I have been obsessed with topic modeling for more than a year now. It is an NLP technique that actually has important applications in social science research. This is a big feast for computational methods, and for a social scientist like me.
When I first learned about topic modeling, I spent a lot of time trying to learn how to make it work. Besides, I also wanted to know how I can use this cool technique in my research concerning race/ethnicity, immigration, etc. I was not concerned at all about the mathematical underpinning of the method.
This all went well until I learned Bayesian statistics this summer. Now I see Bayesian everywhere. I finally understand that under the hood, Bayesian inference makes topic modeling such as LDA, or STM possible. This eureka moment really elated me.
It turns out that I have been using language of tuning hyper parameters without really understanding what goes underneath the entire process. Now with some basic Bayesian statistics, things start to make more sense to me, and I feel more confident in explaining how topic modeling works.