A Predictive Analytics Framework to Anomaly Detection

Online fraudulent activities cost the global economy billions of dollars each year. In the world of big data from large number of transactions, detecting and preventing anomaly cases like fraudulent activities is an essential ingredient to success not only to big banks, insurance companies, small businesses, but also to consumers. Forms and levels of sophistication of online fraud are constantly changing, hence there is a need for a flexible framework that can predict and adjust to new fraudulent cases. One major problem in building a predictive model for anomaly detection, is the scarcity of fraudulent data records in comparison to non-fraudulent, which makes the training data imbalanced. In this paper, we present a comprehensive predictive analytics framework that aims at detecting anomaly cases and most importantly mitigating the problem of imbalanced datasets in training anomaly detection models. The framework is developed through experimentation of an ensemble of sampling algorithms, feature engineering methods, and an array of machine learning algorithms. The framework presented in this study provides a data-driven approach applied to financial credit card data and can be adapted to other domains where deviations from what is normal occur in large datasets.