Prediction of Data Stream Based on Gaussian Process Regression with Online Free Variational Inference Approximation

Abstract

Abstract Online prediction of data stream is currently being used in various fields. The prediction method based on Gaussian Process Regression has obvious advantages since its outputs have a probabilistic significance, which is suitable for dealing with nonlinear and complex regression problem. However, the characteristics of the data stream are real-time and online. If only using the historical data of the previous time as the training sample to construct the prediction model, the model will not be accurately predicted once the distribution of the new data changes. To this end, we employ the Online Free Variational Inference approximation to build the prediction model. The key idea is to introduce a variational distribution and maximizing the Kullback-Leibler Divergence between the variational distribution and the true value of the output’s posterior distribution. Further, we use telemetry data to verify the validity of the Online Free Variational Inference approximation and the experiment shows that the online data stream can be predicted well by the method employed in the paper.