topic Re: Fraud Detection Using Supervised Machine Learning in SAS Data Science

Fraud Detection Using Supervised Machine Learning

hakonstrand — Fri, 15 Mar 2019 09:29:38 GMT

Hello!

I am a student working on a project trying to identify fraud in E-Commerce transactions.

The fraud(target) are rare events (around 5 % of the observations) which leads the model to classify everything as not fraud.

I am thinking of a method of oversampling without losing all the observatoions. When I'm using sample to make observations of fraud and not fraud 50/50, the program takes only the fraudcases, 2000 observations, and 2000 random non fraud observations, making me lose the rest of the dataset, which is almost 25000 observations.

Is there a way of making the dataset 50/50 (fraud, not fraud) without removing observations? My thought is that it might just duplicate fraud observations in the training dataset (making it 50%/50%), leaving the test dataset as it was (95% not fraud, 5% fraud). Any step-by-step method of doing this in SAS?

I also have a lot of variables, almost 600. I am thinking of using PCA to get the most relevant of these. How can I do this using SAS?

Re: Fraud Detection Using Supervised Machine Learning

Ksharp — Fri, 15 Mar 2019 13:19:15 GMT

Use PROC PLS or PROC HPGENSELECT to pick up variables.

Re: Fraud Detection Using Supervised Machine Learning

MBRACH — Tue, 26 Mar 2019 15:03:19 GMT

In here you have some ideas how to deal with rare cases:

http://support.sas.com/documentation/cdl/en/emxndg/67980/HTML/default/viewer.htm#p1w6fewo0jhzxdn1rytuk1kt0pqj.htm

In here is a paper about nice method called SMOTE but unfortunately this version works only on continuous variables:

https://www.sas.com/content/dam/SAS/support/en/sas-global-forum-proceedings/2018/3604-2018.pdf