In the previous method, we didn't make any assumptions about the topic prior to distribution, and this can result in a limitation, because the algorithm isn't driven by any real-world intuition. LDA, instead, is based on the idea that a topic is characterized by a small ensemble of important words, and normally a document doesn't cover many topics. For this reason, the main assumption is that the prior topic distribution is a symmetric Dirichlet one (α1 = α2 = ... = αK = α). The probability density function is defined as follows:
If the concentration parameter α is less than 1.0, the distribution will be sparse, ...
