Bank firibgarligini aniqlashda gradiyent busting modellari asosidagi qavariq vaznli ansambl: IEEE-CIS Fraud Detection to‘plamida tajriba

Detecting online payment fraud is challenging under class imbalance and temporal concept drift. On the IEEE-CIS Fraud Detection dataset (590,540 transactions, 3.5% fraud), five models and their ensemble were evaluated using a time-based split. A stacking ensemble with a logistic-regression meta-learner achieved lower ROC-AUC (0.8369) than every individual model: the unconstrained meta-learner overfit the training period and failed to generalize. Replacing it with a convex combination (non-negative weights summing to one) raised ROC-AUC to 0.9235, exceeding the best individual model. The results show constrained ensembles are more robust under temporal distribution shift.

Hali tarjima qilinmagan