A natural evaluation metric for statistical topic models is the probability of held-out documents given a trained model. While exact computation of this probability is intractable, several estimators for this probability have been used in the topic modeling literature, including the harmonic mean method and empirical likelihood method. In this paper, we demonstrate experimentally that commonly-used methods are unlikely to accurately estimate the probability of held-out documents, and propose two alternative methods that are both accurate and efficient.
%0 Conference Paper
%1 Wallach2009
%A Wallach, Hanna M.
%A Murray, Iain
%A Salakhutdinov, Ruslan
%A Mimno, David
%B Proceedings of the 26th Annual International Conference on Machine Learning
%C New York, NY, USA
%D 2009
%I ACM
%K LDA
%P 1105--1112
%R 10.1145/1553374.1553515
%T Evaluation Methods for Topic Models
%U http://doi.acm.org/10.1145/1553374.1553515
%X A natural evaluation metric for statistical topic models is the probability of held-out documents given a trained model. While exact computation of this probability is intractable, several estimators for this probability have been used in the topic modeling literature, including the harmonic mean method and empirical likelihood method. In this paper, we demonstrate experimentally that commonly-used methods are unlikely to accurately estimate the probability of held-out documents, and propose two alternative methods that are both accurate and efficient.
%@ 978-1-60558-516-1
@inproceedings{Wallach2009,
abstract = {A natural evaluation metric for statistical topic models is the probability of held-out documents given a trained model. While exact computation of this probability is intractable, several estimators for this probability have been used in the topic modeling literature, including the harmonic mean method and empirical likelihood method. In this paper, we demonstrate experimentally that commonly-used methods are unlikely to accurately estimate the probability of held-out documents, and propose two alternative methods that are both accurate and efficient.},
acmid = {1553515},
added-at = {2014-02-14T23:07:56.000+0100},
address = {New York, NY, USA},
author = {Wallach, Hanna M. and Murray, Iain and Salakhutdinov, Ruslan and Mimno, David},
biburl = {https://www.bibsonomy.org/bibtex/25299ad4ceeae704049c4b39dc8a17d63/lopusz_kdd},
booktitle = {Proceedings of the 26th Annual International Conference on Machine Learning},
description = {Evaluation methods for topic models},
doi = {10.1145/1553374.1553515},
interhash = {0d0c6f2f2a090eb9e5e53088ebeb82b9},
intrahash = {5299ad4ceeae704049c4b39dc8a17d63},
isbn = {978-1-60558-516-1},
keywords = {LDA},
location = {Montreal, Quebec, Canada},
numpages = {8},
pages = {1105--1112},
publisher = {ACM},
series = {ICML '09},
timestamp = {2014-02-14T23:07:56.000+0100},
title = {Evaluation Methods for Topic Models},
url = {http://doi.acm.org/10.1145/1553374.1553515},
year = 2009
}