Discretization of numerical meta-features into categorical: analysis of educational and business data sets

• Published in: 2022 45th Jubilee International Convention on Information, Communication and Electronic Technology (MIPRO) pp. 1179 - 1184
• Main Authors: Oreski, Dijana, Visnjic, Dunja, Kadoic, Nikola
• Format: Conference Proceeding
• Language: English
• Published: Croatian Society MIPRO 23-05-2022
• Subjects: business data; Data mining; Education; educational data; meta-features; meta-learning; Prediction algorithms; Social sciences; Task analysis; Taxonomy; Transforms
• ISSN: 2623-8764
• DOI: 10.23919/MIPRO55190.2022.9803574

Meta-learning is learning from previous experience gained while applying learning algorithms to different data. Meta-learning consists of three steps: (i) establishing meta-features, (ii) performing learning, and (iii) prediction. This paper focuses on the first step, meta-features. Meta-features are a mix of numerical and categorical variables. We build upon the idea that learning from numerical meta-features is often less effective and less efficient than learning from categorical meta-features. Thus, the objective of this study is to discretize numerical meta-features into categorical values. An overview of meta-features is given in the paper, along with a taxonomy of discretization methods. In addition, a survey of significant discretization methods is provided. Then, discretization is performed on 58 datasets selected from two domains of social sciences: educational and business domains. Research results are discussed, and contributions for meta-learning process improvement are provided.