Annealing Paths for the Evaluation of Topic Models
Abstract
We investigate new methods for evaluating topic models based on annealed importance sampling (AIS), a Monte Carlo integration technique previously applied to topic model evaluation by Wallach et al. (2009). Given two probability distributions, AIS produces an estimate of the ratio of their partition functions by annealing between them. We introduce ratio-AIS paths and show that they have one to two orders of magnitude lower empirical variances in estimates of per-document perplexity ratios than previous methods, with the convex path having the least variance.
Cite
Text
Foulds and Smyth. "Annealing Paths for the Evaluation of Topic Models." Conference on Uncertainty in Artificial Intelligence, 2014.Markdown
[Foulds and Smyth. "Annealing Paths for the Evaluation of Topic Models." Conference on Uncertainty in Artificial Intelligence, 2014.](https://mlanthology.org/uai/2014/foulds2014uai-annealing/)BibTeX
@inproceedings{foulds2014uai-annealing,
title = {{Annealing Paths for the Evaluation of Topic Models}},
author = {Foulds, James R. and Smyth, Padhraic},
booktitle = {Conference on Uncertainty in Artificial Intelligence},
year = {2014},
pages = {220-229},
url = {https://mlanthology.org/uai/2014/foulds2014uai-annealing/}
}