Transparency and interpretability are important topics within the machine learning community, particularly concerning the impact on decision makers of these technologies. These aspects are crucial for enabling users to leverage their expertise and decide whether to trust these technologies. In ML methods and particularly in the binary classification task, sparse linear methods have been developed and used as scoring systems. However, beyond their accuracy, we want these models to be sparse, with integer coefficients, and manipulable to allow the incorporation of operational constraints. These are known as Interpretable Machine Learning (IML) models for linear classification, where mathematical integer programming emerges as a tool to generate these IML models. Nonetheless, using integer programming generates models that are difficult to solve for large data sets. Consequently, motivated by the aforementioned issues, in this paper we propose new IML models for linear classification, based on two models: the Supersparse Linear Integer Model (SLIM) and Discrete Level Support Vector Machine (DILSVM). This new approach explores modeling the function through the use of SOS1 constraints. Additionally, a “neural network style” approach is used to approximate the function, where auxiliary variables and constraints are used to emulate an artificial neural network. Computational experiments conducted on a set of selected ML datasets demonstrate that our formulation has comparable accuracy and interpretability, but offers higher scalability.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号