Have you ever wondered what it would be like to predict the future? Professor George Chen isn’t psychic, but he can provide analysis modelling that predicts the timing of critical events – such as when a customer might end a subscription service or when a convicted criminal might re-offend.
CMU Australia’s Speaker Series welcomed Professor Chen on Friday as he gave a talk to students, alumni and the public titled Toward Explainable Deep Survival Analysis Models with Guarantees. Professor Chen explained the basics behind survival analysis modelling, which predicts how much time will elapse until a critical event occurs – known as time-to-event outcome.
Originally, this type of modelling was used in healthcare applications to determine the length of time until death in patients. “Looking at time until death – this is classically what people focused on when they were first looking at time-to-event outcomes,” Professor Chen said, “which is why this sub-field of study where modelling time-to-event outcomes is referred to as survival analysis.”
Survival analysis has many applications in healthcare, including time until disease relapse, readmission to hospital or device failure, but Professor Chen was quick to emphasise that the field is not all doom and gloom. Survival analysis can also be used to predict time until a criminal reoffends, a customer ends a subscription service, or even when a PhD student will graduate.
While recent advances in machine learning have focused on collecting large amounts of data to achieve a high level of prediction accuracy, there has been little focus on whether the models are easy for experts to interpret. As an alternative, Professor Chen told students, neural net models could achieve accuracy comparable to existing state-of-the-art (SOTA) models while being explainable and coming with statistical accuracy guarantees.
He showed the results of experiments on healthcare datasets that predicted time until death for patients with various diseases and, on a lighter note, a music subscription dataset which predicted when a customer would end their subscription.