Data Fairness to Find Biases That Influence the Algorithm's Decision Making Results

Tipo
Artigo de evento
Data de publicação
2021
Periódico
3rd European Conference on the Impact of Artificial Intelligence and Robotics, ECIAIR 2021
Citações (Scopus)
0
Autores
Soares L.S.
da Silva L.A.
Orientador
Título da Revista
ISSN da Revista
Título de Volume
Membros da banca
Programa
Resumo
© 2021 3rd European Conference on the Impact of Artificial Intelligence and Robotics, ECIAIR 2021. All rights reserved.The problem addressed in this research aims to ensure that the decision-making process guided by algorithms is equitable. With significant advances in Artificial Intelligence, Machine Learning among others, the algorithm can replace a human being in several tasks. However, there is a discussion about the proper definitions of data justice in which it attracts the attention of researchers from different areas such as: software engineering, law and sociology. Seeking to observe the impacts and influence that data can apply on the results of decision-making algorithms, we used a dataset from an investigative study carried out by ProPublica on the commonly cited Compas tool (Correctional Offender Management Profiling for Alternative Sanctions) in the area of algorithmic bias. Due to the emerging application of decision-making algorithms and other types of models by the scientific and corporate community, coupled with the disposal of various sources to extract or create datasets used to train algorithms, the need for prudence and social responsibility with the results highlights its importance to mitigate the effects of algorithmic discrimination. By investing in the search for the author of the algorithmic bias, one can aim at the structural social inequality present in these data. This due diligence is required for all steps of algorithm construction, from exploring the dataset to development and decision making. Based on the consulted literature, an Exploratory Analysis was applied to the COMPAS dataset to verify the presence of class imbalance, which in this case was confirmed. Subsequently, the Self Organizing Map (SOM) artificial neural network was applied as a visual way to analyze the impact of imbalaced data in the generation of clusters in aspects of class distribution. The experiments were divided into three steps to understand the behavior of imbalaced data: In the first stage, when data is imbalaced by class and intraclass. In the second step, when it is balanced by class. Also, in the third step by intraclass. Thus, it was possible to observe the arrangement of each element in the dataset and its relevance in the generation of the cluster, concluding that when the data balanced by the intraclass has more imbalances, they make the classifications more equal.
Descrição
Palavras-chave
Assuntos Scopus
Algorithmic bias , Algorithmics , Decision-making algorithms , Decision-making process , Decisions makings , Discrimination , Intra class , Kohonen network , Machine-learning , Self-organizing-maps
Citação
DOI (Texto completo)