Influential ML: Towards detection of algorithmic influence drift through causal analysis

Abstract

Machine Learning (ML) algorithms are applied in environments where they can make increasingly more impactful decisions. Such decisions can have great power over the classified instances. Therefore, these classifications need to be accurate, reliable, and explainable. Pushed mainly by regulations, a new ML field for research and practice is gaining traction: MLOps. This consists of practices to reliably and efficiently deploy and maintain ML models in production. Influential ML is part of MLOps, which focuses on understanding the consequences of an ML algorithm. One of these consequences is prediction influence drift, whereby the classifications of an algorithm are the cause for changing instances over time through feedback loop effects. This research aims to create a detection approach that can identify, quantify, and possibly localize prediction influence drift. The result is a developed detection approach that uses causal inference techniques to detect self-fulfilling and self-defeating prophecies on simulated synthetic data. The evaluation shows that this approach is promising.