Dealing with data streams often implies a consideration of concept drift. Monitoring for drift can be done via the error signal of a predictive model. This also provides a mechanism for anomaly detection and related tasks. In the case of a time window or multi-dimensional stream, we require a multi-output predictive model, and from this we get a multi-dimensional error distribution. In this talk we explore the use of multi-output methods, particularly those of the probabilistic chain variety, in the context of data-stream learning and in particular with regard to interpretation.
Jesse Read is Assistant Professor in the Data Science and Mining (DaSciM) team in the Computer Science Department at Ecole Polytechnique (Paris region, France). He obtained his PhD from the University of Waikato in New Zealand, and has carried out postdoctoral research in Univerisdad Carlos III in Spain, and Aalto University in Finland. His research interests are in the areas of machine learning, including multi-label and structured-output learning, and learning from data streams and sequential data.