Online variational Gaussian process method for time series data

Abstract

*The invention discloses an online variational Gaussian process method for time series data, which comprises the following steps of: S1, determining a data set and an observation value by adopting a regression model of a Gaussian process framework; s2, solving variation free energy to perform single data processing, and calculating distribution of corresponding variation lower limits; s3, aiming at the stream data condition, solving the problem by adopting online variational reasoning; s4, solving the model posterior probability by solving the new relative entropy divergence; s5, converting the new relative entropy divergence into minimum variational free energy, and correspondingly solving factor distribution q * (b); s6, solving a variation lower limit to obtain variation free energy; and S7, according to the variational distribution obtained by solving, calculating prediction distribution and a prediction result at any test point. According to the method, the training complexity and the prediction complexity of a traditional Gaussian process algorithm are reduced, the calculation cost is reduced, and the method has the performance equivalent to that of a traditional sparse Gaussian process approximation method.

Description

technical field [0001] The invention relates to the field of computer technology, in particular to an online variational Gaussian process method for time series data. Background technique [0002] Gaussian Processes are a powerful and popular framework for solving machine learning problems. For time-varying data generated in real life, such as the Internet of Things (IoT), Gaussian processes are an attractive option for building real-valued nonlinear models from time-series data due to their flexible and uncertain quantification. However, the Gaussian process method is known to be computationally expensive, with a training complexity of N for N observations The prediction complexity is Therefore, probabilistic modeling becomes computationally intractable when using Gaussian process methods in a big data environment. [0003] When the data generated by the IoT scales to huge volumes or even infinite, traditional Gaussian process methods are unaffordable over time. To so...

AI Technical Summary

Problems solved by technology

Obviously, this scheme is impractical because of limited computing resources

Images