Fusing attribute reduction accelerators

In the fields of rough set and machine learning, attribute reduction has been demonstrated to be effective in removing redundant attributes with clear explanations. Therefore, not only the generalization performances of the derived reducts, but also the efficiencies of searching reducts have drawn much attention. Immediately, various accelerators for quickly deriving reducts have been designed. However, most of the existing solutions merely speed up the procedure of searching reduct from one and only one perspective, it follows that the efficiencies of those accelerators may be further improved with a fusion view. For such a reason, a framework called Fusing Attribute Reduction Accelerators (FARA) is developed. Our framework is specifically characterized by the following three aspects: (1) sample based accelerator, which is realized by gradually reducing the volume of samples based on the mechanism of positive approximation; (2) attribute based accelerator, which is realized by adding multiple qualified attributes into the potential reduct for each iteration; (3) granularity based accelerator, which is realized by ignoring the candidate attributes within coarser granularity. By examining both the efficiencies of the searchings and the effectiveness of the searched reducts, comprehensive experiments over 20 public datasets fairly validated the superiorities of our framework against 5 popular accelerators. (C) 2021 Elsevier Inc. All rights reserved.