AI Case Study

eBay research identifies 40% of credit card fraud with high precision automatically using machine learning

eBay researchers have developed a method using machine learning for anomaly detection to automate detection of fraudulent credit card transactions. They are able to identify 40% of fraudulent instances with high precision. This has obvious beneficial implications for eBay, as the global auction and retail platform experiences large numbers of transactions everyday.



Internet Services Consumer

Project Overview

"We hypothesize that unlike bad behavior good behavior does not change with time and data points representing the good behavior have consistent spatial arrangements under dierent groupings. Based on this hypothesis we propose an ensemble of clustering methods for outlier detection.

Reported Results

The researcher's method was able to "identify 40% of the fraud cases with the proposed algorithm with high precision. This is very useful in industry setting where a high percentage of fraudulent transactions are detected without any manual judgement... our proposed method is highly accurate in identifying true negatives or good consumer behavior. Our method can be immensely helpful as out of 284,807 samples we can safely rule out 139220 as true negatives based on threshold of 0.5. We will make few mistakes that is acceptable in industry at this scale. Moreover, we can also identify about 40% of the false positives with recommending only about 2000 data samples."


"Consider an application of credit card fraud detection. Here, it
is easier to obtain samples with good non-fraud behavior than
samples that exhibit a fraudulent pattern as the latter is scarce and time variant. Once a fraud pattern is accurately determined it is just a matter of time before fraud shifts to a different area exhibiting a totally different pattern. However, if we can develop a method that can estimate a measure of consistent behavior (good behavior) for each data point then we can identify outliers as data points with low consistency score. We refer to the data points that exhibit good non-fraud patterns as consistent data points. We consider data points that belong to the same cluster (or close proximity clusters) under different spatial groupings as being consistent. This is especially true when the data points have features that are highly correlated. Based on this understanding, we attempt the problem of outlier detection by estimating a consistency score for every data point. One way to estimate the consistency for every data point is by running an ensemble of unsupervised clusters. Since the spatial arrangements of consistent points do not change, they should form closed clusters. We run the simple k-means clustering algorithm N times for different values of k where k can range from 2 to K. The N runs of k-means on the sample set will assign each data point to N clusters with N centroids associated with their respective clusters."





"Often the challenge associated with tasks like fraud and spam detection is the lack of all likely patterns needed to train suitable
supervised learning models. This problem accentuates when the
fraudulent patterns are not only scarce, they also change over time. Change in fraudulent pattern is because fraudsters continue to innovate novel ways to circumvent measures put in place to prevent fraud. Limited data and continuously changing patterns makes learning significantly difficult. In credit card fraud detection, the situation aggravates by the fact that the fraudulent behavior patterns keep changing. This is because fraudsters keep innovating novel ways to scam people and online systems. Combination of changing patterns and fewer labeled data points makes it an extremely challenging problem to keep any marketplace safe and secure."



"The dataset contains credit card transactions made in September
2013 by European cardholders in two days. It has 284,807 samples of which 492 samples are fraudulent transactions. The dataset is highly unbalanced as the positive class samples are only 0.172% of all data points. This dataset is anonymized and contains only numerical input variables which are the result of a PCA transformation. It has 30 features out of which the only features which have not been transformed with PCA are Time and Amount features. Time is the seconds elapsed between each transaction and the rst transaction in the dataset and Amount is the transaction amount. Class is the response variable and it takes value 1 in case of fraud and 0 otherwise."