SuperTML-Clustering: Two-Dimensional Word Embedding for Structured Tabular Data

In the tabular data classification task based on deep learning, Embedding is an important research direction, most of the current Embedding research focuses on one-dimensional Embedding, but there are few studies on two-dimensional Embedding. we propose the SuperTML-Clustering method to convert tabular data classification into image classification. SuperTML-Clustering can embed the feature values of each sample on an image, and determine the number of repeated times of the feature values on the image to explicitly show the importance of the feature.So CNN can easily capture the importance of the feature.To reduce the pressure on the CNN model to learn continuous features, we cluster important continuous feature values into buckets, and then embed the index of the bucket into the image. In order to prevent the bucket from breaking the continuity of continuous features, SuperTML-Clustering can dynamically change the pixel value indexed in the image according to the original numerical information of the feature value. Finally it puts embedded image into the CNN model for learning. We carry out experiments on three real-world datasets, and compare the proposed method with the two popular tabular classification models, and the experimental results show that the SuperTML-Clustering method has a good classification effect.