Leveraging uncertainty estimates to improve classifier performance

Binary classification involves predicting the label of an instance based on whether the model score for the positive class exceeds a threshold chosen as per application needs (e.g., maximizing recall at a precision bound). However, model scores are often not aligned with the true conditional probability of the positive class. This is especially true when the training involves differential sampling across classes or there is distributional drift between train and test settings. In this paper, we provide theoretical analysis and empirical evidence of the dependence of model score estimation bias on both uncertainty and score. Further, we formulate the decision boundary selection in terms of both model score and uncertainty, prove that it is NP-hard, and present algorithms based on dynamic programming and isotonic regression. Evaluation of the proposed algorithms on three real-world datasets yield 25%-40% gain in recall at high precision bounds over the traditional approach of using model score alone, highlighting the benefits of leveraging uncertainty.