Thanks for the answer. I was trying to implement [1, 2] which they used PG for modeling the likelihood of stick-breaking processes. Maybe it’s not beneficial in this case too?
[1] Poddar, Lahari, Wynne Hsu, and Mong Li Lee. “Quantifying aspect bias in ordinal ratings using a bayesian approach.” arXiv preprint arXiv:1705.05098 (2017).
[2] Chen, Jianfei, et al. “Scalable inference for logistic-normal topic models.” Advances in Neural Information Processing Systems . 2013.