Modeling Financial Fraud Detection Using Radial Basisfunction Network

Fraudulent activities with intent to defraud are pervasive in every system. The traditional statistical and long investigative auditors’ methods of fraud detection are becoming obsolete in dealing with the present trend of financial malpractices. The existing fraud detection methods are inadequate in terms of cost and time to detect emerging frauds. This study aimed at developing an enhanced Radial Basis Function (RBF) network to efficiently detect financial frauds. The specific objectives were to: (i) determine the input features relevant to fraud occurrences; (ii) create a series of RBF models; (iii) select an optimum model amongst the candidate models; (iv) evaluate the chosen model; and (v) compare the model with some existing ones.rnThe methodology employed RBF algorithm with radially symmetric Gaussian activation functionsto classify previously unseen data features into their respective true categorical classes. A set of 20 input attributes, minimum hidden nodes and weight adjustments were observed for different runs until the network became divergent. The model was implemented in Rvectorized software packages and was trained 200 times with 1000 online German bank credit transactions. The developed RBF model for fraud detection was compared with three other models: Multi-layer perceptron back-propagation (MLP), Dynamic decay adjustment (DDA) and General nonlinear regression (GNLR).rnThe findings of the study were that:rn i. twenty input features were ranked in order of their importance relative to the network’s output; 4 of the variables, namely, credit history, nationality, guarantor and job status were the most important in that order, while 5 variables: accounts type, telephone number, co-applicant’s guarantor and savings bond are the variables with negative contribution;rnii. a total of 200 RBF base models were generated with randomly selected 850 credit transactions out of 1000 at ten different epochs and hidden nodes;rniii. the RBF model for fraud detection was developed and of the 200 candidate base models, an optimum modelwas obtained at 600 iterations alongside 840 hidden nodes with a misclassification error rate of 6.9% and about 93% degree of accuracy. RBF models are of the form:rny(x)= ∑_(i=1)^h▒〖ω_ij φ_i 〗 (x), where ω_ijis output weight, and φi(x) is the Gaussian activation function, 1/(1+e^(-φ_j (X)) ), with output y∈ {0,1}; rn iv. the chosen RBF model was evaluated and its performance yielded 7.18% misclassification error rate, 92.82% prediction accuracy, 89.71% sensitivity and receiver operating characteristic (ROC) of 98.6 %; and rnrnv. RBF had the highest average accuracy of 75.90% compared to DDA (75.04%), MLP (74.60%) and GNLR (74.53%). Further, RBF and DDA attained the highest accuracy score of (85.71%) and (79.22%) respectively, at iteration 4, while MLP and GNLR attained their highest accuracy at iterations 7 and 5 respectively. rnThe study concluded that RBF network compared to other similar models trained faster, minimized the cost and time of fraud detection. The study recommended that RBF network is an efficient fraud detection model and its parameters must be adjusted closely to zero tolerance as a little threshold tolerance could imply a significant cost.

Get Full Work

Report copyright infringement or plagiarism