The Inverse Regression Topic Model (Supplement)

This supplement includes brief elaborations on the main paper that may be of interest to some readers. In Section 1, we explain the minorization procedure underlying MAP inference. In Section 2, we lay out the details of our stochastic subgradient approximation procedure for online MAP inference. In Section 3, we lay out a useful interpretation of MAP prediction. In Section 4, we summarize the results of experiments with the two other prediction methods for the IRTM (MAP and sufficient reduction based) mentioned in the main paper. In Section 5, we discuss exploration of topic variation via the topic families βk(y) themselves and explain why we found it inadequate.