We are trying to explain data anomalies, says BGU professor

Professor Bracha Shapira explains in the framework of the "Intelligence - Cyber, Digital  & IT" conference how to explain anomalies found in data 

A screenshot from Professor Bracha Shapira's speech at the "Intelligence - Cyber, Digital  & IT" conference. 

Professor Bracha Shapira, Vice Dean for the Research Faculty of Engineering Sciences at Ben Gurion University, explains that it is possible to find anomalies more efficiently using artificial intelligence or machine learning. In research she is conducting with other researchers, she is trying to find methods of describing anomalies in non-contextual analysis.  

What is an anomaly? Well, the goal is to look for hidden connections. There are three approaches – contextual (study of the system using examples), semi-contextual (not enough examples, comparison to normal model) and non-contextual (identification of irregularities without preliminary study). In deep learning, the non-contextual method, you receive data input, there is smart compression of the data, and reconstruction of the input in output. When there is an anomaly, the reconstruction will show an error and then you know that there is an anomaly. Another field is explanation of the anomaly that the algorithm found.   

"If we didn't succeed in reconstructing the input, we try to understand what happened," said Shapira. "We look where the error was, in which characteristic, and we try to predict it. And then it is possible to consider why the characteristic was far from or similar to the anomaly." The researchers are still trying to identify what caused the anomaly.  

You might be interested also