Branching Artificial Neural Ensemble (BRANE): Supervised Learning For Missing Data

Abstract

Machine learning, a scientific model, has been developed over the years for predictions and is now very advanced with modern technology of high performance computing. It is widely used in classification and prediction problems. While using Machine Learning for regression we may come across datasets that have large amount of missing values, which have significant impact on the prediction results. The BRanching Artificial Neural Ensemble (BRANE), a machine learning algorithm, has been developed and proposed in this paper to address missing values in attributes. The proposed algorithm is an extension of Ensemble approach and utilizes one multilayer perceptron for data points, one multilayer perceptron for missing values as a flag entity, and one multilayer perceptron to combine the experts and predict the final outcome. Dataset to train and test this model has been taken from Zillow Competition from Kaggle database and is implemented in Python.