MAP All the Things
summary
Embeddings are concrete: numeric vectors plus a distance, built here with UMAP and demonstrated end-to-end on 1.5M BeerAdvocate reviews collapsed to ~104 beer styles and on Yelp, malware and forum sessions. The pipeline uses pandas grouping, scikit‑learn count vectorize / tf‑idf, distances like Jaccard or Hellinger (Fisher information justification), and HDBSCAN/datashader for visualization. Notebooks and a GitHub repo reproduce the experiments; the embeddings reveal geography, style clusters and even three distinct malware strains.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 5:51Fixed-width numeric embedding
- 18:42Categorical embedding with sets
- 29:47Categorical embedding with counts
- 35:54Documents as bag-of-words
- 39:09Word embeddings
- 42:02Generalizing embeddings
- 44:52Installation and resources
- 45:12Q&A
related talks
38mData Mapping for Data Exploration
54mPresentation: Finding Your Meme Twin with Embeddings & Vector Databases
52mA Million Products, One Experience: Using Collaborative and Multimodal Embeddings
25mThe Art of Clustering: The Good, The Bad and The Beautiful
27mHow I hacked UMAP and won at a plotting contest
45mWords in Space: A Visual Exploration of Distance, Documents, and Distributions...
52mModern Problems Require Modern Solutions
30mWords in Space