Multi-objective insights and analysis on data driven classifiers for anomaly detection in water distribution systems

Abstract

[EN] Machine learning techniques have shown to be a powerful tool for extracting and/or inferring complex patterns from data. In the case of the so-called supervised learning, a given learner representation could learn such patterns using labeled data. For example, a helpful approach is to adjust a learner to detect anomalies: historical data can be used, where those events are identified, to find a pattern to classify new data as an anomaly (true event) or not (false event). In this example, the learner's objective is to act as a binary classifier, where a balance between false negatives (predict a typical operation, when in fact an anomaly exists) and false positives (predict an anomaly, when there is not). This balance is attained via an optimization (learning phase), where the learner representation is adjusted. Multi-objective optimization techniques have a natural way of dealing with such problems. They perform a simultaneous optimization of conflicting objectives. As a result, a set of Pareto-optimal solutions, the Pareto front, is calculated. This idea could be used in the training process of binary classifiers. Nevertheless, this requires an integral methodology, merging multi-objective optimization and multi-criteria decision making. While it is true that this idea is not new, methodologies and guidelines are still missing to conduct this process. In this work, we move toward the definition of an integrated methodology of multi-objective learning for binary classifiers for anomaly detection. An anomaly detection database for water distribution systems is used for such a purpose. Preliminary results show to be competitive regarding the F1-score to similar approaches.