Bayesian Mixture Hawkes

Code accompanying the paper "# What Drives Online Popularity: Author, Content or Sharers? Estimating Spread Dynamics with Bayesian Mixture Hawkes" [Calderon P, Rizoiu MA].

Description

This repo contains data and the Stan and Python implementation of the Bayesian Mixture Hawkes (BMH) model in the paper. The repo is divided into two parts: model_inference/ and headline_style_profiling/.

Model Inference

model_inference/ contains the Stan and Python implementation of the BMH-P (popularity) and BMH-K (kernel) models. There are three subdirectories:

• bmhp/ contains the BMH-P implementation and RNIX/CNIX cascade sizes.
• bmhk_rnix/ contains the BMH-K implementation and RNIX inter-arrival distribution data.
• bmhk_cnix/ contains the BMH-K implementation and CNIX inter-arrival distribution data.

Each model_inference/bmh*/ folder consists of the following folders:

• data/ contains the pertinent data files for fitting
• stan_files/ contains the Stan code for the BMH model variants
• log/ holds log files upon running
• output/ holds (fitting) output files from running run_bayesianfit.sh
• preds holds (prediction) output files from running run_bayesianpred.sh
• tracker/ stores tracker files when running run_bayesianfit.sh on PBS
• tracker_pred/stores tracker files when running run_bayesianpred.sh on PBS
• metrics_mean/ and preds_mean/contain holdout likelihood and cascade size predictions for BMH-K and BMH-P, resp.

Note that the data files in bayesian-mixture-hawkes/model_inference/bmhk_cnix/data/, bayesian-mixture-hawkes/model_inference/bmhk_rnix/data/ and bayesian-mixture-hawkes/model_inference/bmhp/data/ are clipped due to storage issues. Please adjust accordingly. Full datasets can be provided upon request.

Pipeline to run inference and prediction:

1. Compile the Stan models using compile_models.py .
2. Run BMH fitting on data with multiple_run_bayesianfit.sh
3. Obtain BMH predictions with multiple_run_bayesianpred.sh
4. Run collect_mean.py to get mean metrics.
5. (BMH-P) Run collect_pred.py to get cascade size predictions.

Headline Style Profiling

headline_style_profiling/contains the Python resources for the headline style profiling case study in the paper. There are two subdirectories.

• inflammatory_fakenix/ contains resources for the CNIX analysis.
• inflammatory_rnix/ contains resources for the RNIX analysis.

Raw headline data is in inflammatory_fakenix/data/. Code to generate the graphs in the case study are in inflammatory_rnix/theta/combined_analysis.ipynb.

License

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 piogabrielle.b.calderon@student.uts.edu.au or Marian-Andrei@rizoiu.eu.