We are designing machine learning algorithms whose results are meaningful and intuitive to human experts, yet have predictive accuracy on par with the state-of-the-art machine learning algorithms. These models are parsimonious (sparse), as humans can handle only a handful of cognitive entities at one time, and are in the forms of a decision list or linear scoring system.
These algorithms have been used to design predictive models for:
i) Medical condition prediction: predicting the next medical condition before the patient experiences it.
A Hierarchical Model for Association Rule Mining of Sequential Events: An Approach to Automated Medical Symptom Prediction. Annals of Applied Statistics.
ii) Stroke prediction in patients with atrial fibrillation: An improvement over the CHADS_2 score
An Interpretable Stroke Prediction Model using Rules and Bayesian Analysis. Proceedings of AAAI Late Breaking Track.
iii) Predicting violent crime in youth raised in out-of-home care.
Supersparse Linear Integer Models for Predictive Scoring Systems. Proceedings of AAAI Late Breaking Track.