Evaluating Classification Models Performance - Accuracy Paradox