Detecting and mitigating bias in training datasets

Veliki porast mogućnosti i sve veća prisutnost sustava utemeljenih na AI-u u životima ljudi postavljaju pitanja o utjecaju, upravljanju, etici i odgovornosti ovih tehnologija. Sustavi temeljeni na strojnom učenju donose odluke koje nekada direktno utječu na ljudske sudbine. Iako bi neživi sustavi trebali biti nepristrani, praksa je pokazala da algoritmi često diskriminirajuće tretiraju određene skupine. Računala donose odluke na temelju naučenog, a uče iz podataka koje im priloži čovjek. Pojava diskriminacije u podacima naziva se pristranost (engl. Bias). Ukloniti u potpunosti pristranost iz podataka nije moguće, stoga je potrebno u fazi razvoja i produkciji modela strojnog učenja obratiti posebnu pažnju na detekciju, vizualizaciju i umanjivanje pristranosti. U sklopu ovog rada testirana je XAI biblioteka za programski jezik Python kroz izradu i korištenje web aplikacije testirane na podacima vezanim za odobrenje kredita s ciljem provjere upotrebljivosti.The vast increase in capabilities and the increasing presence of AI-based systems in people's lives raise questions about the impact, governance, ethics and responsibilities of these technologies. Machine-based learning systems make decisions that may directly affect human destinies. Although non-living systems should be impartial, practice has shown that algorithms often discriminate against certain groups. Computers make decisions based on what they have learned, and they learn from the information served by humans. The occurrence of discrimination in data is called bias. It is not possible to completely eliminate bias from the data, so special attention should be paid to detecting, visualizing and reducing bias in the development and production of machine learning models with the goal of making experts aware of the potential consequences. As part of this work, the XAI library for the Python programming language was tested through the creation and use of a web application tested on credit approval data for usability testing