Functional effects models: Accounting for preference heterogeneity in panel data with machine learning
Repository to reproduce results from our paper on functional effects where we learn individual-specific intercepts and coefficients from the socio-demographic characteristics to account for inter-individual heterogeneity in panel data.
To run the functional effects models, we recommend to clone this project and install its dependencies in a new environement. We recommend to use the conda python package manager. To clone the project, create a new environement and install the required dependencies, run the following commands:
git clone git@github.com:big-ucl/functional-effects-model.git
cd functional-effects-model
conda create -n func_effects_model python
pip install -r requirements.txt
To run a functional effects, navigate in the src folder and run this command:
python main.py --args_names args_values
The main arguments to specify are:
- functional_intercept: "true"/"false" to run with or without functional intercept;
- functional_params: "true"/"false" to run with or without functional slopes;
- dataset: "SwissMetro"/"LPMC"/"easySHARE", which dataset to run the model on; and
- model: "RUMBoost"/"TasteNet"/"GBDT"/"DNN"/"MixedLogit", which model to use: "RUMBoost" for functional effects with GBDT, "TasteNet" for functional effects with DNNs, "GBDT" blackbox GBDT, "DNN" blackbox DNN and "MixedLogit" for a Random intercept model. Note that not all models are implemented on all datasets.
- optimal_hyperparams: "true"/"false", to run the models with optimal hyperparameters form the hyperparameter search, otherwise will run with default or specified hyperparameters.
See the parser for all arguments that can be specified.
To run the functional effects models with new datasets, you will need to add your dataset in the data folder, and add its path and model structure to the constants file. Finally, the data loading in the train and hyperparameter search files will need to be adapted. Note that in case of monotonic constraints, you will have to specify it in here for GBDT and here for DNNs.
The case studies can be reproduced by running the run_models.py script, with correct arguments and dataset in the data folder. Note that to run models on the easySHARE dataset, you will need to add and pre-process it beforehand.
The hyperparameter search can be reproduced by running the hyperparameter_search.py python script.
The synthetic experiment can be reproduced by running the synthetic_experiment.py python script.
The easySHare dataset pre-processing can be done by running the data_preprocessing.ipynb jupyter notebook.
Finally, models_wrapper.py contains all models wrapped in consistent classes from their source code.