Pattern-Adaptive Time Series Prediction via Online Learning and Paralleled Processing using CUDA
Zhifeng He, Meng Han, Bing Han · 2019
In this paper we study the problem of prediction of time series data in a computational efficient way. We proposed a SARIMA (Seasonal AutoRegressive Integrated Moving Average) based-online learning algorithm to update the model parameters via Gradient Descent based algorithm every time when a new data comes, based on the gap between prediction and true value of the time series data. In this way, our model can adapt to the new pattern of the time series if the pattern changes, and generates more accurate prediction results comparing with existing SARIMA implementations. Besides, it has the capability of learning the true pattern and converge to optimal model even if the start point is far from optimal. We also implement our algorithm on CUDA (Compute Unified Device Architecture) platform to support training millions of different online learning models concurrently by leveraging the GPU's capability of parallel computing. The performance of our online learning algorithm is evaluated with simulation results.