lda2vec
summary
Describes a concrete lda2vec hybrid that adds a learned document vector to skip‑gram word2vec, converts document factors to proportions with softmax, and enforces sparsity so topics are human‑interpretable. Implements a PMI+log SVD count‑based word embedding alternative; Stitch Fix shifts some pipelines to counts. Demonstrates on Hacker News (training ~2 days on a Titan GPU), publishes code and notebooks on GitHub, and warns lda2vec is useful for prototyping but not yet a drop‑in production replacement for mature word2vec/LDA tools.
AI-written summary · about
chapters (16)
- 0:00Intro
- 2:47Training
- 4:05Word vectors
- 5:42Dot product
- 8:54Extra negative samples
- 9:43PMI matrix
- 11:05Popular words
- 13:28Counting
- 17:54Semantics
- 19:49LDI
- 20:58Differences between LDA and LDA2vec
- 33:27LDA2vec Notebook
- 38:00Documentation
- 38:30Should I use lda2vec
- 41:40Perplexity
- 44:15Personalization
