Skip to content



Repository files navigation

Multinomial Logistic Regression using PyTorch

This repository provides a Multinomial Logistic regression model (a.k.a MNL) for the classification problem of multiple classes.

It's a fork of liaison/Multinomial-Logistic-Regression that includes the possibility to expand the data if the data provided is at the session-chosen-alternative level instead of session-alternative level.

I also includes an implementation for using this in Stata.

The novelity of this model is that it is implemented with the deep learning framework 'Pytorch'. This allows the possibility to handle multiple choice sets and also multiple alternatives. Current implementations only allow for few alternatives.

A typical scenario to apply MNL model is to predict the choice of customer in a collection of alternatives, which is often referred as Customer Choice Modeling. As a clarification of terminology, the alternative might also be referred as 'option', and the collection of alternatives might be called as 'session'.


  • this python module contains the implementation of Multinomial Logistic Regression model that is implemented with Pytorch.

  • this python module provides a number of auxiliary functions in complement with the model. For instance, one can find a train function with the capabililty of early stopping on the predefined threshold of error delta.

  • this is a minimized model that is intended for inference only, with on dependency on the Pytorch framework. Once one obtains a model with the MNL module, one could *export" the trained model to Mint and deploy it in the running time with minimal dependencies (panda + numpy).

  • nmlogit.ado: TBD


One can inspire from the demo notebook on the usage of API. In general, one only needs to provides a dict of parameters for the training, e.g. loss function, optimizer, regularization etc.

As to the input data format, any data source that could be transformed into Python dataframe will do. There is only one requirement on the content of the data, i.e. each data record should contain a session_id attribute that group the records/options into a particular choice session, and a binary choice attribute that indicates whether the option is chosen (value 1) or not (value 0).

session_id choice feature_1 feature_2 feature_n
10001 0 0.331 0.587 ...
10001 0 0.983 0.223 ...
10001 1 0.732 0.945 ...

If the data is not expanded (i.e. all alternatives are included), there is an option to expand the dataset for given groups (choice_groups). In that scenario, three datasets needs to be provided.

  • choice: identifies which alternative session_id chose
session_id alter_id choice choice_groups
10001 5 1 group x
10002 2 1 group y
10003 20 1 group x
  • alternatives: alternative level features
alter_id choice_group feature_1 feature_2
5 group x 0.331 15675
2 group y 0.983 7987
20 group x 0.732 12
  • session: session level features
session_id feature_1 feature_1 feature_2
10001 group x 0.331 15675
10002 group y 0.983 7987
10003 group x 0.732 12
  • session-alternative level features TBD


Multinomial Logistic Regression model







No releases published


No packages published


  • HTML 94.6%
  • Python 3.0%
  • Jupyter Notebook 1.9%
  • Stata 0.5%