Predicting house prices using Ensemble Learning with Cluster Aggregations

Abstract: The purpose of this investigation, as prescribed by Valueguard AB, was to evaluate the utility of Machine Learning (ML) models to estimate prices on samples of their housing dataset. Specifically,the aim was to minimize the Median Absolute Percent Error (MDAPE) of the predictions. Valueguard were particularly interested in models where the dataset is clustered by coordinates and/or attributes in various ways to see if this can improve results. Ensemble Learning models with cluster aggregations were built and compared against similar model counterparts which do not partition the data. The weak learners were either lazy kNN learners (k nearest neighbors), or eager ANN learners (artificial neural networks) and the test set objects were either classified to single weak learners or tuned to multiple weak learners. The best results were achieved by the cluster aggregation model where test objects were tuned to multiple weak learners and it also showed the most potential for improvement.