Joint embedding space learning with a supervised contrastive loss for item categorization

Abstract

In the real-world, product classification is a significant task. It is because item categorization is combined with the financial element in the flow leading to manufacturers, distributors, sellers, and consumers. In addition, the process of managing products and categories by each company is a very labor-intensive task due to its large scale. To this end, previous studies on product classification generally focused on feature extraction of products and categories and neglected the task of adjusting features in the learning process. Previous studies used product and category information, but there are some improvements. Because, in reality, there are situations in which unlearned categories need to be classified. To solve this problem, we want to take advantage of the benefits of learning products and categories in the joint embedding space. In addition, we propose a method of finely adjusting feature vectors by learning the relationship between products and categories based on supervised contrastive learning rather than simply projecting feature vectors into the same embedding space. We set up two scenarios depending on whether categories appeared in training data, and as a result, our proposed method produces better performance than general classification models.