Multigranulation consensus fuzzy-rough based attribute reduction

摘要

As big data often contains a significant amount of unstructured, imprecise, and uncertain data, the fuzzy-rough-set-based attribute reduction is a valuable technique for uncertainty reasoning and data mining. However, tradition fuzzy-rough-set algorithms take all data correlations into account, which may lead to challenging requirements for computing and memory space resources. In this paper, we propose a novel Multigranulation Consensus Fuzzy-Rough Attribute Reduction (MCFR) algorithm for big data analysis, which assigns tasks to multiple nodes for parallel computing with different multigranulation weights and creates reasonable granular subpopulations for attribute reduction. A multigranulation margin model with a self-evolving consensus scheme is constructed wherein super elitists are classified into different multigranulation populations using the weighted margins’ matrix to calculate the parallel positive regions. This model allows penalizing super elitists with cooperative behaviors from coarsening to refining. A cascade cross-coevolutionary multigranulation learning model utilizes a more efficient coevolutionary attribute classification with multigranulation flexible thresholds. Here, the super elitist crossover learning strategy is employed to search for elitists’ chromosomes in granular subpopulations to yield better objective function values for handling dynamic big data classification tasks. Our experimental results demonstrate that MCFR can achieve a high performance in addressing uncertainty and fuzzy attribute reduction problems for big datasets with increasing noise. We also demonstrate a practical application of MCFR by using it to automatically segment fMRIs of the neonatal brain tissues at both the global and tissue levels. The results were in the good agreement with expert manual classifications, thus introducing a new method for the medical decision support problem of predicting disorders from neonatal brain fMRIs.