Application of convolutional neural network to traditional data.

• Propose a feature grid-based CNN model, FGCN, on traditional data. • Propose methods of converting instance with form of 1-d vector to feature grid. • The performance of FGCN model has reached the state-of-the-art technique XGBoost. • The positions of features in the grid have little influence on prediction accuracy. • Fully connected layers in CNN give little marginal classification performance. Convolutional neural networks (ConvNets) have been applied to various types of data, including image, text, and speech, but not to traditional data. In this study, traditional data are defined as data whose features have no spatial or temporal dependencies but might have statistical correlations. We construct a feature grid-based ConvNet (FGCN) model for classification tasks on traditional data. The FGCN model is composed of two functional parts: The first is used to convert traditional data in the form of a 1-D feature vector into a 1-D, 2-D, or higher-dimensional feature grid; and the second is a ConvNet classifier for the converted data. The experimental results show that the FGCN model performs well; therefore, it is worth considering this model for classification tasks on traditional data. [ABSTRACT FROM AUTHOR]