Ben Letham

In sequential event prediction, a set of events take place sequentially and at each step we use previous events to predict the subsequent events. In many problems, such as those described here, the sequence of events is non-Markovian. Here we present a formulation for sequential event prediction using supervised ranking, a machine learning tool. We apply our methods to two specific applications. First, an online grocery store recommender system, in which we use the set of items that have been added to the basket to predict the items that will subsequently be added. The sequence of items may be influenced by the recommendations, which leads to a non-convex problem. Second, patient symptom prediction, in which we use a patient's medical history to predict the set of symptoms that he or she will present at their next visit. For both applications, we use real data to show that our supervised ranking methods perform well and are a promising approach for sequential event prediction.