How does CTM differ from LDA?

The standard distinction is the logistic-normal distribution for topic proportions in CTM versus a Dirichlet distribution in LDA.

Does correlated mean two topics always appear together?

No. It describes a learned statistical association, not a deterministic rule.

Are topic names supplied by the model?

The model learns word distributions and latent structure. Human-readable names are interpretations that should be checked against documents.

AI Glossary / Probabilistic machine learning / Intermediate

Correlated topic model

Definition

A correlated topic model, or CTM, is a probabilistic topic model that uses a logistic-normal distribution for a document's topic proportions. This lets the model represent flexible patterns of association between topics, such as two themes tending to become prominent together across documents.

What it means

A document can discuss several themes. A report about electric vehicles may also discuss charging infrastructure, while a report about payroll may involve different combinations. A topic model describes documents as mixtures of latent topics, each represented by a distribution over words.
CTM models relationships among those mixtures through a Gaussian latent vector transformed into proportions. The topics and their relationships are learned statistical structures, not predefined truths about the documents. Human readers still need to interpret the word distributions and inspect representative passages.

How it works

For each document, draw a latent vector from a multivariate normal distribution with a mean and covariance matrix. Transform its coordinates with a softmax so the resulting topic proportions are positive and sum to one.
For each word position, draw a topic assignment from those proportions, then draw a word from that topic's vocabulary distribution. Fitting reverses this generative story: observed words provide evidence about latent assignments, proportions, and model parameters.
The logistic-normal distribution is not conjugate to the categorical word-assignment structure, making inference less straightforward than in some Dirichlet-based models. The original CTM uses approximate variational inference. Its flexibility introduces additional parameters and does not guarantee better predictions for every corpus.

Three practical examples

Research discovery: explore which scientific themes tend to appear together and use those associations to guide document browsing.
Feedback analysis: investigate whether delivery, packaging, and support themes co-occur in customer comments, checking that inferred themes match actual wording.
Internal knowledge organization: map recurring combinations of project topics to help readers find related material, without treating a model's topic labels as an official classification.

Think of it this way

Think of documents as recipes assembled from different ingredients. Some ingredients often appear together, so knowing one is present helps anticipate another. The analogy is limited: topic proportions are latent statistical summaries constrained to sum to one. An observed association is not proof that one theme causes another, and a topic need not correspond to a single clearly named concept.

Worked example

Consider an original three-topic illustration. Transform latent coordinates eta = (log(2), log(2), 0) with softmax. Exponentiating gives weights (2, 2, 1), whose sum is 5, so the topic proportions are (0.4, 0.4, 0.2).
For eta = (0, 0, 0), the proportions are (1 / 3, 1 / 3, 1 / 3). For eta = (−log(2), −log(2), 0), the weights are (0.5, 0.5, 1), giving proportions (0.25, 0.25, 0.5).
Across these illustrative coordinate settings, the first two topic shares rise together while the third falls. They show how shared movement in latent coordinates can express joint topic prominence. They are not fitted observations or an estimated covariance matrix.
If a document has 100 word positions and fixed proportions (0.4, 0.4, 0.2), the expected topic-assignment counts are 40, 40, and 20. Actual sampled counts can differ, and observing the words requires posterior inference rather than simply reading off these expectations.

Illustrative case study

Fictional case: Aspen Library uses a CTM to organize an archive of technical reports. Analysts find two topics whose prominent words suggest batteries and charging infrastructure. They inspect documents where both are prominent before assigning those descriptions.
The team notices that several scanned documents share OCR errors, producing another apparent topic association. It improves preprocessing and compares the resulting model with a simpler baseline. The library presents the map as an exploratory aid, preserving access to source documents. This case illustrates model interpretation and data-quality checks, not a reported discovery.

Common misconceptions and mistakes

Do not say that latent Dirichlet allocation has fully independent topic proportions. Dirichlet proportions sum to one and have a restricted dependence structure, including negative pairwise covariance for distinct components. CTM replaces that restriction with a more flexible logistic-normal construction.
The Gaussian covariance of latent coordinates is not numerically identical to covariance of the transformed proportions. Normalization matters. Topic correlation also does not establish causation, semantic equivalence, or a relationship that must hold outside the analyzed corpus. The usual model is a bag-of-words representation, so it does not explicitly preserve sentence order.

Business use

CTM can help teams navigate large document collections and investigate relationships that a flat category list misses. Its usefulness depends on whether the topics and associations remain interpretable and stable enough for the intended task.
Evaluate held-out prediction, inspect examples, and compare different initializations or topic counts. Record preprocessing choices because vocabulary filtering, duplicated documents, and boilerplate can shape the discovered structure. For an operational taxonomy, combine model suggestions with domain review instead of automatically treating every learned topic as a permanent business category.

Check your understanding

A document has latent coordinates (log(3), 0, 0). What proportions does softmax produce?

Answer: The positive weights are (3, 1, 1), totaling 5, so the proportions are (0.6, 0.2, 0.2). Adding the same constant to all three latent coordinates would leave these proportions unchanged.

Related terms

Softmax
Covariance matrix
Bag of words
Topic model
latent Dirichlet allocation
logistic-normal distribution
variational inference
document-topic distribution

Learn next

Learn how LDA generates words from topic mixtures, then compare Dirichlet and logistic-normal proportions. Calculate several softmax examples and examine how normalization affects dependence. Next, inspect topic-word lists alongside real documents before interpreting a correlation graph.

Softmax

Sources

Blei and Lafferty, “Correlated Topic Models,” original author-hosted paper, model definition and variational inference: https://www.cs.columbia.edu/~blei/papers/BleiLafferty2006.pdf Independent source review: Blei and Lafferty sections 2–2.1 were read for logistic-normal generation, softmax invariance and nonconjugate variational inference.

Previous: Correlated sampling Next: Correlation