Noise Labels in Machine Learning: Evaluation measures and machine learning algorithms

Abstract

Supervised Machine Learning aims to automatically build classification models from a set o labeled data. However, highly quality labeled data is not available in numerous real world problems, an issue known as label noise. This label noise may cause numerous problems for Machine Learning algorithms. Although some research has been conducted in the model generation phase, few studies focus on the evaluation process in presence of noise labels. This is an important research topic, as Machine Learning research is based on empirical studies, and misleading conclusions may be drawn when comparing learning algorithms in class noise context. This research proposal aims to investigate this open-ended topic, developing a systematic study regarding the influence of some class-noise patterns in the evaluation phase of the machine learning process.