In recent years, while the use of personal data has become increasingly active, the importance of privacy protection has also been emphasized. Differential Privacy (DP) is a technology that protects privacy by adding noise. While DP manages privacy centrally on a central server, Local Differential Privacy (LDP) manages privacy locally. From the viewpoint of user privacy protection, the use of LDP is desirable, but it has the problem of degrading the usefulness of data. A machine learning framework, SUPM, has been proposed using LDP and a machine learning privacy mechanism called WALDP. SUPM uniformly handles various attributes in data for simple use. However, the experiments in existing studies were conducted only on datasets containing only numerical data. In this paper, we conduct experiments on discrete data and propose a privacy mechanism that reduces data degradation by utilizing the properties of attributes.