UC Irvine

Software applications that digitally collect and store individual activity data are increasingly prevalent in our daily lives. With the widespread availability of such data, there is a growing demand for predictive models that can provide a personalized experience, i.e., adapt the application to an individual's behavioral patterns. In recent years, such models focused on problems in which the predictive task is to recommend new items that the user has not yet interacted with. Yet, there is also a large number of applications in which users interact with both old (i.e., that a user has interacted with in the past) and new items, such as location tracking services, online music streaming platforms, and e-commerce websites. In this dissertation, we investigate models that can predict an individual's interactions with both previously interacted (``old'') and novel (``new'') items. We develop novel probabilistic models that combine memory-based components with components that can use contextual information to make predictions that generalize to new items. We show that this approach not only models detailed aspect of both repetitive and novelty-seeking behavior but also infers the balance between the two from data.