A consideration of the prediction of input coefficient matrix with deep learning

In this study, we present a new method for estimating input coefficient matrix with deep learning. We use the data of prefectures of Japan as the training data. However, the data is small and is likely to cause overfitting. Therefore, we attempt to avoid the over-fitting by creating new data of virtual areas aggregating multiple prefectures. The approach is based on the concept of data augmentation. As a result of predicting the input coefficient matrix of Yamaguchi prefecture and Gujo city, we showed that the method using deep learning can estimate the matrix with more stable accuracy than RAS method for these areas.