This document discusses using latent Dirichlet allocation (LDA) to classify restaurants from online reviews into topics and recommend restaurants to users. It shows that LDA reduces the feature dimension by over 99% compared to bag-of-words, improves computation efficiency by over 90%, and achieves a classification accuracy between 73-81% depending on the model. LDA identifies 15 topics from reviews including Japanese, Mexican, brunch, bar atmosphere, pizza, Indian, Asian, fast food, sweets, BBQ, compliments, and service. It also evaluates recommendation performance using a normalized distance measure.