Human-participatory interactive model training

摘要

A method is described for training a predictive model which increases the interpretability and trustworthiness of the model for end-users. The model is trained from data having multitude of features. Each feature is associated with a real value and a time component. Many predicates (atomic elements for training the model) are defined as binary functions operating on the features, and typically time sequences of the features or logical combinations thereof. The predicates can be limited to those functions which have human understandability or encode expert knowledge relative to a predication task of the model. We iteratively train a boosting model with input from an operator or human-in-the-loop. The human-in-the-loop is provided with tools to inspect the model as it is iteratively built and remove one or more of the predicates in the model, e.g. if it does not have indicia of trustworthiness, is not causally related to a prediction of the model, or is not understandable. We repeat the iterative process several times ultimately generate a final boosting model. The final model is then evaluated, e.g., for accuracy, complexity, trustworthiness and post-hoc explainability.