STOR Colloquium: Tong Wang, University of Iowa
Dr. Tong Wang
Tippie College of Business
University of Iowa
Hybrid Predictive Model: When an Interpretable Model Collaborates with a Black-box Model
Interpretable machine learning has received increasing interest in recent years, especially in domains where humans are involved in the decision-making process. However, the possible loss of the task performance for gaining interpretability is often inevitable, especially for large datasets or complicated tasks. This performance downgrade puts practitioners in a dilemma of choosing between a top-performing black-box model with no explanations and an interpretable model with unsatisfying task performance. In this work, we propose a novel framework for building a Hybrid Predictive Model (HPM) that integrates an interpretable model with any black-box model to introduce interpretability in the decision-making process at no or low cost of the predictive accuracy. The interpretable model substitutes the black-box model on a subset of data where the black-box model is overkill or nearly overkill. We design a principled objective function that considers predictive accuracy, model interpretability, and model transparency, which is the percentage of data processed by the interpretable model. This framework brings together the advantages of the high predictive performance of black-box models and the high interpretability of interpretable models. We instantiate the proposed framework with two types of models, one using decision rules as the interpretable collaborator and one using linear models. For both models, we develop customized training algorithms with theoretically grounded bounds to reduce computation. We test the hybrid predictive models on structured datasets and text data. In these experiments, the interpretable models collaborate with state-of-the-art black-box models including ensemble models and neural networks. We propose to use efficient frontiers to characterize the trade-off between transparency and predictive performance. Results show that hybrid models are able to obtain transparency at no or low cost of predictive performance.
Refreshments will be served at 3:00pm in the 3rd floor lounge of Hanes Hall