Anomaly Server typically takes an input which is comprised of rows and columns. One row for each record and several observations (or dimensions) for said record.


The complete dataset can be submitted directly from a selection in Microsoft Excel or via an API call.


Anomaly Server returns a value between 0 and 1 for each record indicating how anomalous this record is in the context of all other records in the submitted dataset.


Anomaly Server combines machine learning and statistical methods for calculating the anomaly score, but the main principle is that the harder it is for Anomaly Server to memorise (or classify) the differences between points, the less anomalous they are.


The observations that are easy to remember are outliers.


Imagine you from a balcony look down on a room filled with hundreds of people at a conference reception.


Once you look away you may not be able to produce a description of everyone detailed enough, so a colleague can positively identify who you are talking about.


But if there was a person 6’ 8” tall, in a fluorescent suit wearing a yellow hat, chances are you might remember that person quite easily. That’s because the person is an anomaly in this context.


Once you look away you may not be able to produce a description of everyone detailed enough, so a colleague can positively identify who you are talking about.


On the other hand, the more like the rest of the group a person is, the longer a description it will require to positively identify them. That’s because they are (much) less anomalous in this context.


Anomaly Server works by calculating how much effort is involved in distinguishing each record from the rest. The less effort required, the more anomalous they are, and vice versa.


Anomaly Server is domain-independent and does not require training with known anomalies.