Regarding how the user click data is used in the Google News paper, I went back and looked at a referenced paper, "Latent Semantic Models for Collaborative Filtering." The E-step is only calculated for (user, story) pairs that actually occurred. So the q* weights calculated in e-step are non-zero only for existing pairs, and these weights can be used to calculate conditional probabilities for stories given clusters and clusters given users.

This paper has more details. I highly recommend section 3.2 if you need to understand this algorithm.

Thomas Hofmann, Latent semantic models for collaborative filtering, ACM Transactions on Information Systems (TOIS), v.22 n.1, p.89-115, January 2004
http://delivery.acm.org/10.1145/970000/963...85&CFTOKEN=22509318