Author of the code: Théo Moins

This repository contains the source code and the results of "RecSeats: A Hybrid Convolutional Neural Network Choice Model for Seat Recommendations at Reserved Seating Venues".

The article propose a hybrid framework combining choice modelling and convolutional networks for the recommendation of locational choice data (RecSeats). We build a discrete choice model based on an individual-choice-level feature space, and then combine this model with a deep learning stage which not only improves accuracy but also makes it robust to the amount of data available per customer.

For reasons of data anonymization, only the files concerning Locational Choice Experiment are available here for reproducibility.

This folder contains all the data used in the project. To date, only one available folder containing locational choice experiment data is available online. Details about those datasets can be found here : https://seatmaplab.com/public/locationalchoicedatasets/

These are csv files where each line represents an available seat in a given configuration, which allows the input to be reconstructed, with a label indicating whether or not it has been chosen. To avoid having to rebuild the input at each run, two subfolders numpy and dataloader allow to save them, respectively for our individual-level and deep models.

Each experiment also has a JSON file containing all these parameters (filenames, room size, padding, etc.).

To reproduce the results, just download the associated dataset on https://seatmaplab.com/public/locationalchoicedatasets/ and copy the two csv files (train/test) on this Data folder.

This folder contains all the saved outputs related to the models:

• save/accuracies: store the metrics results for the models.
• save/best_models: store the pytorch trained models (for CNN/CDNN/Hybrid).

This folder contains all the source code.

• src/deep: contains code for torch tensor preprocessing and deep models training and evaluation
• src/models: contains all the tested models (Individual-levels, CNN, CDNN, Hybrid), and a folder containing JSON files for the hyperparameters of each of them.
• src/preprocessing: contains code for input generation from csv file and feature computation for individual-level part.
• src/visualisation: contains visualization functions.

Each models is turned in one of the jupyter notebook.

Here is a table for each model of the selected hyperparameters. To retrieve the results obtained, simply change the values for the one in the corresponding table, in the parameters file in src/model/parameters

The same hyperparameters are kept for both models combined.