Enabling inference on held-out data in the author-topic model

[olavurmortensen]

At the moment, it is not possible to make inference on held-out data in the AuthorTopicModel, and as a result it is not possible to evaluate model fit (bound) on new data either.

In LDA, we infer on held-out documents by calling gammad, _ = self.inference([doc]), learning the document's topic distribution gamma (local parameter), without updating the model (sstats, global parameter), by (implicitly) setting collect_sstats=False. This allows us to compute the bound on those documents.

It is not 100% clear what inference on held-out data means in the author-topic model. I suggest this definition, analogous to LDA: computing the topic distribution gamma for a new author with documents docs without updating the model (i.e. no change to sstats). Then computing the bound on these held-out documents and authors.

---

The inference algorithm used in the AuthorTopicModel, as well as the model class, is very similar to LdaModel. Therefore, anyone with experience with LdaModel should find it relatively easy to jump into the AuthorTopicModel.

A report detailing the theory as well as the implementation is available here. This is my masters thesis.

I will offer my assistance with anything related to the AuthorTopicModel to anyone willing to take on this issue. For example on a GitHub PR and on Gitter.

[nickkimer]

has this been implemented anywhere?

[menshikh-iv]

@nickkimer now possible to infer vector for unseen author #1766 (in develop now), I think this addition resolve current issue, you can install current develop branch or wait next 3.5.0 release .