deborahmit / tastenet-mnl Goto Github PK

Dear Yafei Han,

First of all, thank you for providing the replication files of your paper publicly available. I am checking your procedures to apply your model to a paper we are working on with some colleagues, and I have a question regarding how you split the data into training, development, and testing.

In this line of code, as far as I can see, you split the data at the level of choice set (or choice situations) that are available on the data and not at the level of individuals, because len(x) corresponds to the number of choice sets in the entire sample.

Then, when you split the data, you use:

train_ind, dev_test_ind = model_selection.train_test_split(range(N), train_size = 0.7, test_size = 0.3, random_state=8)

which, I think, uses a range() over the number of choice situations.

Hence, as far as I understand, this is not preserving the panel structure of the data since the different answers from a single individual (choice sets) can be assigned to training, development, or testing.

This seems even more problematic given that later you benchmarked your model with a Mixed Logit model in which we should account for the panel structure of the data. In particular, repeated choice situations from the same individual have special treatment as the product of Logit formulas for the entire sequences of choices of the same individual (see equations (6.2) and (6.3) of Kenneth Train's book Chapter 6)

Sadly, I couldn't find the code that replicates the Multinomial Logit (MNL) or Mixed Logit (MIXL) models reported in your article in this repository, so, unfortunately, I cannot reproduce your results.

Am I interpreting your programs correctly, or am I missing something here?

Thank you for your time :)!
Best regards,
Álvaro