Type: dissertation
Abstract: This thesis investigates semi-supervised learning (SSL) methodologies under inductive, fairness, and privacy constraints.In a supervised learning setting, every instance is labeled (i.e. it has the target variable).In a semisupervised learning setting, there are both labeled and unlabeled instances.In order to make a prediction, SSL algorithms leverage the observed mapping between features and labels of labeled instances, and the distribution of features of unlabeled ones.There are two main families of SSL algorithms: transductive and inductive.In the first part of this thesis, we propose a SSL framework to make inductive any transductive method.The typical SSL setting, where there is a scarcity of labels, could lead to the amplification of biases already present in the data: any machine learning algorithm should avoid biases present in historical data and this task is even more challenging when there is only a small number of labeled instances.For this reason, in the second part of this thesis, we investigate a debiasing training method to provide fair representations in a SSL setting.Lastly, in the final part of this manuscript, we study a particular setting of SSL where the labeled instances have privacy constraints.We build up on the definition of differential privacy and conduct different experiments to prove the effectiveness of our approach.
Author: YANG, SHUYI
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7217904405