Admixture of Poisson MRFs: A Topic Model with Word Dependencies

Authors: David Inouye, Pradeep Ravikumar, Inderjit Dhillon

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

Reproducibility Variable Result LLM Response
Research Type Experimental We present a tractable method for estimating the parameters of an APM based on the pseudo log-likelihood and demonstrate the benefits of APM over previous models by preliminary qualitative and quantitative experiments. and Finally, we provide qualitative as well as quantitative evidence for the benefits of APM by training the APM model on both a subset of the Grolier encyclopedia and the CMU 20 Newsgroup dataset.
Researcher Affiliation Academia David I. Inouye DINOUYE@CS.UTEXAS.EDU Pradeep Ravikumar PRADEEPR@CS.UTEXAS.EDU Inderjit S. Dhillon INDERJIT@CS.UTEXAS.EDU Dept. of Computer Science, University of Texas, Austin, TX 78712, USA
Pseudocode No The paper describes the optimization algorithms used (e.g., 'a proximal optimization algorithm'), but does not include any structured pseudocode or algorithm blocks (e.g., Algorithm 1, or numbered steps formatted as code).
Open Source Code No The paper mentions using 'the MATLAB Topic Modeling Toolbox' with a URL (www.psiexp.ss.uci.edu/research/programs_ data/toolbox.htm), but this is a third-party tool used for comparison (LDA), not the open-source release of the APM model developed in this paper.
Open Datasets Yes training the APM model on both a subset of the Grolier encyclopedia and the CMU 20 Newsgroup dataset. and www.cs.nyu.edu/ roweis/data.html (referring to Grolier). The APM model was applied to the CMU 20 Newsgroup dataset evaluated with the two metrics explained next. and To compute the probabilities for the PMI metric, a recent dump of Wikipedia was used with a sliding window of 20 words.
Dataset Splits No The paper refers to 'training data' in the context of the UMass coherence metric ('computing co-occurrence statistics from the training data'), but does not provide specific details on how the dataset (e.g., CMU 20 Newsgroup) was split into training, validation, or test sets (e.g., percentages, sample counts, or explicit standard splits).
Hardware Specification No The paper states that 'the models trained for this paper required many iterations to converge (> 5000)' implying computational resources, but it does not provide any specific hardware details such as GPU models, CPU types, or memory specifications used for the experiments.
Software Dependencies No The paper mentions using the 'MATLAB Topic Modeling Toolbox' (www.psiexp.ss.uci.edu/research/programs_ data/toolbox.htm) and the 'graph visualization program Gephi' (www.gephi.org), but it does not specify version numbers for these or any other software dependencies.
Experiment Setup Yes For both LDA and APM, the hyperparameters α and β were set to α = 200/p and β = 50/k respectively as suggested by the documentation of the toolbox. For APM, the parameter λ was set near 10 7, which was chosen so that there would be some edges in the initial iterations however, as discussed in the following section, the final converged APM solution did not have any edges.