Implementation of the PMBP(D,E) model introduced in "Linking Across Data Granularity: Fitting Multivariate Hawkes Processes to Partially Interval-Censored Data" [Calderon, et al. '21].
Current implementation focuses on the D=3 case. Other cases are left for future work.
Create a conda environment from the provided YAML file environment.yml using the command
conda env create
A sample script is provided in sample.py. Running this triggers the end-to-end pipeline on a sample dataset. The pipeline consists of PMBP (1) initialization, (2) fitting, and (3) evaluation.
The PMBP class in code/pmbp.py is a wrapper for defining, fitting, and evaluating the PMBP(D,E) given input data and configuration.
The basic commands to define and fit the model are:
pmbp = PMBP()
pmbp.initialize(data_label, history, E, end_train, end_validation, end_test)
pmbp.fit()
Without any input to the fit() method, hyperparameter tuning on the validation set is skipped and the default hyoerparameters are used. If one wishes to perform hyperparameter tuning, a grid of hyperparameters values must be provided. The fit() command should be replaced with
pmbp.fit(perform_hyperparameter_tuning=True, grid=grid)
where grid is a list of hyperparameter tuples (dimension_weights, nu_regularization_weight, gamma_init). Once the model is fitted, a _fittedmodel.p file is exported in output/, containing the PMBP object with the optimized parameters in pmbp.parameters. Log files are also written in log/ for diagnostics.
Once the model is fitted, the model can be evaluated in terms of RMSE error on the dimension 1 time series predictions using
pmbp.evaluate()
This function implicitly samples the PMBP(D,E) process on the test set, with the number of samples controlled by the pmbp.n parameter (default set to 5). Once the model is evaluated, an _evaluatedmodel.p file is exported in output/, containing the PMBP object with the sampled test set histories in pmbp.sampled_histories. Log files are also written in log/ for diagnostics.
Performance metrics and predictions can be printed out using the following two commands:
pmbp.print_performance_metrics()
pmbp.print_parameters()
Lastly, the fit of the model on the train-val set and predictions on the test set can be plotted using
pmbp.plot_predictions()
This saves a PDF plot to the output/ folder.
An example Python script is provided to run the entire fitting pipeline for a PMBP(3,2) or PMBP(3,3) model on sample data.
To run the example, go to the command line and run the Python script.
python sample.py
on the same directory. Ideally, run the script on a tmux or screen tab so that the entire procedure runs in the background.
Both dataset and code are distributed under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. If you require a different license, please contact us at [email protected] or [email protected].