FI-GEM networks for incomplete time-series prediction

Abstract

This paper considers the problem of incomplete time-series prediction by FI-GEM (fill-in-generalized ensemble method) networks, which has two steps. The first step is composed of several fill-in methods for preprocessing the missing value of time-series and the outcome are the complete time-series data. The second step is composed of the several individual multilayer perceptrons (MLP) whose their outputs are combined by the generalized ensemble method. There are five fill-in methods that are explored: cubic smoothing spline interpolation, and four imputation methods: EM (expectation maximization), regularized EM, average EM, average regularized EM. Mackey-Glass chaotic time-series and sunspots data are used for evaluating our approach. The experimental results show that the prediction accuracy of FI-GEM networks are much better than individual neural networks.