Struct TopicModeler 

pub struct TopicModeler {
    pub config: TopicModelConfig,
    pub vocabulary: Vec<String>,
    pub word_index: HashMap<String, usize>,
}

LDA-based topic modeler using collapsed Gibbs sampling.

Build the vocabulary from a corpus, run Gibbs sampling, and produce per-topic word distributions and per-document topic distributions.

Fields

config: TopicModelConfig

Model configuration.

vocabulary: Vec<String>

Sorted vocabulary list (index → word).

word_index: HashMap<String, usize>

Word → vocabulary index lookup.

Implementations

impl TopicModeler

pub fn new(config: TopicModelConfig) -> Self

Creates a new TopicModeler with the given configuration.

pub fn fit( &mut self, documents: &[ModelDocument], ) -> Result<TopicModelResult, TopicModelError>

Fit the topic model to documents using collapsed Gibbs sampling.

Returns a TopicModelResult with topic word distributions, per-document topic distributions, perplexity, and metadata.

pub fn transform( &mut self, documents: &[ModelDocument], result: &TopicModelResult, ) -> Result<Vec<DocumentTopics>, TopicModelError>

Infer topic distributions for documents given an already fitted model.

Runs a short round of Gibbs sampling on the new documents, keeping topic_word_counts frozen (using only the fitted model’s topic-word distributions as priors).

pub fn coherence_score( topic: &LdaTopic, corpus_word_counts: &HashMap<String, u32>, ) -> f64

Compute a simplified PMI-based coherence score for a topic.

For each pair of top words (w_i, w_j), adds log((co_count + 1) / (count_j + 1)) where co_count is approximated as min(count_i, count_j) (simplified co-occurrence). Returns the mean over all pairs; returns 0.0 for single-word topics.

pub fn most_similar_topics( topic_a: usize, topic_b: usize, result: &TopicModelResult, ) -> f64

Compute cosine similarity between the word-probability distributions of two topics.

Both topics must exist in result.topics. Returns 0.0 if either topic id is out of bounds or the distributions are zero.

pub fn top_documents_for_topic( topic_id: usize, n: usize, result: &TopicModelResult, ) -> Vec<&DocumentTopics>

Return references to the top n documents ranked by topic_distribution[topic_id].

Returns an empty slice if topic_id is out of range.

pub fn stats(result: &TopicModelResult) -> TopicModelerStats

Compute aggregate statistics from a fitted model result.