PROD - People Recommenders on Opinion Dynamics

Reference code for "The Effect of People Recommenders on Echo Chambers and Polarization" accepted for publication @ICWSM'22.

The repository structure has three main components:

1. the folder src where you can find the .py scripts that implement all the ingredients of the paper (e.g.: random network model, recommenders, etc.).
2. The folder Simulation_scripts contains a unique .py file that you need to invoke to run all the experiments.
3. The folder notebooks where you can find the .ipynb notebooks to reproduce the visualizations shown in the paper.

Reproducibility

If your goal is to reproduce the paper experiments, just follow these steps.

1. Install the conda environment: conda env create --file environment.yml
2. Run cd Simulation_scripts && python run_simulation.py with the proper configurations.
3. Run the Jupyter notebook notebooks/Network-visualizations.ipynb.

Re-usable components

However, PROD is a highly-customizable procedure with several reusable components:

• At src/recommenders.py you can find the implementation of the people-recommender algorithms.

  • twitter_wtf() is the implementation of what in the paper is called SALSA.
  • personalized_pagerank() is the implementation of Personalized PageRank algorithm.
  • calculate_DJ_row() is the implementation of Directed Jaccard heuristic.
  • biased_link_predictor() is the implementation of what in the paper is called OBA.
  • All the other recommenders are either inefficient versions of previous algorithms (used for debug purposes) or other algorithms similar to the ones used (e.g.: Adamic-Adar with Directed Jaccard).

• At src/synthetic_generator.py you can find the implementation of the random network model.

  • the main function is generate_G_and_opinions() which returns the networkX graph, the opinions vector, and the community assignments. If you want to reuse this module, we strongly warn that the LFR benchmark (the base algorithm to generate graphs with communities) is highly unstable, so stick with our values for the parameters avg_deg and power_law_coef. The parameters mu and conformism tune the amount of modularity and initial homophily.

• At src/measures.py you can find the implementation of the echo chambers/polarization measures adopted in the paper.

  • The functions compute_cont_correlation_neighbors() and compute_RWC() allows to compute the NCI and the RWC measures.
  • We left other implemented measures for those interested.

• At src/utils.py you can find auxiliary functions.

• At src/PROD_EPISTEMIC.py you can find the main class that allows running our PROD procedure under the Epistemological opinion update rule.

  • in compute_metrics() you can decide which metric you want to measure at the end of each round. Here, you could add your metric. Implement your function in measures.py, add it to the list, and you are done!
  • in recommender_sys() you can find how the various implemented algorithms are chosen in a PROD run. Here, you could add your recommender algorithm. Implement your function in recommenders.py, then add a specific if statement, and you are done!
  • the main function is simulate() that you can use to execute one single PROD run and allows you to input a random seed to assess the robustness of your results.

• At src/BCM.py you can find the extension of PROD with the bounded confidence opinion update rule.

  • You can clone this module to realize your own opinion update rule. It is quite simple, (i) extend the class contained in src/PROD_EPISTEMIC.py and (ii) override the method update_node_opinion() as done for BCM.