With the emergence of advanced Internet technology, online banking has become a major channel for business and retail banking. In the last two decades, online banking fraud has been found to be a serious concern in financial crime management for all banking services. Credit card fraud has become a major problem in banking financial transactions and are responsible for the loss of billions of dollars every year. Credit card fraud detection is an interesting issue for the computational intelligence and machine-learning communities. In credit card fraud detection, three aspects namely: imbalanced data, feature selection and selection of appropriate learning algorithms, play the vital role in enhancing the prediction performance. Credit card fraud data sets are usually found to be imbalanced, which results in the classifier to be biased toward majority class. Feature selection is applied as a key factor of credit card fraud detection problem that aims to choose more relevant and nonredundant data features and produce more explicit and concise data descriptions. Furthermore, a suitable learning algorithm can enhance the prediction of fraud in credit card fraud data. In this chapter, SMOTE (Synthetic Minority Oversampling Technique) is employed as an oversampling technique to convert imbalanced data sets into optimally balanced data sets. Furthermore, fuzzy and intuitionistic fuzzy rough sets assisted feature selection approaches are implemented to choose relevant and nonredundant features from the credit card fraud data sets as the fuzzy and intuitionistic fuzzy rough set theories have been widely applied to cope with uncertainty in realvalued data or even in complex data. Moreover, various learning algorithms are applied on credit card fraud data sets and performances are analyzed. Finally, we observe that kernel logistic regression (KLR) is the best performing learning algorithm on reduced optimally balanced credit card fraud data sets for the prediction of fraud. From the experimental results, it can be inferred that the performance of different learning algorithms for the classification of fraud and nonfraud data sets can be easily improved by selecting optimally balanced reduced training data sets consisting of credit card fraud, which can be achieved by suitably modifying the class distribution followed by fuzzy and intuitionistic fuzzy rough set based feature selection techniques.