Abstract
This talk reviews the work of our group on the problem of inverse classification, i.e., the perturbation of a test case to minimize its posterior probability of an undesirable class label, such as the predicted onset of a disease. Starting from exhaustive search on a k-nearest neighbor classifier, we develop mathematical optimization models to handle both smooth classifiers (e.g., SVMs) and general non-smooth classifiers (e.g., random forests). The ideas are further extended to causal modeling framework with a deep learning ANN. Results are applied to a recommendation system for patient risk minimization, incorporating a realistic and customizable cost model.