Grant: Discovery Grants Program - Individual

Our program will develop an adaptive learning framework that addresses the data-stream challenges of e-payment and fraud detection systems, such as rapid and continuous data flow, highly imbalanced data and changing data distributions. The hidden and unpredictable data changes are known as Concept Drift (CD). We focus on the e-payment sector because it incurs monetary losses for users in case of fraud and studies on CD are few and general. Learning in such an evolving environment is challenging. Therefore, we will investigate this sector to understand and determine the CD characteristics. CD occurs in legitimate and fraudulent data because users' spending habits often change and fraudsters update their strategies to keep pace with new technologies. Monitoring changes in user behavior and responding to those changes is critical for preserving the e-payment models from becoming obsolete and the prediction quality from deteriorating. Our framework consists of: 1. An ensemble-based CD detection approach that captures explicitly CD characteristics in new data, which are unknown in the e-payment sector. CD is a complex notion involving several characteristics, such as the type (real, virtual and mixed), transition (sudden, gradual, incremental and reoccurring), severity level and drift region.  These characteristics are essential for re-calibrating the e-payment models and acquiring the most contributing training data. 2. A cost-sensitive deep neural network approach that handles the dynamic, extreme class distributions. Learning with fraud data is problematic because the class distribution is highly imbalanced,  classifiers misrepresent the fraud class that is the most critical and the class imbalance ratio varies over time. Cost-sensitive learning is powerful to tackle highly imbalanced data, but unexplored for data stream. We will propose an approach to adapt the costs dynamically for both classes and include the class costs into the neural-network loss function. 3. An adaptive feature learning approach that responds to the extracted CD characteristics. E-payment models that learn under false data assumptions will fail because the past concept is no longer valid. We will define several strategies to update the deep neural network to the various CD characteristics. Lastly, we will incorporate the cost-sensitive learning scheme into the adaptive learning approach.   We will conduct extensive empirical studies to evaluate our approaches using real datasets. We will also use synthetic datasets by varying their sizes, dimensionality, CD characteristics and class imbalance ratios. We will assess how CD impacts the prediction models and how they recover from it. Our framework will adjust to the ever-changing factors, such as e-payment channels, products, sellers and fraud strategies, providing better tools to detect financial fraud. Accounting for these factors is crucial to build highly accurate e-payment systems and to protect them from fraud.