Learning Hidden Markov Models from Pairwise Co-occurrences with Application to Topic Modeling

Authors: Kejun Huang, Xiao Fu, Nicholas Sidiropoulos

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

Reproducibility Variable Result LLM Response
Research Type Experimental 5. Validation on Synthetic Data", "We show the total variation distance between the ground truth probabilities Pr[Xt+1|Xt] and Pr[Yt|Xt] and their estimations c Pr[Xt+1|Xt] and c Pr[Yt|Xt] using various methods. The result is shown in Figure 4. As we can see, the proposed method indeed works best, obtaining almost perfect recovery when sample size is above 108.
Researcher Affiliation Academia 1University of Minnesota, Minneapolis, MN 55455 2Oregon State University, Corvallis, OR 97331 3University of Virginia, Charlottesville, VA 22904.
Pseudocode Yes Algorithm 1 Proposed Algorithm
Open Source Code No The paper mentions 'The in-line implementation of this tailored Newton s method THETAUPDATE and the detailed derivation can be found in the supplementary material.', but it does not provide an unambiguous statement or link for the open-source code of the entire described methodology.
Open Datasets Yes On the Reuters21578 data set obtained at (Mimaroglu, 2007)
Dataset Splits No The paper mentions using the Reuters21578 dataset and synthetic data but does not explicitly provide details about training, validation, or test dataset splits or cross-validation setup.
Hardware Specification No The paper mentions that 'simulations are conducted in MATLAB using the HMM toolbox' but does not provide any specific hardware details such as GPU or CPU models, or memory specifications.
Software Dependencies No The paper mentions 'MATLAB', 'HMM toolbox', and 'Tensorlab', but it does not provide specific version numbers for these software components.
Experiment Setup Yes Fixing N = 100 and K = 20, the transition probabilities are synthetically generated from a random exponential matrix of size K K followed by row-normalization; for the emission probabilities, approximately 50% of the entries in the N K random exponential matrices are set to zero before normalizing the columns... We let the number of HMM realizations go from 106 to 108... initialize M using (Huang et al., 2016a) 2: initialize Θ 1 K(K+1)(I + 11 )