Unsupervised Neural Network-Naive Bayes Model for Grouping Data Regional Development Results

TL;DR

This paper introduces a hybrid unsupervised-supervised neural network model combining SOM-NN and Naive Bayes to improve regional development quadrant clustering accuracy, achieving 98.1% accuracy on test data.

Contribution

It presents a novel hybrid clustering approach using SOM-NN for feature generation and Naive Bayes for classification in regional development analysis.

Findings

Clustering accuracy of 98.1% on test data

Compared to manual Klassen analysis, the model's accuracy is 29.63%

Model performance depends on dataset size and density

Abstract

Determination quadrant development has an important role in order to determine the achievement of the development of a district, in terms of the sector's gross regional domestic product (GDP). The process of determining the quadrant development typically uses Klassen rules based on its sector GDP. This study aims to provide a new approach in the conduct of regional development quadrant clustering using cluster techniques. Clustering is performed based on the average value of the growth and development of a district contribution compared with the average value and contribution of the development of the province based on data in comparison with a year of data to be compared. Testing models of clustering, performed on a dataset of two provinces, namely Banten (as a data testing) and Central Java (as the training data), to see the accuracy of the classification model proposed. The proposed model consists of two learning methods in it, namely unsupervised (Self Organizing Map / SOM-NN) method and supervised (Naive Bayess). SOM-NN method is used as a learning engine to generate training data for the target Class that will be used in the machine learning Naive Bayess. The results showed the clustering accuracy rate of the model was 98.1%, while the clustering accuracy rate of the model results compared to manual analysis shows the accuracy of the typology Klassen smaller, ie 29.63%. On one side, clustering results of the proposed model is influenced by the number and keagaraman data sets used.