For my CS229 Machine Learning class project, I experimented with using undirected graphs of label co-occurrences to predict labels for multi-topic text. Overall, I was happy with the results. The approach I came up with performed on-par with a Naive Bayes mixture model but with only the computational cost of using per-topic Naive Bayes models.

See the project page for an overview and the paper for more details. My Python source code for the project is available on GitHub.