Isolation Forest Based Multi-Source Unsupervised Transfer Learning for Missing GDP Prediction

The rapid growth in industrialization has proportional effect on the increase in carbon emission as well as economic growth of a nation. Nevertheless, there are many nations with unavailable information on their gross domestic products (GDPs). Therefore, primarily, this paper addresses the problem of predicting missing GDP of these nations with the help of their carbon emission data. However, the available data of these countries are insufficient for training a predictive machine learning model. So, we have focused on the emerging yet under-explored area of multi-source unsupervised transfer learning to enlarge the training domain by introducing the detection and removal of anomalies in order to build a robust prediction framework. This is empirically evaluated over the carbon emission and per capita GDP data, collected from the World Bank repository, of a number of developing countries as well as over a set of mixed (developed and developing) countries. Five different domains generated using multi-source unsupervised transfer learning framework are evaluated using three different machine learning models. The best among them is then used to predict the missing per capita GDP of a nation.