Detecting Exceptions

Most systems detect anomalies. Unfortunately, most anomalies are false positives (e.g., alarms with no fire). As false positives dominate alerts in a system, people begin ignoring them. We call this the alarm syndrome. 

When the alert is wrong and the person reacts, this is a Type 1 error. Conversely, when an alert is right and the person ignores it, this is a Type 2 error. We develop apps intended to reduce both Type 1 and 2 errors. In this case, the app does more than detect exceptions, it helps people more accurately accept detections.

We apply combinatorics to decrease the false positive rate on alerts

We then guide analysts through a machine learning based diagnostic

We track the performance of both the system and analyst in correct calls

Then make adjustments and continue monitoring anomaly detection errors