Parallel attribute reduction algorithm for unlabeled data based on fuzzy discernibility matrix and soft deletion behavior

In attribute reduction algorithms, discernibility matrix-based methods and heuristic-based methods are two highly effective approaches. While the prevailing view is that heuristic algorithms are faster than discernibility matrix-based methods, the rise of GPUs and other matrix-based parallel computing devices has enabled discernibility matrix reduction methods to achieve faster computation speeds by leveraging their matrix characteristics. However, few discernibility matrix-based methods can directly adapt to GPU devices, and for unlabeled data, existing discernibility matrix-based methods fail to fully utilize the fuzzy information, resulting in unsatisfactory outcomes. In this paper, we propose a parallel attribute reduction algorithm based on fuzzy discernibility matrices and soft deletion behavior. To achieve parallel computing, we transform the traditional 3-dimensional discernibility matrix into a 2-dimensional matrix. To maximize the use of fuzzy discernibility information, we introduce a fuzzy deletion function, which can effectively update the discernibility matrix by incorporating fuzzy discernibility information. Finally, we propose a stopping mechanism for the algorithm, enabling it to select fewer attributes under appropriate conditions. Experiments demonstrate that our algorithm significantly increases computation speed compared to traditional heuristic algorithms and reduces the number of attributes while maintaining and enhancing the effectiveness of downstream tasks.