Customer profiling

Question

I have a credit card fraud dataset. there are two populations, transactions that are fraud, and transactions that are not fraud. Can you suggest what ML algorithm I can use to model the main characteristics of these two populations. I need to create two profiles:

For example:

Fraud Transactions - transactions having an amount < 90 $, transactions happen during a particular time in the day

Transactions that are not fraud - transactions having an amount > 90 $, transactions happen during a particular time in the day.

I have used descriptive statistics, and tried to look at these two populations separately. But is there any ML model I can use to distinguish between the two distinctly like in 1) and 2)

I have more than 2 features for each population.

Answer 1

My recommendation would be to start simply and try to build something like a logistic regression model to classify fraud/no-fraud for your dataset. This may help you quickly gain some insight into how to engineer the best features to separate the two populations.

Logistic Regression in sklearn: https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html

An alternate approach would be to explore the data more at this point using something like UMAP to visualize what structure may exist in the data and could be used for feature engineering.

Original UMAP Paper: https://arxiv.org/abs/1802.03426

Example blog post on UMAP: https://pair-code.github.io/understanding-umap/