Rethinking LDA: Moment Matching for Discrete ICA

Authors: Anastasia Podosinnikova, Francis Bach, Simon Lacoste-Julien

NeurIPS 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, (a) we compare experimentally the GP/DICA cumulants with the LDA moments and (b) the spectral algorithm [3], the tensor power method [4] (TPM), the joint diagonalization (JD) algorithm from Algorithm 1, and variational inference for LDA [1]. Real data: the associated press (AP) dataset, from D. Blei s web page,4 with N = 2, 243 documents and M = 10, 473 vocabulary words and the average document length b L = 194; the NIPS papers dataset5 [28] of 2, 483 NIPS papers and 14, 036 words, and b L = 1, 321; the KOS dataset,6 from the UCI Repository, with 3, 430 documents and 6, 906 words, and b L = 136.
Researcher Affiliation Academia Anastasia Podosinnikova Francis Bach Simon Lacoste-Julien INRIA Ecole normale sup erieure Paris
Pseudocode Yes Algorithm 1 Joint diagonalization (JD) algorithm for GP/DICA cumulants (or LDA moments)
Open Source Code Yes We use our Matlab implementation of the GP/DICA cumulants, the LDA moments, and the diagonalization algorithms. The datasets and the code for reproducing our experiments are available online.7 Footnote 7: https://github.com/anastasia-podosinnikova/dica
Open Datasets Yes Real data: the associated press (AP) dataset, from D. Blei s web page,4 with N = 2, 243 documents and M = 10, 473 vocabulary words and the average document length b L = 194; the NIPS papers dataset5 [28] of 2, 483 NIPS papers and 14, 036 words, and b L = 1, 321; the KOS dataset,6 from the UCI Repository, with 3, 430 documents and 6, 906 words, and b L = 136. Footnote 4: http://www.cs.columbia.edu/ blei/lda-c Footnote 5: http://ai.stanford.edu/ gal/data Footnote 6: https://archive.ics.uci.edu/ml/datasets/Bag+of+Words
Dataset Splits No The paper mentions 'held out documents' for evaluation but does not specify explicit training, validation, and test dataset splits with percentages or sample counts. It describes how semi-synthetic data is generated and evaluated, and how real data is evaluated using 'held out documents', but without precise split information.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models, memory specifications, or cloud computing resources used for running the experiments. It only mentions a 'Matlab implementation'.
Software Dependencies No The paper mentions using a 'Matlab implementation' but does not specify the version of Matlab or any other software dependencies with their respective version numbers.
Experiment Setup Yes All the sampling models have the parameter c which is set to c = c0 c/ c 1, where c is the learned c from the real dataset with variational LDA, and c0 is a parameter that we can vary. The GP data are sampled from the gamma-Poisson model (3) with b = c0/b L so that the expected document length is b L. For the AP data, K {10, 50} topics are learned and, for the NIPS data, K {10, 90} topics are learned.