Browsing by Subject "fairness"

Goals. The goal of this research was to find out, how the use of the non-identifying dynamic algorithm would affect fairness experience; and through it, behavioral intentions, in rebating context. Besides that, it was assessed how the provision of detailed information on algorithm's logic affects the fairness experience. Dynamic pricing, especially based on identification, has been shown to negatively affect fairness. The dynamic algorithms are better to companies due to their profitability potential. It is of vital importance to find out the conditions, on which they might be employed, while taking into account the possible reactions of the customers. A differential assessment of entity and event fairness through the lens of fairness heuristic theory is chosen as a backbone of this research to extend the mosaic empirical evidence of their mutual interaction paths. The fairness experience is also closely connected to affects; incidental affects and integral emotions, which are evoked by the fairness experience itself. Because of this close relationship, to complement general picture, the affects were assessed as well. Methods. The manipulations were performed on two levels. The first level, the exposure to dynamic algorithm or seeing the human-set pre-determined rebate rates, happened on the company's site when the algorithm trial was run. The second manipulation level, the amount of the available information, was performed during the gathering of the survey data. There were three conditions in the information manipulation: no information (the control), bare information about the ongoing trial and trial information including a detailed algorithm's logic description. The size of the final sample, used for the analysis, consisted of 404 participants. The main analysing technique employed was SEM. Results and conclusions. Effect paths between entity and event fairness areas were in accordance with the fairness heuristic theory - event fairness mediated the change in entity fairness partially. The subjects that were exposed to the algorithm, event fairness was affected negatively by the bare trial information as expected. The provision of the detailed information did not affect fairness. Entity fairness was connected to both, incidental affects and integral emotions. There were no analogous connection between event fairness, and affects and emotions. Fairness mediated only partially the change from incidental affects to integral emotions. Integral emotions were not connected to the behavioral intentions. Entity fairness mediated fully the effect of event fairness on the behavioral intentions. The provision of the detailed information affected directly positively on pro-active behavioral intentions without a mediation of fairness. None of the manipulations affected directly complaining intentions. The results provide important information about the dynamic algorithm exposure in real life, outside the laboratories. Despite the dynamic pricing being seen as unfair in principle, the exposure to the detailed information might have positive effects on the outcomes. There was only a limited support for the role of affects in the pricing fairness context.

As privacy gains consensus in the field of machine learning, numerous algorithms, such as differentially private stochastic gradient descent (DPSGD), have emerged to ensure privacy guarantees. Concurrently, fairness is garnering increasing attention, prompting research aimed at achieving fairness within the constraints of differential privacy. This thesis delves into algorithms designed to enhance fairness in the realm of differentially private deep learning and explores their mechanisms. It examines the role of normalization, a technique applied to these algorithms in practice, to elucidate its impact on fairness. Additionally, this thesis formalizes a hyperparameter tuning protocol to accurately assess the performance of these algorithms. Experiments across various datasets and neural network architectures were conducted to test our hypotheses under this tuning protocol. The decoupling of hyperparameters, allowing each to independently control specific properties of the algorithm, has proven to enhance performance. However, certain mechanisms, such as discarding samples with large norms and allowing unbounded hyperparameter adaptation, may significantly compromise fairness. Our experiments also confirm the critical role of hyperparameter values in influencing fairness, emphasizing the necessity of precise tuning to ensure equitable outcomes. Additionally, we observed differential convergence rates across algorithms, which affect the number of trials needed to identify optimal hyperparameter settings. This thesis aims to offer detailed perspectives on understanding fairness in differentially private deep learning and provides insights into designing algorithms that can more effectively enhance fairness.