Treatment of Missing Data in Intelligent Lighting Applications

The concept of intelligent lighting facilitates the use of machine learning models to adapt the lighting application behavior based on changing context. Ideally, a complete dataset without missing values is used to train the learning algorithm. Nevertheless, it is common to have missing data values in the dataset, e.g. due to lack of rich enough user interfaces such as smart phones. In this paper, we study various probabilistic approaches to treat missing feature values in a dataset collected from an office breakout area. This dataset is used to train the learning model to provide intelligent lighting solutions. We evaluate the performance of five different approaches by simulation, using four rule-based classification algorithms and various proportions of missing data. We find that none of these approaches gives best performance over the necessary range of conditions, and that an adaptive strategy is more suited.