Random Ensemble MARS: Model Selection in Multivariate Adaptive Regression Splines Using Random Forest Approach
Yazarlar (2)
Arş. Gör. Dilek SABANCI
Tokat Gaziosmanpaşa Üniversitesi, Türkiye
Ondokuz Mayıs Üniversitesi, Türkiye
| Makale Türü |
Özgün Makale
|
| Makale Alt Türü | Ulusal alan endekslerinde (TR Dizin, ULAKBİM) yayınlanan tam makale |
| Dergi Adı | Journal of New Theory |
| Dergi ISSN | 2149-1402 |
| Dergi Tarandığı Indeksler | TR DİZİN |
| Makale Dili | İngilizce |
| Basım Tarihi | 01-2022 |
| Sayı | 40 |
| Sayfalar | 27 / 45 |
| DOI Numarası | 10.53570/jnt.1147323 |
| Makale Linki | https://doi.org/10.53570/jnt.1147323 |
| Özet |
| Multivariate Adaptive Regression Splines (MARS) is a supervised learning model in machine learning, not obtained by an ensemble learning method. Ensemble learning methods are gathered from samples comprising hundreds or thousands of learners that serve the common purpose of improving the stability and accuracy of machine learning algorithms. This study presented REMARS (Random Ensemble MARS), a new MARS model selection approach obtained using the Random Forest (RF) algorithm. 200 training and test data set generated via the Bagging method were analysed in the MARS analysis engine. At the end of the analysis, two different MARS model sets were created, one yielding the smallest Mean Square Error for the test data (Test MSE) and the other yielding the smallest Generalised Cross-Validation (GCV) value. The best model was estimated for both Test MSE and GCV criteria by examining the error of measurement criteria, variable importance averages, and frequencies of the knot values for each model. Eventually, a new model was obtained via the ensemble learning method, i.e., REMARS, that yields result as good as the MARS model obtained from the original data set. The MARS model, which works better in the larger data set, provides more reliable results with smaller data sets utilising the proposed method. |
| Anahtar Kelimeler |
BM Sürdürülebilir Kalkınma Amaçları
| Atıf Sayıları | |
| Google Scholar | 14 |