Topic Modeling: LDA, NMF

Best for: Discovering themes in text

How it works

Latent Dirichlet Allocation is a generative model: each document $d$ draws a topic mixture $\theta_d\sim\text{Dir}(\alpha)$, each token draws a topic $z_{dn}\sim\theta_d$, then a word $w_{dn}\sim\text{Dir}(\beta_{z_{dn}})$ from that topic’s distribution. Topics and per-document mixtures are inferred by variational EM or collapsed Gibbs sampling, maximising the marginal likelihood $p(W\mid\alpha,\beta)$. NMF offers a non-probabilistic alternative by factorising the term-document matrix $X\approx WH$ under non-negativity constraints that yield interpretable parts.

Common fields

Research · customer feedback · news analysis