The In-depth 2020 Guide to E-commerce Fraud Detection

Learn how Machine Learning can provide Fraud Prevention and Protection for E-Commerce industries. These algorithms can work as well for any other industry, but in this case we look in the scope of just E-Commerce. If you have any questions, suggestions or comments - feel free to contact me. I would love to discuss this topic with you and learn something new myself!

E-commerce Fraud Solutions with Machine Learning
We know that conventional rule-based solutions work according to specific rules written by programmers, which does not allow them to be flexible and smart with new fraud patterns. At the same time, E-commerce fraud solutions built with Machine Learning improve themselves over time with the input of new information; in other words, they can “learn.”

There are two major classes of Machine Learning algorithms — supervised and unsupervised. Both can be used for fraud detection and prevention, but each has its pros and cons.

Machine Learning grounded detection solutions scan transactions and evaluate their threat score, such as between 0 and 1. The score is then compared to a pre-established threshold that will mark the transaction as fraudulent or not. Let’s take a closer look at the nature of some of these algorithms:

Supervised Decision Tree
After being fed data on fraudulent and normal transactions, a supervised Decision Tree will then make a classification (a prediction). The fraudulence score computation starts from the root node of the tree when it is split into child nodes; other nodes are also split into child nodes with binary or multi-fashion conditions. This is done depending on the value of the input variable.

When the tree is built, a new data input (a transaction) is classified by going through the root of the tree starting from the root node according to the feature values of the input.

Supervised Support Vector Machine (SVM)
A Support Vector Machine (SVM) works in another way — it separates transaction data samples into two classes on a plane graph in such an order that the formula needed for it shows the smallest error as compared to the ground truth dataset (real transactions labeled). The main idea behind an SVM is to draw a line between classes that will leave the biggest margins between fraudulent and non-fraudulent transactions to achieve a high level of detection.

Anomaly Detection Using Autoencoder
In the event that a customer has a very few examples of fraudulent transactions, it is better to use Autoencoder — where fraudulent samples are excluded on the step of model training, but are still used for testing. All anomaly detection techniques are aimed at denoting unusual or unexpected events in the data.

A neural autoencoder is a type of architecture that is trained on one class of events and used to notify us about unusual events. The process of training implies an equal number of input and output units that have a certain number of layers in between. The final decision on whether a transaction is fraudulent or not is based on the threshold value and the distance between the input and its reproduced output layer.

Outlier Detection: Isolation Forest
The other technique that tackles cases where there are very few or no fraudulent transactions in a dataset is Isolation Forest, which belongs to the outlier techniques class. The idea behind the Isolation Forest is that the outlier can be defined through making less random splits than a data point that belongs to the normal class; outliers happen much more rarely than normal samples and have values that are not typical for the average values of a data set.

The algorithm chooses a split value out of a randomly selected value range of a randomly selected feature. As a result of the selections, a tree is grown. The tree depth is measured with the number of required random splits (called mean length). When a forest consisting of such trees is grown, the mean length number is measured over all trees and becomes a measure of normality, or in the other words, the function we use to trace outliers.

Random splits have significantly shorter tree depth in cases with outliers than in cases with normal data samples. This helps us identify which data points are likely to be outliers.

Explore the full article here: https://spd.group/machine-learning/e-commerce-fraud-detection/