Bee: towards a robust attribute reduction

Yining Chen

doi:10.1007/s13042-022-01633-4

Return

Bee: towards a robust attribute reduction

2022-09-07

PRE

DOI：10.1007/s13042-022-01633-4

Original

Original request for help

Save

Abstract

En 中文

The problem solving of attribute reduction is popular in reducing dimensions of data. Note that besides the efficiency of searching expected reducts, the performance related to the derived reducts should also be paid much attention to. Among various representative performance, it is worth mentioning that the robustness of reduct is crucial to downstream learning tasks. The reason is contributed to the fact that unstable results of attribute reduction may shake the confidence of domain experts when experimentally validating the selected attributes in reducts. In view of this, a novel framework called Bucket based Ensemble sElector (Bee) was developed, which outputs robust reduct with higher stability. Firstly, raw sample space was partitioned by a bucket mechanism. Secondly, over each bucket, candidate attributes were evaluated and then an appropriate attribute was identified. Finally, a voting was executed to identify a universal attribute which should be added into reduct pool for each iteration in the process of searching. Additionally, our framework was introduced into not only the searchings of approximation quality, regularization loss, unsupervised relevance related reducts, but also a quick searching procedure called attribute group. By testing 20 UCI benckmark data sets with raw label and 4 different ratios (10%, 20%, 30%, 40%) of noisy label, comprehensive experiments demonstrated the superiorities of our Bee: it not only offers robust results of attribute reduction but also guarantees comparable predictions than some other popular algorithms.

Keywords:

Attribute reduction

Bee

Bucket

Ensemble selector

Rough set