Predicting smart home lighting behavior from sensors and user input using very fast decision tree with Kernel Density Estimation and improved Laplace correction

Abstract

One way to predict the behavior of smart home lighting is by using machine learning. Currently many methods of supervised learning that used for this problem, one of them is decision tree method. Very Fast Decision Tree (VFDT) as one of the decision tree method that has advantages in online machine learning that may useful in smart home, but there are still some room of improvisation that can improve accuracy of VFDT. The experiment result is obtained that VFDT is better than Naïve Bayes (NB) and Artificial Neural Network (ANN) in offline and online experiment. In addition, Kernel Density Estimation (KDE) and improved Laplace correction that is used as improvisation of VFDT is able to increase the accuracy and Matthews Correlation Coefficient (MCC) of VFDT in predicting smart home lighting switch usage.