Meta-learning based industrial intelligence of feature nearest algorithm selection framework for classification problems

Li Li

doi:10.1016/j.jmsy.2021.03.007

Return

Meta-learning based industrial intelligence of feature nearest algorithm selection framework for classification problems

2022-01-01

PRE

DOI：10.1016/j.jmsy.2021.03.007

Original

Save

Abstract

En 中文

As more and more novel algorithms being proposed, it is difficult for engineers to choose suitable algorithms to solve engineering problems in manufacturing systems. According to the No Free Lunch theorems, there is no universal algorithm applicable to all problems. To solve the problem, this paper proposes a meta-learning based industrial intelligence of feature nearest algorithm selection framework for classification problems which are common in manufacturing. The proposed framework consists of three parts: dataset module, classification algorithms module and recommendation module. Firstly, 32 dataset characteristic measurement indicators are extracted based on meta-learning from 34 popular datasets. Then 13 common classification algorithms are applied in each dataset for algorithm optimizing. Through comparative analysis, it is found that the bat algorithm is better than the differential evolution algorithm in parameter optimization. Finally, in recommendation module, the relatively better algorithm is recommended to the new dataset according to the similarity between the new dataset and the reference datasets, which is measured by four distance methods. The result shows that the recommendation accuracy and hit rate of the new dataset are not less than 84.10 % and 85.29 % respectively, which can provide a good algorithm recommendation for engineers.

Keywords:

Algorithm selection

Meta-learning

Classification problems

Characteristic measurement indicators

Similarity