| < Day Day Up > |
|
In order to more closely approximate the training and testing methodology utilized by neural networks, the jackknife method of discriminant analysis was utilized. This "leave-one-out" principle is a sophisticated method based on estimation with multiple subsets of the sample. In other words, the discriminant function is fitted to repeatedly drawn samples of the original sample. The jackknife method yielded 50.4% of original grouped cases correctly classified. These results are shown in Table 1.
No Fraud Detected | Fraud Detected | |
---|---|---|
No Fraud Detected | 172 | 75 |
Fraud Detected | 123 | 30 |
Ungrouped | 10 | 0 |
50.4% of grouped cases correctly classified 172 + 30/400 = 50.4% 123+75/400 = 49.5% |
In comparison, the neural network method classified 75.9% good parts. The training set consisted of 200 respondents. The data was normalized and ran for one hour and 20 minutes before it resulted in 100 good parts. A weight was obtained, and the network was saved. The testing data set (composed of 210 respondents) was then placed into the saved network and run. After 24 hours, the neural network had achieved 75.9% good parts. These results of the neural net testing are shown in Table 2.
Fraud/No Fraud | |
---|---|
Correctly Identified | 75.9% |
Incorrectly Identified | 24.1% |
Total | 100% |
| < Day Day Up > |
|