Unveiling Clustering in BERTopic Topic Modeling
summary
BERTopic is shown end-to-end on Amazon Alexa reviews from Kaggle using sentence-transformer embeddings, UMAP for reduction, clustering with HDBSCAN, and topic representation with C-TF-IDF; a demo extracts a top topic with 284 occurrences and a “gift” topic (topic 4) about purchases as presents. HDBSCAN builds a minimum spanning tree, prunes by stability score, yields outliers as noise, and benchmarks faster than DBSCAN (K-means is fastest but less accurate); a CUML GPU option addresses scale.
AI-written summary · about
chapters (23)
- 0:00intro
- 0:22preface
- 0:36who are we?
- 1:42agenda
- 2:30topic modeling use case
- 4:02why bertopic?
- 6:47bertopic end-to-end flow
- 7:36clustering
- 8:33dataset description
- 8:56demo
- 13:07what is hdbscan?
- 13:37to understand hdbscan we need to know dbscan
- 15:39what if there was no fixed radius?
- 15:54k-nn algorithm to define radius
- 18:00minimum spanning tree finds density and hierachy
- 19:49density based spatial clustering
- 20:09stability score "λ"
- 22:03final clusters
- 22:32hdbscan steps
- 23:05hdbscan - performance comparison
- 23:58hdbscan - strenghts and weaknesses
- 24:47conclusion and future scope
- 26:28references & ressources
related talks
34mTopic modeling with BERTopic
45mDecoding Topics: A Comparative Analysis of Python’s Leading Topic Modeling Libraries Using Climate C
57mDiscover semantic topics in a collection of documents
13mThinking of Topic Modeling as Search
25mThe Art of Clustering: The Good, The Bad and The Beautiful
01h 19mGenerating Customers Insights with Topic Modelling and HuggingFace SetFit Method
01h 01mWorkshop: Intro to topic modeling using Python
35mLeveraging LLM for text augmentation in Topic Modeling