XAI Personalized Recommendation Algorithm Using ViT and K-Means

Abstract

This study presents an unsupervised learning instance segmentation methodology using self-supervised learning Vision Transformer (ViT) and K-means methodology. The proposed instance segmentation is attracting attention as a core field of computer vision that assigns all pixels of an image to an appropriate class and localizes objects to bounding boxes. However, the task of producing high-accuracy pixel-level labels is more important than image classification and object detection. It requires high costs and a lot of time. Therefore, this study provides a clear and easy-to-understand explanation for the personalized decision-making process by using an iterative object mask refinement technique that performs class agnostic unsupervised instance segmentation using K-means clustering and self-adaptive supervised learned ViT. do. The proposed method generates pseudo labels that can be used to learn commercial instance segmentation models. The generated pseudo labels have higher accuracy than other methodologies, and the instance segmentation model learned with pseudo labels improves the existing highest performance by more than 50–80%. Therefore, in this study, without changing the structure of the learning function or model, we proposed a personalized ViT recommendation algorithm through single object discovery, multi-object discovery, and supervised learning segmentation using K-means, a simple clustering methodology, and self-supervised learning ViT.