Response surface methodology to tune artificial neural network hyper-parameters

Bozkurt Keser S., Buruk Şahin Y.

Expert Systems, cilt.38, sa.8, 2021 (SCI-Expanded) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 38 Sayı: 8
  • Basım Tarihi: 2021
  • Doi Numarası: 10.1111/exsy.12792
  • Dergi Adı: Expert Systems
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, ABI/INFORM, Applied Science & Technology Source, Biotechnology Research Abstracts, Business Source Elite, Business Source Premier, Compendex, Computer & Applied Sciences, INSPEC, Library, Information Science & Technology Abstracts (LISTA), Psycinfo
  • Anahtar Kelimeler: artificial neural network, central composite design, desirability function, hyper-parameter tuning, machine learning, CUSTOMER CHURN, OPTIMIZATION, HYPERPARAMETERS, CLASSIFICATION
  • Eskişehir Osmangazi Üniversitesi Adresli: Evet

Özet

© 2021 John Wiley & Sons Ltd.Artificial neural network is a machine learning algorithm that has been widely used in many application areas. The performance of the algorithm depends on the type of the problem, the size of the problem, and the architecture of the algorithm. One of the most important things that affect ANN's performance is the selection of hyper-parameters, but there is not a specific rule to determine the hyper-parameters of the algorithm. Although there is no single well-known method in hyper-parameter tuning, this issue has been discussed in many studies. In this study, a central composite design which is a successful response surface methodology technique that considers factor interactions is used for hyper-parameter optimization. A categorical central composite design that has 39 experimental runs was used to predict accuracy and F-score. The effect of ANN hyper-parameters on selected performance indicators is investigated using two different size customer churn prediction data set, which is widely used in the literature. Using the desirability functions, multiple objectives are combined and the best hyper-parameter levels are selected. As a result of verification tests, the accuracy values are 85.79% and 83.29% for the first and second data set, respectively. And, the F-score values are 86.15% and 82.33% for the first and second data set, respectively. The results show the effectiveness of the adopted RSM technique.