Held-out inference¶

After fitting, transform infers topic proportions (θ) for new, unseen documents while holding the fitted topics (φ) fixed: a held-out test set, or freshly collected texts.

import topica

model = topica.LDA(num_topics=20, seed=42)
model.fit(train_docs, iterations=1000)

theta = model.transform(new_docs, seed=0)     # (len(new_docs), num_topics)
theta.argmax(axis=1)                          # dominant topic per new document

transform accepts a list[list[str]] or a Corpus. Out-of-vocabulary tokens are dropped; a document with no in-vocabulary tokens gets the prior. Rows sum to 1, and results are deterministic for a fixed seed.

Available across the model families¶

Each model uses the same inference it uses at fit time:

Model	Inference
`LDA`, `LabeledLDA`, `SupervisedLDA`	collapsed Gibbs against fixed φ
`HDP`	collapsed Gibbs over the discovered topics
`DMR`	collapsed Gibbs with `α_d = exp(Xγ)`; pass held-out `features`
`CTM`, `STM`	Laplace variational E-step against the logistic-normal prior

For CTM / STM, the variational transform reproduces the model's own training θ to ~1e-3. It is the same inference R's stm runs in fitNewDocuments, not an approximation.

# DMR with held-out covariates:
theta = dmr.transform(new_docs, features=new_X)

# STM variational inference (no Gibbs parameters needed):
theta = stm_model.transform(new_docs)