Integrating optimization and learning to prescribe interventions for tuberculosis patients

Abstract:

Creating impact in real-world settings requires agents which navigate the full pipeline from data, to predictive models, to decisions. These components are typically approached separately: a machine learning model is first trained via a measure of predictive accuracy, and then its predictions are used as input into an optimization algorithm which produces a decision. However, the loss function used to train the model may easily be misaligned with the end goal of the agent, which is to make the best decisions possible. We focus on combinatorial optimization problems and introduce a general framework for decision-focused learning, where the machine learning model is directly trained in conjunction with the optimization algorithm to produce high-quality decisions. Technically, our contribution is a means of integrating common classes of discrete optimization problems into deep learning or other predictive models, which are typically trained via gradient descent. The main idea is to use a continuous relaxation of the discrete problem to propagate gradients through the optimization procedure. We instantiate this framework for two broad classes of combinatorial problems: linear programs and submodular maximization. We then provide an application of such techniques to a real problem of societal importance: improving interventions in tuberculosis treatment. Using data on 17,000 Indian patients provided by the NGO Everwell, we consider the problem of predicting which patients are likely to miss doses of medication in the near future and optimizing interventions by health workers to avert such treatment failures. We find the decisionfocused learning improves the number of successful interventions by approximately 15% compared to standard machine learning approaches, demonstrating that aligning the goals of learning and decision making can yield substantial benefits in a socially critical application.