Event streams are pervasive and essential in many disciplines. Examples of such data include web search query logs, the firing patterns of neurons, gene expression data, and systems error logs. Understanding the dependencies between types of events in such streams is of general interest to those who aim to understand the behavior of the systems that generate these event streams. In this talk, I describe Graphical Event Models (GEMs), models that can be viewed as a graphical model for continuous time event processes. I also describe a particular family of GEMs, the Piecewise-Constant Conditional Intensity Models (PCIM), that are amenable to both learning and inference. In particular, I describe a closed-form Bayesian approach to model selection and estimation of these models. In addition, I describe an importance sampling algorithm based on a Poisson superposition proposal distribution for forecasting future sequences of events using these models. I will present empirical results using synthetic data, supercomputer event logs, and web search query logs to illustrate that our learning algorithm can efficiently learn nonlinear temporal dependencies. I will also demonstrate that our importance sampling algorithm can effectively forecast future event sequences. Time permitting I will discuss current work on modeling event streams with a large number of structured event types and learning causal relationships.
This talk is based on work with Asela Gunawardana, Puyang Xu, and Ankur Parikh.