Fraud detection is considered an archetypal application of machine learning within the banking sector. However, whilst the internet is littered with jupyter notebooks and blog posts tackling the problem as one of imbalanced classification, in real-life applications financial institutions detecting fraud often do not have the luxury of labelled data. The reason for this is partly practical – it can be time consuming even for a domain expert to label a transaction as fraudulent or legitimate – but also of a more fundamental kind – fraudsters are constantly adapting their behaviour so as to avoid detection, meaning that a model trained on yesterday’s fraudulent activities may not be sufficient to detect today’s fraud.
Van Lanschot Kempen would like the students to build an anomaly detection algorithm to detect suspicious transactions in an anonymized credit card transaction dataset. Assessment of the students’ work will consider two elements:
(1) Guided by their models, they will choose the 100 most suspicious transactions within the dataset and these will be compared with the true labels to calculate a precision score
(2) Students will prepare a short presentation describing their methodology for anomaly detection. Presentations highlighting novel approaches and/or attempts to visualize anomalies will be scored highly.
The dataset is located in the /data/ folder of this repo. It consists of 284,807 transactions, each of these
transactions is associated with a euro value ('Amount') and a unique ID ('Transaction_ID'). Each transaction is further
characterized by 28 features, labelled X01-X28. No data cleaning is required of this dataset before starting your
analysis.
If you're familiar with the git lfs library, you can simply clone this repo and the data in it, if you're not familiar
with this, simply download the data from the repo by right clicking on the download button, and choosing Save Link As as shown below:
then
Once you have your list of 100 transactions, rank them from 0 to 99 (where 0 is the transaction most likely to be
fraudulent (i.e. is the biggest anomaly)), save this as a csv with the file name team_<your-team-name>.csv and send it to
a.biffin@vanlanschotkempen.com. You can see an example file in the /submission/ folder if you are unsure about how it should be formatted.
As well as your preferred data/analytics libraries (e.g. pandas, numpy, scipy), some useful python libraries for solving this usecase include:
- Note that the level of contamination (i.e. proportion of transactions which are fraudulent) in datasets such as this is very small, around the 0.1% level.
- If your dataset is too large to visualize, you can try picking out the transactions you are most confident are anomalies and label them '1', then sample from the remaining transactions and label them '0'. Plotting both these populations might give you away to visualize anomalous transactions.