UTZ Certified

Machine Learning

UTZ Certified


UTZ certified
Industry: Labeling
Headquarters: Amsterdam, Netherlands


Unsupervised Fraud Detection

Initial situation

UTZ certified is a non-governmental organization (NGO) that certifies products such as coffee, tea, cocoa and hazelnuts from farmers worldwide. This labeling confirms the standards set by UTZ certified sustainable farming practices, social criteria and strict monitoring of the supply chain to ensure transparency throughout the process.


In order to increase transparency even further and to exclude possible fraud that may have been committed at the beginning of the process by providing false information about the harvest, cimt ag supports UTZ certified in using machine learning to exclude it as completely as possible. This is a so-called fraud detection or anomaly detection.


In order to solve fraud detection with machine learning, the availability of the labels, i.e. the labeling of the data, of the data to be examined must first be checked.

A distinction is made between supervised, semi-supervised and unsupervised detection.

In this case, no labels are available at all, so we are directly in the field of unsupervised machine learning. Due to the lack of labels, the unsupervised fraud detection method is the most complex of all three machine learning branches and can lead to inaccurate results under certain circumstances.

Technologies and Methods

To meet the challenges named above, we work with different unsupervised fraud detection algorithms. Algorithms that are based on different principles, e.g. neighbor-, density-based or clustering. The generated results of the algorithms provide a score for each instance, which is merged with all other score results of the algorithms to a final score. Such a method reduces the possible inaccuracy of the unsupervised machine learning method enormously.

Visualization of the outlier score of the algorithm kNN Global

To enable the process to run quickly and smoothly, we used RapidMiner, which is a widely used data mining tool. 

Extract from RapidMiner process


With our unsupervised fraud detection method, we can provide UTZ certified with a stable result, which clearly shows by means of scores which instances in the data set represent a possible fraud. These scores can be converted into labels in the future, so that training of supervised models is possible and provides even more accurate results.


We will gladly advise you

Do you have a similar case? Arrange an appointment with us. We are looking forward to meeting you!


Would you like to learn more?

Scroll to Top