Analysis of tabular data based on graph neural network using supervised contrastive loss

Abstract

In many applications in the industry, tabular data are the most commonly used data type. Tabular data have the advantage of being easy to understand and interpret. Machine learning algorithms are mainly used to deduce a mapping function that maps input data to output in tabular data. Machine learning methods for dealing with tabular data are classified into two categories: similarity-based approach and feature-based approach. The feature-based approach explores a mapping function by clarifying the relationships between features, while the similarity-based approach uses similarities of observations to explore a mapping function. Both aspects have their pros and cons. Feature-based models are easy to understand and intuitive to use and deploy but generally cannot utilize the relationships between observations. Similarity-based models are most suited for exploiting the relationships among observations, but their availability is usually limited. In order to take advantage of both aspects, we propose an algorithm to combine feature-based and similarity-based approaches using a graph neural network. Additionally, to complement the shortcomings of cross-entropy loss, the proposed method applies supervised contrastive learning. Through supervised contrastive learning, the proposed method boosts the efficiency of learning the generalized representation for each class. Experimental results show that the proposed method provides more precise results for classification tasks, implying that it may improve the generalization capability.