Struct HierarchicalTopicModel 

pub struct HierarchicalTopicModel { /* private fields */ }

Hierarchical LDA-style topic model with a tree-structured topic hierarchy.

See the module-level documentation for a full overview.

Implementations

impl HierarchicalTopicModel

pub fn new(config: HtmModelConfig) -> Self

Create a new model with the given configuration.

pub fn add_topic_node( &mut self, parent: Option<HtmTopicNodeId>, label: Option<String>, ) -> Result<HtmTopicNodeId, String>

Add a new topic node to the hierarchy.

If parent is None the root node is used as the parent. Returns the new node’s ID. Returns an error string when the parent doesn’t exist or depth/children limits are exceeded.

pub fn add_document(&mut self, tokens: &[&str]) -> HtmDocId

Add a document given a slice of string tokens.

Tokenisation: tokens are lower-cased; empty tokens are skipped. Returns the new document’s HtmDocId.

pub fn run_inference(&mut self, n_iter: usize)

Run collapsed Gibbs sampling for n_iter iterations.

Each iteration sweeps all documents and re-samples each token’s topic assignment from the conditional posterior restricted to nodes on the document’s current root-to-leaf path.

pub fn get_topic(&self, id: HtmTopicNodeId) -> Option<HtmTopic>

Retrieve an aggregated HtmTopic for the given node ID.

Returns None if the node doesn’t exist. The top-N words are computed using Laplace-smoothed probabilities (β smoothing).

pub fn document_topics(&self, doc_id: HtmDocId) -> Vec<(HtmTopicNodeId, f64)>

Return proportional topic distribution along a document’s current path.

Each element is (topic_node_id, proportion) where proportion is the fraction of the document’s tokens assigned to that node. Sums to 1.0 (or 0.0 for empty documents).

pub fn topic_hierarchy( &self, ) -> Vec<(u32, HtmTopicNodeId, Option<HtmTopicNodeId>)>

Return the full topic hierarchy as a flat list of (depth, node_id, parent_id) tuples sorted by depth then by node ID.

pub fn compute_coherence(&self, topic_id: HtmTopicNodeId, top_n: usize) -> f64

Compute PMI-based coherence for a topic node using its top-top_n words.

The coherence score is the average pointwise mutual information for all pairs of top words, computed over the document corpus.

Returns 0.0 if the topic node doesn’t exist, has fewer than 2 words, or no documents.

pub fn prune_empty_topics(&mut self)

Remove leaf topic nodes that have zero total word counts and no documents assigned.

The root node is never removed. Pruning is repeated until no further nodes can be removed.

pub fn model_stats(&self) -> HtmModelStats

Return summary statistics for the current model state.

pub fn n_topic_nodes(&self) -> usize

Return the number of topic nodes (including root).

pub fn topic_node(&self, id: HtmTopicNodeId) -> Option<&HtmTopicNode>

Return a reference to a topic node, if it exists.

pub fn vocab_size(&self) -> usize

Return the current vocabulary size.

pub fn word_at(&self, idx: u32) -> Option<&str>

Return the word at a given vocabulary index.

pub fn word_index(&self, word: &str) -> Option<u32>

Return the vocabulary index of a word (None if not present).

pub fn get_document(&self, doc_id: HtmDocId) -> Option<&HtmDocument>

Retrieve a document by its ID.

pub fn root_id(&self) -> HtmTopicNodeId

Return the ID of the root topic node.

pub fn hash_bytes(data: &[u8]) -> u64

FNV-1a 64-bit hash helper (exposed for testing / external hashing needs).