Windy Gambetta
Usually, in data science process of obtaining patterns and knowledge from data, accuracy is the sole performance measurement to assess the quality of model it generated. This approach leads to another problem when the model generated unintentionally shows discriminating pattern. The model often bias to the majority class. This problem should be addressed early during the model development phases. We followed a standard approach for building a discrimination aware classifier from datasets with sensitive attributes as described below and use several datasets including an Indonesian criminal cases dataset. For each experiment we use different algorithms and measure the accuracy and fairness of the generated models. Based on the results of the experiments, for pre-processing techniques, it is shown that the Uniform Sampling, Massaging the Dataset, and Reweighing preprocessing techniques tend to reduce the level of discrimination in the dataset significantly while others algorithms even though reducing the level of bias they tends to reduce the accuracy of the system. For in-processing techniques adversarial debiasing, prejudice remover, additive counterfactually fair, and decision boundary fairness measurements gave the best results but their performances are sensitive to the selection of its hyperparameter, for some hyperparameter setting their performances are the worst. For post-processing system, reject option classification gave better results compared to other techniques such as equalized odds post- processing, or calibrated equal odds post- processing. Comparing all results it is found that no technique show the best for every datasets. Some techniques give better results for certain datasets but perform badly for other datasets.
Penerapan Teknologi Tepat Guna, Penerapan Karya Tulis
Data science process of obtaining patterns and knowledge from data, accuracy is the sole performance measurement to assess the quality of model it generated. This approach leads to another problem when the model generated unintentionally shows discriminating pattern. The model often bias to the majority class.