Topic Modeling: A Provable Spectral Method
summary
Provides a provable pipeline that recovers K topic vectors by combining threshold then SVD with k‑means and a catchword extraction step, under four explicit assumptions: catchwords and dominant topics (dominant weight α ≈ 0.4), a small fraction of nearly-pure documents, and a unimodal word‑frequency curve. Thresholding sharply reduces the signal strength needed versus raw SVD, the method identifies ~20 catchwords per topic with total frequency ≈0.05, and semi‑synthetic experiments beat prior anchor‑word algorithms in L1 reconstruction.
AI-written summary · about
chapters (3)
related talks
13mThinking of Topic Modeling as Search
29mThinking of Topic Modeling as Search
01h 01mWorkshop: Intro to topic modeling using Python
38mDimension Reduction and Extracting Topics - A Gentle Introduction
56mTensor Decompositions for Learning Latent Variable Models I
31mTopic Modeling Approach for Rank Aggregation
45mDecoding Topics: A Comparative Analysis of Python’s Leading Topic Modeling Libraries Using Climate C
34mTopic modeling with BERTopic