BiLSTM model based on multivariate time series data in multiple field for forecasting trading area

Abstract

An artificial neural network-based model is widely used for analyzing and predicting multivariate time series data. However, the study on the analysis and prediction of multivariate time series data in multiple fields has limitations in that it does not take the features of the fields into account. In this paper, we propose a Bi-directional Long Short-Term Memory model based on multivariate time-series data in multiple fields that considers the fields’ features. This model differs from the existing model in that data input into the input layer is divided into fields to learn the features of those fields. In addition, we tried to learn the trend of the time series data at the same time by simultaneously learning the value of the data and its variation. We applied the model proposed in this paper on trading area forecasts to collect purchasing data and SNS data on “restaurants” in the trading area to progress the learning. Experiments and performance evaluation were performed based on the Root Mean Square Error, which was based on whether the learning was done by each field and whether there was variation in the input value. Experimental results show that the proposed model performed better than other models.