返回Google圖書搜尋

Reliability in Ensemble Learning and Learning from Crowds

出版	2019
URL	http://books.google.com.hk/books?id=YHwe0AEACAAJ&hl=&source=gbs_api

註釋The combination of several human expert labels is generally used to make reliable decisions. However, using humans or learning systems to improve the overall decision is a crucial problem. Indeed, several human experts or machine learning have not necessarily the same performance. Hence, a great effort is made to deal with this performance problem in the presence of several actors, i.e., humans or classifiers. In this thesis, we present the combination of reliable classifiers in ensemble learning and learning from crowds. The first contribution is a method, based on weighted voting, which allows selecting a reliable combination of classifications. Our algorithm RelMV transforms confidence scores, obtained during the training phase, into reliable scores. By using these scores, it determines a set of reliable candidates through both static and dynamic selection process. When it is hard to find expert labels as ground truth, we propose an approach based on Bayesian and expectation-maximization (EM) as our second contribution. The aim is to evaluate the reliability degree of each annotator and to aggregate the appropriate labels carefully. We optimize the computation time of the algorithm in order to adapt a large number of data collected from crowds. The obtained outcomes show better accuracy, stability, and computation time compared to the previous methods. Also, we conduct an experiment considering the melanoma diagnosis problem using a real-world medical dataset consisting of a set of skin lesions images, which is annotated by multiple dermatologists.