Partially Monotonic Learning for Neural Networks

In the past decade, we have witnessed widespread adoption of Deep Neural Networks (DNNs) in several Machine Learning tasks. However, many critical domains, such as healthcare, finance, or law enforcement, transparency is crucial. particular, lack ability to conform with prior knowledge greatly affects trustworthiness predictive models. This paper contributes DNNs by promoting monotonicity. We develop a multi-layer learning architecture that handles subset features dataset that, according knowledge, monotonic relation response variable. use two alternative approaches: (i) imposing constraints on model’s parameters, and (ii) applying an additional component loss function penalises non-monotonic gradients. Our method evaluated classification regression tasks using datasets. model able known relations, improving decision making, while simultaneously maintaining small controllable degradation ability.