Skip to contents

WDL and WIG Model Specs

Source: vignettes/specs.Rmd
specs.Rmd

In this vignette, I will show how to set up the control parameters (hyper-parameters) needed for the WDL and WIG models.

The wdl_specs() is a list of lists, and consists of 5 parts (lists): wdl_control, tokenizer_control, word2vec_control, barycenter_control, optimizer_control.

The wig_specs() is the same as wdl_specs(), with additional wig_control.

wig_control

This is the options only needed for wig_specs(). By default, it is

wig_control = list(
  group_unit = "month",
  svd_method = "docs",
  standardize = TRUE
)
  1. group_unit dictates at which level of time to group the documents, and it will be passed to lubridate::floor_date() as the unit argument. The default option is “month” to obtain monthly time series index, and other options can be specified following the unit argument in lubridate::floor_date().
  2. svd_method can be either “docs” or “topics”. The “doc” method means the Truncated SVD will be applied on the reconstructed documents to get the index directly; whereas “topics” means TSVD will be applied to the topics matrix before the construction of the index. The latter one is the one originally proposed in Xie (2020).
  3. standardize: bool, whether or not to standardize the result index as mean 100 and standard deviation 1. This is default to be true, following Baker et al. (2016), Xie (2020).

wdl_control

This is the options supplied to the WDL modelling, and is used for both wdl_specs() and wig_specs().

  1. num_topics: number of topics for the topic modeling
  2. batch_size: batch size for the training purpose
  3. epochs: epochs (i.e. number of passes) for the training data
  4. shuffle: bool, whether to shuffle the input data randomly
  5. verbose: bool, whether to print out useful diagnostic information

tokenizer_control

Arguments for tokenizers::tokenize_word_stems().

word2vec_control

Arguments for word2vec::word2vec(), but with the following default parameters:

type = "cbow"
dim = 10
min_count = 1

barycenter_control

Identical to barycenter_control in barycenter() function, but with default

with_grad = TRUE

optimizer_control

Parameters to control the optimizer (SGD, Adam, AdamW).

optimizer_control = list(
  optimizer = "adamw",
  lr = .005,
  decay = .01,
  beta1 = .9,
  beta2 = .999,
  eps = 1e-8
)

The default optimizer is AdamW (“adamw”), but you can also choose vanilla SGD (“sgd”) or the vanilla (“adam”). You can also set the learning rate lr in your hyper-parameter search.

The other default parameters should mostly be untouched for most people, unless you know exactly what you are doing. For a reference, you can see Section 7.1 in Xie (2025), and the references within.

See Also

See also vignette("wdl-model"), vignette("wig-model").

References

Baker, S. R., Bloom, N., & Davis, S. J. (2016). Measuring economic policy uncertainty. The Quarterly Journal of Economics, 131(4), 1593–1636. https://doi.org/10.1093/qje/qjw024

Peyré, G., & Cuturi, M. (2019). Computational Optimal Transport: With Applications to Data Science. Foundations and Trends® in Machine Learning, 11(5–6), 355–607. https://doi.org/10.1561/2200000073

Schmitz, M. A., Heitz, M., Bonneel, N., Ngolè, F., Coeurjolly, D., Cuturi, M., Peyré, G., & Starck, J.-L. (2018). Wasserstein dictionary learning: Optimal transport-based unsupervised nonlinear dictionary learning. SIAM Journal on Imaging Sciences, 11(1), 643–678. https://doi.org/10.1137/17M1140431

Xie, F. (2020). Wasserstein index generation model: Automatic generation of time-series index with application to economic policy uncertainty. Economics Letters, 186, 108874. https://doi.org/10.1016/j.econlet.2019.108874

Xie, F. (2025). Deriving the Gradients of Some Popular Optimal Transport Algorithms (No. arXiv:2504.08722). arXiv. https://doi.org/10.48550/arXiv.2504.08722