniques for systems with shared (openMP) and distributed memory (Open MPI). Performance tests of the implementation were performed with synthetic and real data sets and compared with pcalg (R). Also, we reproduced the results of applying PC-stable to simulations of advection and diffusion processes. Results of the performance tests show that our implementation with synthetic data is up to 30 times faster than pcalg (serial) and the parallel implementation (openMP) is up to 3.7 times faster than the serial implementation in C. Furthermore, our implementation (openMP) with real datasets is 11 times faster than parallel pcalg. Although the results of our implementation compared to pcalg differ in 6 of the 1200 synthetic datasets tested, in up to 22 links (of 910) of the resulting DAG. We considered that these differences have no impact on the conclusions obtained from the resultant DAGs. When applying the implementation to simulations of processes, we found that our implementation is more drastic deleting links at different times. Although, the resulting links in the DAG have the same behavior as in the replicated results.