GRU The design of GRU-based cell structure robust to missing value and noise of time-series data in recurrent neural network

摘要

The present invention provides a recurrent artificial neural network model capable of simultaneous noise alleviation and missing value replacement of time-series data in accordance with a problem to be predicted. A single cell structure includes: (a) a step of alleviating noise by a weighted average method using a noise alleviation filter capable of learning in time-series data; (b) a step of replacing missing values; and (c) a step of storing information to be remembered at the current time in a potential state vector by GRU operations. Also, in constructing a recurrent artificial neural network model, in a process where a weight parameter for noise alleviation included in the cell structure learns the recurrent artificial neural network model to be suitable for a prediction project in the step (a), learning is performed to be optimized for the project. By the method, the recurrent artificial neural network model simultaneously performing noise alleviation and missing value replacement of time-series data without separate preprocessing can be used in various machine learning projects.