Topic Models from Reviews_​BERTopic

Dimensionality Reduction: Employs UMAP with cosine distance metric to project high-dimensional text embeddings into a 2D space.

Clustering (HDBSCAN): Groups similar documents into topics. Note that Euclidean distance is used here instead of Cosine; this is because HDBSCAN’s high-performance algorithms require a Minkowski metric. Direct Cosine distance is not mathematically compatible with these acceleration structures, which are necessary to efficiently calculate cluster density and document probabilities.

Keyword Optimization: Applies Maximal Marginal Relevance (MMR) with a diversity score of 0.3 to extract the most representative and non-redundant keywords for each topic.

GroupBy (Summary words): Concatenates the top 10 words per topic to provide a readable summary of the identified themes.

Bubble Chart UMAP: Generates an interactive visualization where each colour represents a topic and the size determines the number of terms of the document; its position in the 2D space reflects semantic similarity to other topics. If the Topic Extractor is re-run and the number of topics changes, the Color Manager will lose its configuration for the new Topic IDs. This will cause the Bubble Chart to fail.

To fix: Open the Color Manager node, click 'OK' to refresh the color mapping, and re-execute the downstream nodes.

Hierarchical View: Produces a dendrogram to visualize the relationships and hierarchy between topics, showing how specific clusters (like "Front Desk") relate to broader categories.

Topic View / Comparison: Allows for the comparison of topic distributions across different segments, such as individual hotels. NOTE: change topics name in the node before running it