Overview / Usage

The goal of the project is to propose a topic model that is efficiently trainable and can infer the keywords for each document. The most immediate application of the project is to monitor the discussions on social media that do not only directly mention a certain brand or company but also mention the most prominent keywords that are picked up by the model.

Methodology / Approach

We propose a new probabilistic graph model that revises LDA model to incorporate the generation of keywords, which dictates the prior for topic-word matrices and topic weights. Several computational tricks in variational inference or sampling will be used to improve the speed and robustness during training.