| dc.description.abstract | The high demand for machine learning (ML) in our lives brings concern about the possible impacts that such tools may cause on society. Nowadays, algorithms are being used to assist hiring processes and identify criminals through images. From a societal standpoint, decisions made in these situations are critical. Therefore, it is essential that there is no discriminatory behavior against certain groups. To assess this issue, this work proposes to build a bridge between fairness in ML under a new perspective, from survival analysis (e.g., Cox models and variations with deep learning). To accomplish our goal, we have reviewed literature from both areas. Next, we introduced four proposal definitions of fairness for survival analysis. The first and second proposal are nominated divergence in demographic parity. Both of them focus in the difference between empirical and predicted curves. The third proposal, called casual discrimination, verifies the error of calculationg the c-index when we change data for specifics group. The last proposal is a new metric, called "justiça de filas", which compares individuals from different groups at the same time. After that, we applied these proposals on three different databases: the first one was from the heath domain, MIMIC-III, and the other two were from the criminal domain, Rossi and COMPAS. In MIMIC-III database, bias appeared in the divergence in demographic parity proposal, divergence in conditional demographic parity proposal and "justiça de filas". For example, there was a difference between the empirical and predicted curves for blacks with cancer. The same happened with black women and black men with cancer. In Rossi and COMPAS databases, situations with bias appeared using the "justiça de filas"metric, which was responsible for identified bias in all databases analyzed. In addition to the discussion in these three different contexts, at this moment, this dissertation is the first one to bring this to Cox models. | |
