Detection of Abnormalities by Means of Autoencoders

Anomaly Detection in Accounting Data

Machine Learning (ML) for Anomaly Detection

A major media topic in 2018 was the fraud cases in the German care system. On 16.10.2018 wrote Spiegel, that according to expert estimates, "around two billion euros will be lost annually through fraud in outpatient care alone". The interest in counteracting this is high. Finding fraud cases well in data therefore appears to be a worthwhile goal.

We have set ourselves the task of analysing, optimising and implementing ML algorithms for detecting conspicuousness. In particular, we focus on assistance systems for users. This branch of research is called Anomaly Detection.


Solve Statistical Problems With Machine Learning

Machine Learning (ML) today offers solutions and assistance in various areas that influence our everyday life. For example, speech and image recognition on smartphones works impressively well and autonomous driving is no longer just theory. These techniques are also finding their way more and more into originally statistical domains. For example, ML algorithms can be used to calculate credit ratings.

The hope is always that ML algorithms will solve the tasks of classical statistics better or at least more conveniently. This development is also politically desired and encouraged.

The focus is e.g. on the explainability of such algorithms or on applicability in everyday consumer life.

Adversarial Autoencoders Offer the Possibility to Visualize a Large Variety of Data

Specifically, we are currently working on finding a robust and interpretable representation for large amounts of data - with an unmanageable number of characteristics. Above all, this should enable the user to compare data observations with each other and put them into relation. For this purpose we use for example Adversarial Autoencoders.
These models from the group of Unsupervised Learning offer the possibility to actively influence the presentation of data.

The picture shows the development of the data display during the training of one of our autoencoders. Conspicuous objects of our data set are marked dark.

The property »conspicuous« is usually not known, but you can see how exactly these objects gather around the edge. Furthermore certain coherent structures can be seen.

Detektion von Auffälligkeiten mit Hilfe von Autoencodern