artefactory / choice-learn Goto Github PK

Important to fix

contexts_features -> features_fixed_by_choice
contexts_items_features -> items_features_by_choice
contexts_items_availabilities -> available_items_by_choice
fixed_items_features -> features_fixed_by_item

Change default behaviour when instantiated with DF

• Dataset instantiation warning if no names > conditional MNL not possible
• check on conditionalMNL instantiate with features names not compatible

The example in the README file does not run as expected.

transport_df is not defined.

I also tried to load the SwissMetro dataset:

dataset = load_swissmetro(add_items_one_hot=False, as_frame=False, return_desc=False, preprocessing=None)

but I get the following error:

ValueError: Feature income has an attributed coefficient
                            but is not in dataset

[Thread] Improve / increase catched errors to help for debug

Endpoints using a dataset or a batch ?

ChoiceDataset:

• .fit
• .evaluate
• .compute_utility_from_dataset ?
• .compute_probas_from_dataset ?

batch of features:

• .compute_utility
• .predict_probas

https://docs.google.com/presentation/d/1Og-g6kmp_50uIlmnd00qv5kee2y-0rI93walDfrxNsw/edit#slide=id.g2c2e17c10ee_0_133

Comparison with latent class modelisation here.

Results slightly different:

Choice-Learn's:

gmnl's:

and not a specific object as currently

one_hot = FeaturesStorage(one_hot=True, **kwargs)

• Clean functions attributes
• Add datasets descriptions, citations, links to papers, etc...
• Make sure new datasets are listed

•  Keep track of items_id in ChoiceDataset
• Used in parametrisation for:
  • Conditional Logit
  • Nested Logit

• Base Class utilitary
• A simple example on dataset ?

🚀 Feature Request

Suggestions regarding the Experiments and examples section of the paper submitted to JOSS :

• The (Ben Hamner, Friedman, 2013) citation seems to be very Kaggle specific (for instance I didn't find a correspondance on Google Scholar), so could you add on the reference to add the link to the data. Kaggle website suggests:

Adam, Ben Hamner, Dan Friedman, SSA_Expedia. (2013). Personalize Expedia Hotel Searches - ICDM 2013. Kaggle. https://kaggle.com/competitions/expedia-personalized-sort

In general having links for each reference is a good practice in such journal.

• Is there a notebook somewhere on this repo that shows the code used to generate this benchmark? If so, I would add the link to it in a footnote of the paper or at least a comment that it is in the repository. There is a lot of content here (congrat's!), but it would really help the reader to find its way.
• The "customize choice models" is indeed a nice feature, however I'm not sure it is necessary to show code regarding this in the paper. It is probably sufficient to just mention it and link to the documentation of the package.

📎 Additional context

This comment are here to ensure data is accessible and reproducibility.

Connected to openjournals/joss-reviews#6899

Not very critical but the authors and special thanks sections in the README are empty.

🚀 Feature Request

Add Loss update in TQDM progress bar in [model].fit()

🔈 Motivation

In order to follow NLL value over epochs during fit

🛰 Alternatives

None

📎 Additional context

None

• from_single_wide df: ADD prefixes possibilitiy
• remove automatic "_" > Finalement make it an argument of the function

https://docs.google.com/spreadsheets/d/1INO-mYIFt6bs4U_1cWEy84gdvqFoW-4uL4ZD2N-smxk/edit#gid=0

• Naming question: item <-> alternative / alt ?

🚀 Feature Request

Suggestions regarding JOSS paper :

• p.3 L.45: I think it should be made clearer that the FeaturesStorage structure comes from this packages and not numpy
• p.3 L.51: L-BFGS is not a second-order algorithm as it doesn't use any second-order information on the problem. It is true though that it uses first-order information to build an approximation of the second-order.
• The paper is quite long for JOSS format, and it would be great (also optional IMO), one suggestion would be to remove the Streamlined signature section. In general, the paper should focus on the main points.

📎 Additional context

Connected to openjournals/joss-reviews#6899

• better handling of dependencies
• PyPI formatting
• poetry
• item/item-full signature to be kepts ?
• Possibility to load SwissMetro on Long format for example
• conditional Logit example main on SwissMetro
• ADD: Regularization ?
• ENH: naming of weights in .report

Retrieving values in a FeaturesStorage should increase matrix / vector dimension

• If all values are either float or int
• If indexes are well ordered

Then automatically transform storage into a np.ndarray

Should help salvage some memory usage