Variational Inference for Sequential Distance Dependent Chinese Restaurant Process

Authors: Sergey Bartunov, Dmitry Vetrov

ICML 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5. EXPERIMENTS In this section we empirically compare our variational algorithms with a set of baselines. We start from comparison with Gibbs sampler as it is currently the only available inference algorithm for dd CRP. Next we compare our variational inference for CRP-equivalent mixture model and variational inference for Dirichlet Process. Then we evaluate distance-dependent mixture model against CRP mixture. We release our software implementation used for the experiments.
Researcher Affiliation Academia Sergey Bartunov SBOS@SBOS.IN Dorodnicyn Computing Centre of the Russian Academy of Sciences, Moscow RUSSIA Dmitry P. Vetrov VETROVD@YANDEX.RU Moscow State University, Moscow RUSSIA Higher School of Economics, Moscow RUSSIA
Pseudocode Yes Algorithm 1 Variational inference for seqdd CRP
Open Source Code Yes We release our software implementation used for the experiments. http://github.com/sbos/seqddcrp.jl
Open Datasets No Our dataset consisted from 2246 news articles from the Associated Press, we performed word stemming, but did not remove stop-words. (No citation, link, or specific source for this AP dataset is provided.)
Dataset Splits No The paper only mentions a train/test split ("Each dataset was split into 200 train data points and 200 test data points.") but does not explicitly describe a validation split.
Hardware Specification No The paper does not provide any specific details about the hardware used for the experiments.
Software Dependencies No The paper mentions its implementation is in Julia (implied by the GitHub link `seqddcrp.jl`), but it does not specify the Julia version or any other software dependencies with version numbers.
Experiment Setup Yes We set parameter = 0.1 for both models to encourage small number of clusters and provided weak informative priors for covariance matrix to slightly suggest it s spherical form. Truncation level for DP VB was set to 50.