Predicting Happiness - Comparison of Supervised Machine Learning Techniques Performance on a Multiclass Classification Problem

Abstract

In the modern world, especially in contemporary economies and politics, a population's subjective well-being is a frequent subject of the public debate. As comparisons of happiness levels in different countries are published, different circumstances and their effect on the value of the subjective well-being reported by people are also analysed. However, a significant amount of the research related to subjective well-being and its determinants is still based upon survey answers and employing conventional statistical methods providing details regarding correlations and causality between different factors and subjective well-being. Application of Supervised Machine Learning techniques for prediction of subjective well-being may provide new ways of understanding how individual factors contribute to the concept value and allow for addressing any issues, which may potentially affect mental and physical health. The focus of this research is to use the survey data and make predictions regarding subjective well-being (a multiclass target) using Supervised Machine Learning models. In particular, the study is aimed at comparing the performance of two techniques: Decision Tree and Neural Networks. The „C4.5 algorithm‟ used by the Decision Trees is considered as the benchmark algorithm, to which other supervised learning algorithms should be compared. At the same time, Neural Networks were previously proven to have high predictive power, even with multiclass categorisation problems. Two experiments are conducted as part of this research, one using original highly imbalanced data; the other using the dataset balanced using SMOTE. The experimental results gathered show that for the first experiment there is no statistically significant difference (p