Time-Series Cross-Validation Parallel Programming using MPI
Abdulrahim Mohammed, Ahmed Khedr, Duaa AlHaj, Reem Al Khalifa, Abdulla Alqaddoumi · 2021 International Conference on Data Analytics for Business and Industry (ICDABI) · 2021
Time-series data has a natural chronological arrangement with modeling and cross-validation techniques, highly dependent on sequential processing of data, which challenges its parallelization. Since the running time of the modeling algorithm largely consists of a main operation of solving the modeling algorithm several times with various training and testing dataset sizes, it is the prime target for optimizing the running time. This research presents a parallel implementation of time-series cross-validation on a rolling basis that involves using a subset of the dataset for training purposes, expanding each dataset with each run to obtain test accuracy. The evaluation is conducted using parallel speedup, parallel efficiency, and computational time. The proposed algorithm distributes the task of modeling among several processors to run independently and obtain results. The computational time recorded ranged from 15.47 seconds at best with five processors to 28.33 seconds at worst when done with a single processor. The parallel system speedup was found to be sub-linear with an efficiency decreasing with the increase in the number of processors.