structural-topic-model 0.3.0

Python implementation of the Structural Topic Model (STM), a port of the R stm package with a scikit-learn style API

stm — Structural Topic Model in Python

日本語 README

A Python port of the R stm package (Roberts, Stewart & Tingley) with a scikit-learn style API.

What is STM?

The Structural Topic Model (STM) is a logistic-normal topic model where document metadata ("prevalence covariates") shifts the prior mean of each document's topic proportions. Without covariates it reduces to the Correlated Topic Model (CTM). Estimation uses semi-collapsed variational EM, the same algorithm as the R stm() function.

Installation

pip install structural-topic-model

Quick Start

import numpy as np
from stm import StructuralTopicModel

# X: (n_docs, n_vocab) word count matrix (dense or scipy.sparse)
# covar: (n_docs, n_covariates) prevalence covariate matrix (intercept added automatically)

model = StructuralTopicModel(n_components=10, init="spectral")
model.fit(X, prevalence=covar)

model.theta_        # topic proportions of training docs (n_docs, K)
model.components_   # topic-word distributions (K, V), each row sums to 1
model.gamma_        # prevalence regression coefficients (1+P, K-1)
model.sigma_        # topic covariance matrix (K-1, K-1)

# Inference on new documents (cf. fitNewDocuments)
theta_new = model.transform(X_new, prevalence=covar_new)

# Top words per topic (cf. labelTopics)
model.top_words(n_words=10)               # by probability
model.top_words(n_words=10, kind="frex")  # FREX: frequency-exclusivity balance

Without covariates, the model is estimated as a CTM:

model = StructuralTopicModel(n_components=10).fit(X)

Content Covariates

Content covariates allow the vocabulary used within topics to vary by document category. Pass one categorical label per document:

model = StructuralTopicModel(n_components=10)
model.fit(X, prevalence=covar, content=party_labels)

model.aspect_components_   # per-level topic-word distributions (n_levels, K, V)
model.kappa_["params"]     # sparse deviations from baseline (lasso-estimated)
model.content_levels_      # sorted unique levels

model.transform(X_new, prevalence=covar_new, content=labels_new)

Estimated via Distributed Poisson regression (equivalent to the R package's default kappa.prior="L1").

Covariate Effects (estimateEffect)

Regress topic proportions on covariates using method of composition, returning coefficients with measurement uncertainty:

from stm import estimate_effect

eff = estimate_effect(model, covar, uncertainty="Global", nsims=25)
tables = eff.summary()      # {topic: structured array with estimate/std_error/t_value/p_value}
tables[0]["estimate"]       # coefficients for topic 0 (first entry is intercept)

Topic Selection (searchK)

from stm import search_k

res = search_k(X, K_values=[5, 10, 15], prevalence=covar,
               model_params={"max_iter": 100})
res["heldout"]   # heldout log-likelihood (document completion)
res["residual"]  # residual dispersion — Taddy (2012), closer to 1 is better
res["semcoh"]    # semantic coherence
res["exclus"]    # exclusivity

Diagnostics

from stm import topic_corr, semantic_coherence, exclusivity, check_residuals

tc = topic_corr(model, cutoff=0.01)    # topic correlation graph (simple method)
tc.posadj                               # positive correlation adjacency matrix
semantic_coherence(model, X, M=10)     # semantic coherence per topic
exclusivity(model, M=10)               # exclusivity per topic
check_residuals(model, X)              # residual dispersion test {dispersion, pvalue, df}

R–Python API Reference

R	Python
stm(docs, vocab, K, prevalence=~x, data=meta)	StructuralTopicModel(n_components=K).fit(X, prevalence=design)
init.type="Spectral" (recommended, default)	init="spectral" (default)
init.type="Random"	init="random"
gamma.prior="Pooled" (default)	implemented (automatic when covariates are present)
sigma.prior	sigma_prior
emtol / max.em.its	tol / max_iter
model= (warm restart)	warm_start=True
content=~group (kappa.prior="L1", default)	fit(X, content=labels)
interactions	content_interactions
fitNewDocuments()	transform()
labelTopics()	top_words()
estimateEffect() / summary()	estimate_effect() / .summary()
searchK()	search_k()
make.heldout() / eval.heldout()	make_heldout() / eval_heldout()
topicCorr(method="simple")	topic_corr()
semanticCoherence() / exclusivity() / checkResiduals()	semantic_coherence() / exclusivity() / check_residuals()
$theta / $beta / $sigma / $mu$gamma	theta_ / components_ / sigma_ / gamma_
$beta$logbeta (content model)	aspect_components_ (probability scale)
$beta$kappa	kappa_

Differences from scikit-learn LDA

• components_ contains normalized probability distributions (sklearn LDA stores unnormalized pseudo-counts).
• Covariates are passed via fit(X, prevalence=...) / transform(X, prevalence=...). R formulas are not supported — encode categorical variables numerically beforehand (e.g. pandas.get_dummies or patsy).
• perplexity(X) is provided, computed from the variational lower bound (exp(-bound / n_tokens)); lower is better.
• warm_start=True enables continued learning across repeated fit() calls (cf. R's model= argument).

Not Implemented

• gamma.prior="L1" (prevalence-side glmnet mode)
• kappa.prior="Jeffreys" (legacy content estimation)
• fixedintercept=FALSE
• LDA (collapsed Gibbs) initialization, ngroups memoized inference, K=0 (Lee & Mimno)
• estimateEffect() with uncertainty="Local", formula interface (pass pre-expanded basis matrices instead)
• topicCorr(method="huge"), selectModel(), permutationTest(), plot functions

Development

uv sync
uv run pytest tests/

References

• Roberts, M., Stewart, B., & Tingley, D. (2019). stm: An R Package for Structural Topic Models. Journal of Statistical Software, 91(2).
• Arora, S. et al. (2013). A Practical Algorithm for Topic Modeling with Provable Guarantees. ICML.