MISS: Multiclass Interpretable Scoring Systems

In this work, we present a novel, machine-learning approach for constructingMulticlass Interpretable Scoring Systems (MISS) - a fully data-drivenmethodology for generating single, sparse, and user-friendly scoring systemsfor multiclass classification problems. Scoring systems are commonly utilizedas decision support models in healthcare, criminal justice, and other domainswhere interpretability of predictions and ease of use are crucial. Priormethods for data-driven scoring, such as SLIM (Supersparse Linear IntegerModel), were limited to binary classification tasks and extensions tomulticlass domains were primarily accomplished via one-versus-all-typetechniques. The scores produced by our method can be easily transformed intoclass probabilities via the softmax function. We demonstrate techniques fordimensionality reduction and heuristics that enhance the training efficiencyand decrease the optimality gap, a measure that can certify the optimality ofthe model. Our approach has been extensively evaluated on datasets from variousdomains, and the results indicate that it is competitive with other machinelearning models in terms of classification performance metrics and provideswell-calibrated class probabilities.