GDPM is a gradient descent-based polarization minimization algorithm used to reduce online media polarization by finding the optimal user-interest matrix given under a limited budget.
We release two anonymized graph datasets, TwitterSmall and TwitterLarge, with ground-truth opinions, user interest and influence. TwitterSmall contains more than 1000 nodes and TwitterLarge has more than 27,000 nodes (the previously largest publicly available dataset contains less than 550 nodes). For details of data collection and processing, please check Appendix B.1 in the [paper](THIS IS A LINK TO ARXIV).
The datasets are in main/data/. TwitterLarge.txt (TwitterSmall.txt) is the anonymized edge list. We provide 2 data formats .jld and .csv in the dataset. The code is written in Julia and it takes .jld as the default data format. The s.jld and opinion.csv are the ground-truth opinions. X.jld and user_interest.csv are the user-to-topic matrices that encode users' interest information towards topics. Y.jld and topic_influence.csv is the topic-influence matrix that encodes users' influence scores among topics. The two matrices are both row-stochastic.
The following table summarizes the dataset in two data formats.
| Julia format | CSV format | Description |
|---|---|---|
| s.jld | opinion.csv | innate opinion data, vector |
| X.jld | user_interest.csv | user interest matrix |
| Y.jld | topic_influence.csv | topic influencer matrix |
Real-world datasets are publicly available from Network Data Repository. In this project, we only provide Advogato.txt in /graph_data/ folder. Please check Appendix B.1 for details about how to generate synthetic opinion, user-interest and topic-influence matrices.
We provide Erdős–Rényi random graphs for Convex.jl experiments. The data are located in main/data/Random_*.
Our code is written by Julia 1.7.2. Run the following command to install packages:
julia packages.jl
Run the following command to start optimization on two Twitter datasets with ground truth data.
sh GDPM_BLs.sh
Run the following command to generate synthetic innate opinion, user-topic matrix, and topic-influence matrix and start optimization on synthetic data, here we use Advogato dataset as an example:
julia main_simulation.jl Advogato 0.1 10 100 polarized powerlaw
Convex.jl is a Julia package for Disciplined Convex Programming (DCP). We provide Convex.jl program with SCS solver to solve our optimization problem on random graphs. Run the following command to start optimization:
sh convex_box.sh
The output results of experiments are saved in folder output/. The approximation error during optimization is saved in output/Errors/. The output/XTs/ folder contains user_interest matrix after optimization. The output/Opts/ folder contains the intermediate result of polarization-disagreement index during optimization. The output/Logs/ folder contains output from main_GDPM_BLs.jl and main_simulation.jl. The output/figs/ folder contains figures to visualize the aforementioned results.
This project uses .jld file format to save all output results. In order to use those results in other programming languages, we provide a tool to convert .jld files to .csv files. The following example shows how to use the tool:
julia convert2csv.jl "data/TwitterSmall/Y.jld" "data/TwitterSmall/topic_influence.csv"
where the "data/TwitterSmall/Y.jld" is the Julia data path and "data/TwitterSmall/topic_influence.csv" is the output file path.