Abstract:
Various models exist to predict a numerical value in supervised learning problems. One of the challenges in predicting an outcome with high degree of precision involves dealing with numerical data points which can be represented using differently. To solve for such challenge and in order to predict the logerror value in Zillow’s competition on Kaggle, we have developed a new model, BRanching Artificial Neural Ensemble (BRANE). This ensemble network uses a number of multilayer perceptrons (MLP) to predict the outcome and combines the results using an additional MLP. This approach not only allowed us to use different datatypes as inputs, but also predicted better and converged faster than traditional MLP models.
