A Framework for Constrained Bayesian Interference with Applications
In many practical applications, it is not easy to capture domain-specific knowledge or constraints with a suitable “prior," which makes using Bayesian methods challenging. WNCG Prof. Joydeep Ghosh and his students Sanmi Koyejo, Rajiv Khanna, Shalmali Joshi and Cheng Lee are developing a broad framework that poses constrained Bayesian inference as optimization problems. In addition to including standard Bayesian inference as a special case, this framework can express and solve for a variety of other constraints, such as a low rank interaction matrix, and how multiple clusterings of a common set of objects based on multiple views should align themselves. For some of these problems, the WNCG research team developed efficient optimization procedures and provided both additional flexibility and computation speed. A version that applies to transposable data readily outperforms state of the art solutions on recommender systems with side information, and on the difficult problem of identifying candidate disease-genes from extremely sparse known-association data.
This research was funded by the National Science Foundation.
Paper 1: A Constrained Matrix-Variate Gaussian Process for Transposable Data