Sára Jurovatá

· Applied 3 ML models (Logistic Regression, Random Forest, and LightGBM) and computed fairness metrics such as Demographic Parity Ratio, Statistical Parity Difference, and Equalized Odds Difference using AIF360 and Fairlearn on both datasets. · Implemented 6 mitigation techniques (preprocessing, in-processing, postprocessing) and compared the performance of the frameworks on both datasets. · Fairlearn showed a slight advantage in metrics, along with superior usability and visualization features.

· Trained 4 predictive ML models (Linear Regression, Decision Tree, Random Forest, and LightGBM) on weather (temperature and air quality) data. · LightGBM was the best-performing model, explaining 82.8% of the variance, with a MAE of 2.463. · After feature selection, the LightGBM model improved slightly, explaining 83.3% of the variance with a lower MAE of 2.437. · Used SHAP to identify features like ozone levels, date, and SO₂ as the most influential in the predictions.

· Processed over 23 million anonymized transactions to identify fraud in a highly imbalanced dataset. · Used ROSE to undersample the majority class and SMOTE to synthesize fraudulent cases, balancing the dataset. · Compared 4 machine learning models (Logistic Regression, Naïve Bayes, Random Forest, and XGBoost) using performance metrics (accuracy, recall, precision, and F1 score), along with the ROC curve. · XGBoost balanced recall and precision with faster training, ideal for scaling. Random Forest had higher recall (73.16%) but more false positives.