LAUSR

Internet of Things applications can greatly benefit from accurate prediction models. The performance of prediction models is highly dependent on the quantity and quality of their training data. In this paper, we investigate the creation of a dynamic ensemble from distributed deep learning models by considering the spatiotemporal patterns embedded in the training data. Our dynamic ensemble does not depend on offline configurations. Instead, it exploits the spatiotemporal patterns embedded in the training data to generate dynamic weights for the underlying weak distributed deep learners to create a stronger learner. Our evaluation experiments using three real-world datasets in the context of the smart city show that our proposed dynamic ensemble strategy leads to an improved error rate of up to 33% compared to the baseline strategy even when using $\frac {1}{3}$ of the training data. Moreover, using only 20% of the training data, the error rate of the model slightly increased by up to 2 in terms of mean square error. This increase is 82% less than the 11.3 increase seen in the baseline model. Therefore, our approach contributes to the reduced network traffic while not hindering the accuracy significantly.