Application of Synthetic Minority Over-Sampling Technique (SMOTE) to Outlier Data for Probabilistic Neural Network (PNN)

Ramdan Hayati (Program Studi Statistika, Universitas Negeri Gorontalo, Bone Bolango 96119, Indonesia)
Isran Hasan (Program Studi Statistika, Fakultas Matematika dan Ilmu Pengetahuan Alam, Universitas Negeri Gorontalo)
Novianita Achmad (Program Studi Matematika, Universitas Negeri Gorontalo, Bone Bolango 96119, Indonesia)

Jurnal Matematika Sains dan Teknologi
Vol. 25 No. 1 (2024)

One common model of Artificial Neural Network (ANN) used in classification tasks is the Probabilistic Neural Network (PNN). PNN is an algorithm that utilizes probability functions, eliminating the necessity for a large dataset during its development process. In this research, the best model parameters were initially determined using the sigma parameter and Kernel Density Estimation (KDE) function on a randomly sampled dataset employing the Stratified Random Sampling (SRS) method. The optimal sigma parameter obtained from this process is 0.075, with a Gaussian KDE function. The data used in this study is related to direct marketing campaigns (phone calls) from Portuguese banking institutions collected by S ́ergio. Subsequently, PNN is applied to this dataset to determine its Accuracy and F1-Score values. The results indicate an accuracy rate of 87.117% and an F1-Score of 92.755%. Following this, Synthetic Minority Over-Sampling Technique (SMOTE) is applied to the dataset to balance the data. PNN is then implemented on the oversampled data, and in this phase, an evaluation of the Accuracy and F1-Score values is conducted, resulting in respective figures of 93.437% and 93.511%.