Machine learning is a key technology in the area of online services and our work focuses on principled machine learning methods at web scale. Our expertise includes kernel methods, learning theory and graphical models. The drivers behind our research are applications in online services and our work is aimed at understanding those aspects of machine learning that can help make use of the vast amount of data available online.

Research Areas

• Data-centric Modelling Languages: We are developing systems to facilitate rapid prototyping and deployment of graphical models on large scale data by aligning the modelling language with the data representation. Current work is focussing on a probabilistic query language a la SQL which we call PQL.
• Inference in Web-Scale Graphical Models: Inference in large scale graphical models requires distributed computations. We develop algorithms for distributed inference that carry out message passing based inference on distributed factor graphs with an eye on balancing approximation accuracy with computational constraints.
• Machine Learning in Closed Loop Systems: In online services such as search and advertising the predictions of machine learning algorithms are used to drive decisions about what to display. These decisions determine the future composition of the training sample for the machine learning algorithm. We investigate aspects of dynamics, exploration, and incentives related to this closed feedback loop.
• Learning to Rank:Ranking is a fundamental task in information retrieval and skill estimation from game outcomes. We investigate learning algorithms for this task with a focus on graphical models.