Machine learning methods for outliers detection in unequally distributed data sets

• Grant number: 22/11854-0
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Effective date (Start): November 01, 2022
• Effective date (End): October 31, 2023
• Field of knowledge: Physical Sciences and Mathematics - Probability and Statistics - Statistics
• Principal Investigator: Mario de Castro Andrade Filho
• Grantee: Matheus Vinícius Barreto de Farias
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil

Abstract

Prediction and classification problems are increasingly relevant in the nowadays data-driven world. In this context, we highlight the relevance of outliers prediction. An outlier is an observation that distances itself from the bulk, which can be characterized, for example, as attempts at bank fraud or intrusions into a communications system that appear to be legitimate. In this sense, the proposed project deals with the study of state-of-the-art algorithms for outliers detection, with emphasis on the BCOPS algorithm developed by Guan and Tibshirani (2022). BCOPS is a conformal prediction algorithm combined with supervised learning that seeks to build prediction sets with a certain level of coverage, without any loss in classification performance. We aim to unravel the details of this method and compare it with the alternatives already existing in the literature, as well as explore the cases of best and worst performance of the algorithm.