Article Preview
Top2. Recent Development In Banking Fraud Detection
In the field of credit card fraud detection and computer intrusion, most of the work carried out so far use learning techniques, such as SVM, NN, Decision Trees, Logistic Regression and Markov models etc.
SVM binary classification is well suited for detecting the legal and illegal frauds. SVM has the advantage over NN when the learning objective is a non-convex problem with multiple local minima, in which a NN tends to get stuck at a saddle point (Cristianini and Taylor, 2000; Vapnik, 1999). Another weakness of a Neural Network model is its large amount of hyperparameters to be tuned (Cristianini and Taylor, 2000). Sahin and Duma (2011) explore the use of SVM for credit card fraud detection based on feature engineering and profiling techniques, and then compare the model with a decision tree. The results show that SVM generalise better than the decision tree with a limited dataset. However, the growth in dataset size resulted in its performance decreasing. Several studies (Benard, 2007; Ando, 2016; Li, 2017; Batani, 2017) address the fraud detection as a sequence classification problem. A variant of Recurrent Neural Networks (RNN) called Long Short-Term Memory (LSTM) is used to profile the customer behaviours over timesteps. It processes sequences of values and shares the parameters across different part of the model (Goodfellow et. al., 2013). The statistical strength of sharing over timesteps across different input sequence lengths allows it to be generalised well for time-series problems. The state of the system at given point in time can be given by:
where S
t is the state of the system at time t,
is a deterministic function and X
t is the input vector. Ando et al. (2016) apply RNNs and SVMs to profile fraudulent behaviours from web log data and show that the RNN model with LSTM outperforms other models. Wise (2007) also demonstrates the advantage of LSTM to model customer behaviours to better detect the credit card fraud over Feed-Forward Neural Networks and SVM models.