How can I optimize my credit card fraud outlier detection process?
Hello everyone,
I am very new to rapidminer. Im currently working on a process in which I wish to detect credit card fraud with one of the detect outlier operators. I have found the best success with the Densities operator. I have taken a sample of 1000. The denseties operator finds there to be 381 outliers and 619 not outliers. Actual amount of Fraud is at 83 though. How can I optimize my process, so there are not as many transactions getting flagged as outliers when they aren't fraudulent? I am aware that maybe a different operator/ process could be more efficient but I am tasked with operating on the detect outliers. Any input would be helpful, thank you very much!
Explanation of steps: Numerical to binominal to change "fraud" to true/ false, Set role to put fraud as the label, sample the size to 1000, normalize the data, cross validation with decision tree to see how it does with deciding on true/ false, finally detect outliers (distance) with distance 1.0 and proportion 0.95 and squared distance.
The data set i use.
Answers
-
Hello there,
To improve your process, try tweaking the settings of the Densities operator. Lowering the distance threshold can help reduce false alarms, and decreasing the proportion of outliers may filter out less important cases. Make sure your data is scaled properly (normalized) so that no single feature has too much influence, and focus only on the features most related to fraud by removing irrelevant ones.
You can also make the data simpler by using methods like PCA to remove unnecessary details.
I’d also suggest using outlier detection as a first step to identify potential fraud cases, then running those flagged cases through a Decision Tree model to confirm if they are actually fraudulent.
0
Altair Employee