PREPROCESSING TECHNIQUES FOR HANDLING DISCRIMINATION IN BINARY LABEL DATASETS

• Main Author: Zabrina Pramata, Nella
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/51429
• Institution: Institut Teknologi Bandung

Data Mining is a process for gaining pattern and knowledge from data (Han etc., 2012). This process can help users because it can be used as a consideration to determine the next business steps. However, the prediction results are not 100% reliable. One of the reasons is the possibility of unfairness in the results of the prediction made by the model. The unfairness in the results could occur because the training data that is used for training process contains sensitive information. The pattern obtained from the training process is influenced by sensitive information which could potentially cause discrimination against that sensitive information. That way, this result is likely to harm the specific groups of people due to discrimination. The Artificial Intelligence system that could potentially discriminate against sensitive information is called Unfair Artificial Intelligence which will henceforth be abbreviated as Unfair AI. Several ways have been proposed by various researchers to handle unfair AI in the form of techniques in the preprocessing, inprocessing, and postprocessing stages. In this research, the Fairness AI techniques to handle discrimination is focused on the preprocessing techniques so that sensitive attributes in the dataset can be handled before the training step. As for the dataset used, only focuses on Binary Label Dataset. The Fairness AI preprocessing techniques used in this research consisted of Uniform Sampling, Preferential Sampling, Preferential Sampling, Massaging the Dataset, Reweighing, Suppression, and four modified techniques of Suppression. Based on the results of this study, it is known that Uniform Sampling, Massaging the Dataset, and Reweighing techniques tend to reduce the level of discrimination. However, the other six Fairness AI preprocessing techniques can still be used to reduce the level of fairness even though the results are not always effective. Based on these results, there is no most suitable technique for all datasets so that these nine techniques still need to be used to find out which technique is suitable for the dataset to be used.