Predictive Modelling of Binary Classification Task on Imbalanced Data Using Supervised Learning Algorithm (Case Study: The House of Representatives of the Republic of Indonesia Member Electability in Election)

• Main Author: Yudha Pranata, Anugrah
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/43458
• Institution: Institut Teknologi Bandung

General election to select members of the House of Representatives (DPR) in Indonesia is done once in every 5 years. In every election, only a few of candidates nominated by political parties is elected and serves as legislative member in the House of Representatives. Ananda, Arifin & Suryadinata (2005) documented that there were 7,756 candidates for 550 seats in the House of Representatives in 2004. Then, General Election Comission documented that there were 11,301 candidates for 560 seats in the House of Representatives in 2009, 6,606 candidates for 560 seats in the House of Representatives in 2014 and 7,968 candidates for 575 seats in the House of Representatives in 2019. Since the early beginning of the election process for every election period, there are many people/institutions trying to predict the election results, either using historical data or state affairs at the time. The condition where proportion of classes that are not equal in predictive modelling e.g. in this case of Indonesian legislative election is usually called as imbalanced data. One method that can be used to model prediction of election results is learning algorithm or usually called as machine learning. This research’s goal is to create predictive modelling for the House of Representatives members’ electability with using supervised learning algorithm. Predictive models are created using data of factors influencing election candidates’ electability from previous research about estimation model in case study of legislative election in 2004 with methods Ordinary Least Square (OLS) and Logistic Regression by Darawijaya (2014). This research is improving previous model with adding Neural Network implementation based on research reference of predictive modelling for case of president election in United States by Zolghadr, Niaki, & Niaki (2018). This predictive modelling of election candidates’ electability is created for binary classification task because there are only 2 classes in the output variable: ‘Elected’ or ‘Not Elected’. To evaluate the results of created predictive models, metrics that are used are accuracy and Receiver Operating Characteristic (ROC)/Area Under Curve (AUC). Result of this research is the predictive model created that gives the best value for all evaluation metrics is Neural Network model with data of combined candidates’ public personal information and campaign fund which results accuracy of 91.48% and AUC of 91.59%.