A proposal to reduce false alarms in anomaly detection systems with counterfactuals generated through multi-objective constrained optimization

Abstract Anomaly detection systems are a field of study responsible for implementing techniques capable of detecting abnormal changes in time series. These abnormal changes have a particular usage in industrial applications such as water monitoring and manufacturing. One of the challenges in these systems is to reduce the number of false-positives: it is not desired to trigger false alarms to the users frequently. In order to solve this challenge, we propose a strategy using a counterfactual generation algorithm to be applied as a post-processing method on a previously trained, standard anomaly detection anomaly. This counterfactual generation algorithm is modeled as a multi-objective optimization constrained problem. These counterfactuals are then evaluated by a classifier trained to find members of false-positives. Results show that this proposal can successfully reduce the number of false-positives.