{"id":5226,"date":"2023-02-21T10:18:19","date_gmt":"2023-02-21T18:18:19","guid":{"rendered":"https:\/\/live-cometml.pantheonsite.io\/?p=5226"},"modified":"2025-04-29T14:06:37","modified_gmt":"2025-04-29T14:06:37","slug":"credit-card-fraud-detection-with-autoencoders","status":"publish","type":"post","link":"https:\/\/www.comet.com\/site\/blog\/credit-card-fraud-detection-with-autoencoders\/","title":{"rendered":"Credit Card Fraud Detection With Autoencoders"},"content":{"rendered":"\n
\"credit
Photo by Karolina Grabowska on pexels.com<\/figcaption><\/figure>\n\n\n\n

In this article, we\u2019ll leverage the power of autoencoders to address a key issue for banks and their customers: credit card fraud detection. Feel free to follow along with the full code tutorial in this Colab<\/a> and get the Kaggle dataset here<\/a>.<\/p>\n\n\n\n

Credit Card Fraud<\/h2>\n\n\n\n

Credit card fraud represents an important, yet complex challenge for banks. In 2021 alone, Nilson<\/a> estimated that credit card fraud had surpassed $28.5 billion dollars worldwide. And while effective fraud detection efforts, like the use of machine learning and deep learning algorithms, have helped bring about a steady decline in fraud since its peak in 2016, fraud still represents about 6.8% of international credit card transactions.<\/p>\n\n\n\n

But fraudulent financial data comes with its own set of unique challenges. Fraudulent transactions represent anomalous data that make up a very small percentage of total transactions. This results in significant class imbalances that make many traditional evaluation metrics, like accuracy<\/a>, irrelevant.<\/p>\n\n\n\n

\n

If you aren\u2019t already, you can follow along with the full code in this Colab<\/strong><\/a>.<\/p>\n<\/blockquote>\n\n\n\n

The Data<\/h2>\n\n\n\n

Our data consists of two days\u2019 worth of real bank transactions made by European cardholders in 2013. In this dataset, fraud makes up just 0.172% of all transactions, meaning a model that predicted \u201cno fraud\u201d for every single observation would still achieve an accuracy of over 98%! In the pie chart below, fraudulent transactions are represented by the tiny orange sliver.<\/p>\n\n\n\n

\"A
Image by author; note the tiny proportion of fraudulent samples<\/figcaption><\/figure>\n\n\n\n

Additionally, in order to protect all personal identifiable information (PII<\/a>) in this dataset, all features (besides Time<\/code> and Amount<\/code>) have undergone a PCA<\/a> transformation. This means that we cannot use any domain knowledge for the purposes of feature engineering or selection.<\/p>\n\n\n\n

Data Visualization and Pre-processing<\/h2>\n\n\n\n

Let\u2019s start out by visualizing our data. It would be impossible to plot 30 dimensions, so first we\u2019ll apply Scikit-Learn\u2019s implementation of t-SNE<\/a> to the data. t-SNE<\/a> is a dataset decomposition technique. Here we only plot the first two components with maximum information.<\/p>\n\n\n\n

\"Scatterplot
Image by author; the two classes are effectively indistinguishable<\/figcaption><\/figure>\n\n\n\n

We can observe in this graph that there is little difference between fraudulent and non-fraudulent transactions. Most machine learning models would struggle to classify this data as-is.<\/p>\n\n\n\n

First, we\u2019ll perform some very basic transformations to the data before feeding it to our model. The original Time<\/code> feature represents the number of seconds elapsed between each transaction and the first transaction in the data. We\u2019ll convert this relative time measure to hour-of-the-day.<\/p>\n\n\n\n

We also scale the Time<\/code> and Amount<\/code> features, as all other features (V1<\/code>, V2<\/code>, \u2026 V28<\/code>) were previously scaled during the PCA transformation. Finally, because we are training an autoencoder to \u201clearn\u201d the embeddings of a normal transaction, we\u2019ll subset just normal samples as our training dataset, and use a 50:50 split of normal and fraudulent samples for our validation set. We\u2019ll also set aside 250 samples for a test dataset of completely unseen data.<\/p>\n\n\n\n

 <\/p>\n\n\n\n