Closed-form Marginal Likelihood in Gamma-Poisson Matrix Factorization

Authors: Louis Filstroff, Alberto Lumbreras, Cédric Févotte

ICML 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 6. Experimental results We now compare the three MCEM algorithms proposed for MMLE in the Ga P model, first using synthetic toy datasets, then real-world data.
Researcher Affiliation Academia 1IRIT, Universit e de Toulouse, CNRS, France.
Pseudocode No The paper describes the steps of the MCEM algorithm but does not present them in a clearly labeled 'Pseudocode' or 'Algorithm' block.
Open Source Code Yes Python implementations of the three algorithms are available from the first author website.
Open Datasets Yes Finally, we consider the NIPS dataset which contains word counts from a collection of articles published at the NIPS conference.1 The number of articles is N = 1, 500 and the number of unique terms (appearing at least 10 times after tokenization and removing stop-words) is F = 12, 419. 1https://archive.ics.uci.edu/ml/datasets/bag+of+words
Dataset Splits No The paper does not specify explicit train/validation/test dataset splits. For synthetic data, it states N=100 samples, and for NIPS, N=1,500 articles and F=12,419 terms without mentioning how these were partitioned for training, validation, or testing.
Hardware Specification No The paper mentions 'CPUtime' but does not provide any specific hardware details such as GPU models, CPU models, or memory specifications used for running the experiments.
Software Dependencies No The paper mentions 'Python implementations' but does not specify any version numbers for Python or any specific software libraries or dependencies used.
Experiment Setup Yes We proceed to estimate the dictionary W using hyperparemeters K = K + 1 = 3, αk = βk = 1 with MCEM-C, MCEM-H and MCEM-CH. The algorithms are run for 500 iterations. 300 Gibbs samples are generated at each iteration, with the first 150 samples being discarded for burn-in (this proves to be enough in practice), leading to J = 150. The Gibbs sampler at EM iteration i + 1 is initialized with the last sample obtained at EM iteration i (warm restart). The algorithms are initialized from the same deterministic starting point given by Wfk = 1, as suggested by Equation (49).