Ensemble Based Predictive Model for Streaming Data

The Internet of Things (IoT) is employed in diverse applications to provide intelligent solutions. While IoT improves the user experience, the primary difficulty in realizing such applications is the large volume of data generated from the devices. Suitable analytics techniques are demanded to derive valuable insights from this data. the analysis becomes further challenging when real-time processing is expected. This paper evaluates the performance of regression models in an IoT dataset utilizing ensembling and adaptive windowing concepts. One of the critical factors that impact prediction accuracy in real-time streams is the training window size. We examine the impact of stream window size on prediction error by applying fixed-sized and adaptive windows. Also, a rolling variance-based model training is utilized to detect the change in the data stream. The results show that the ensembling technique with an adaptive windowing scheme and rolling-variance-based change detection provides better predictions than independent models.