1
$\begingroup$

I'm about to start a project regarding Fraud Detection in Insurance but there is no dependant variable to train the model and classify it.

Please provide the algorithms that I can use to detect Fraud in both Auto and Health insurance industry.

Thanks in advance!

Improve this question
$\endgroup$
2
  • $\begingroup$ I understand it could be a business constraints, but it seems like a supervised problem naturally. Just assume the ML is a very smart person very good at sorting things. So you want that person to tell you what records are possibly fraud, without telling her how you would define as fraud. If you have ways to verify, then you can probably collect data for learning. If you have no ways to verify... hum.. not sure how you can optimize your model $\endgroup$
    – The Lyrist
    Commented Jan 22, 2019 at 18:30
  • $\begingroup$ Thanks for your answer! Sure, it's a supervised problem at the origin. But when you look at Credit Card Fraud detection they use Isolation Forest to detect anomalies in the database, it's unsupervised. Is there any method (semi-supervised for example) to detect fraud in insurance Claims? $\endgroup$
    – Soufiane S
    Commented Jan 22, 2019 at 20:05

4 Answers 4

Reset to default
2
$\begingroup$

Fraud, by definition, is a supervised concept.

No unsupervised algorithm will reliably detect this. At best you'll get some suspicious cases with unsupervised approaches such as anomaly detection.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Yes I know... But the problem is, there is no Fraud/Not Fraud data in the database. Can we create a Decision Tree manually? If we know what variables are contributing to Fraud Detection... $\endgroup$
    – Soufiane S
    Commented Jan 23, 2019 at 9:57
  • 2
    $\begingroup$ Well. The algorithm is not going to come up with a definition of fraud for you. It's more likely to split the data into women vs. men. Unsupervised learning can't do magic... $\endgroup$
    – Has QUIT--Anony-Mousse
    Commented Jan 23, 2019 at 23:08
2
$\begingroup$

Usually it is a binary classification problem, i.e. supervised learning. One thing to take into account in these projects is the imbalance betweeen FRAUD - NO FRAUD labels, being the former instance much less frequent. But with no labels, I would start by doing some feature engineering and applying PCA, to check the existance of instances getting slightly away.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thanks a lot for your answer! PCA is used to reduce dimensionality, right? So you are suggesting reducing the dimensions to - let's say - 4. Then what do we do next? Can we apply PCA to hetorogenous data knowing that not all variables are quantitative... I will give it a try and let you know... $\endgroup$
    – Soufiane S
    Commented Jan 23, 2019 at 9:54
  • $\begingroup$ Hi again, I have tried to apply PCA (sklearn) to the heterogenous dataset but as expected it didn't work... because the values must be quantitative... Is there any approach to solve this issue? Or maybe a method to convert qualitative variables to quantitative? $\endgroup$
    – Soufiane S
    Commented Jan 23, 2019 at 11:10
0
$\begingroup$

Autoencoder can be used for finding anomalies (via the reconstruction error) and therefore are also applied in fraud detection. One example with the goal to detect accounting fraud with autoencoder is this paper: Detection of Anomalies in Large Scale Accounting Data using Deep Autoencoder Networks

However, you need some way to validate the results you get from this approach. If you have no labeled data at all, someone with expertise would need to look at the anomalies detected to check if they are really fraud or not.

Improve this answer
$\endgroup$
1
  • $\begingroup$ Ok sir, I will try autoencoders. $\endgroup$
    – Soufiane S
    Commented Feb 8, 2019 at 0:25
0
$\begingroup$

I think clustering techniques can be applied here.

Think of it this way, the fraud points will most likely not be part of any clusters as the majority of points will not be fraud.

Check out Local Outlier Factor, you will get an idea.

Also if you have the labels beforehand, run the unsupervised algorithm and then see how well it catches the fraud cases(outliers).

Improve this answer
$\endgroup$
1
  • $\begingroup$ Thanks for your answer, I will look into that! $\endgroup$
    – Soufiane S
    Commented Feb 8, 2019 at 0:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.