Words in Space: A Visual Exploration of Distance, Documents, and Distributions...
summary
Argues that picking a distance metric matters more than the topic name and shows concrete comparisons: Euclidean, Manhattan, Chebyshev, Minkowski, Mahalanobis, cosine distance and Canberra on three UCI corpora, with yellowbrick visualizers producing t-SNE plots so viewers can see separability change. Claims Mahalanobis rescales dimensions usefully and Canberra wins for short ad tags, while Jaccard and Hamming fail here. Demonstrates code that plugs metrics into scikit-learn, shows plots for ~400 documents, and reports which metrics actually produce distinct clusters.
AI-written summary · about
chapters (23)
- 0:00Introduction
- 1:08Audience background
- 3:26Tabular to spatial
- 9:14Text modeling challenges
- 12:24High dimensionality in text
- 16:56Distance metrics for text
- 19:04Yellowbrick diagnostics
- 20:10t-SNE visualizer demo
- 22:20Cosine distance limits
- 24:49Euclidean on corpora
- 27:20Manhattan on text
- 28:44Chebyshev distance
- 29:54Minkowski for text
- 30:38Mahalanobis normalization
- 32:02Cosine distance limits
- 33:26Canberra and Jaccard
- 34:49Hamming ineffectiveness
- 35:11Model selection iteration
- 35:52Other Yellowbrick visualizers
- 37:45Distance in scikit-learn
- 38:36Human vs automated search
- 39:26Q&A stop words
- 43:33Q&A visualizers contributions
related talks
30mWords in Space
40mTopological Techniques for Unsupervised Learning
38mData Mapping for Data Exploration
40mTools and Tricks from a Pragmatic Data Scientist
01h 28mDocument/Sentence Similarity Solution
49mMetric Representations: Algorithms and Geometry
31mTaking AI/ML to the next level: Text!
38mA Bi-metric Framework for Fast Similarity Search