With the advent of Web 2.0 and the rapid technological advances, there is a plethora of data in every field; however, more data does not necessarily imply more information, rather the quality of data (veracity aspect) plays a key role. Data quality is a major issue, since machine learning algorithms are solely based on historical data to derive novel hypotheses. Data may contain noise, outliers, missing values and/or class labels, and skewed data distributions. The latter case, the so-called class-imbalance problem, is quite old and still affects dramatically machine learning algorithms. Class-imbalance causes classification models to learn effectively one particular class (majority) while ignoring other classes (minority). In extend to this issue, machine learning models that are applied in domains of high societal impact have become biased towards groups of people or individuals who are not well represented within the data. Direct and indirect discriminatory behavior is prohibited by international laws; thus, there is an urgency of mitigating discriminatory outcomes from machine learning algorithms.In this thesis, we address the aforementioned issues and propose methods that tackle class imbalance, and mitigate discriminatory outcomes in machine learning algorithms. As part of this thesis, we make the following contributions:• Tackling class-imbalance in semi-supervised learning – The class-imbalance problem is very often encountered in classification. There is a variety of methods that tackle this problem; however, there is a lack of methods that deal with class-imbalance in the semi-supervised learning. We address this problem by employing data augmentation in semi-supervised learning process in order to equalize class distributions. We show that semi-supervised learning coupled with data augmentation methods can overcome class-imbalance propagation and significantly outperform the standard semi-supervised annotation process.• Mitigating unfairness in supervised models – Fairness in supervised learning has received a lot of attention over the last years. A growing body of pre-, in- and postprocessing approaches has been proposed to mitigate algorithmic bias; however, these methods consider error rate as the performance measure of the machine learning algorithm, which causes high error rates on the under-represented class. To deal with this problem, we propose approaches that operate in pre-, in- and post-processing layers while accounting for all classes. Our proposed methods outperform state-of-the-art methods in terms of performance while being able to mitigate unfair outcomes.