Beyond Content: Behavioral Policies Reveal Actors in Information Operations

This repository contains the code to reproduce the experiments and figures for the paper:

Beyond Content: Behavioral Policies Reveal Actors in Information Operations. XXX, TBD (Under Review), 2025.

The codebase builds on the data-processing pipeline introduced in:

Lanqin Yuan, Philipp J. Schneider, and Marian-Andrei Rizoiu. 2025. Behavioral Homophily in Social Media via Inverse Reinforcement Learning: A Reddit Case Study. In Proceedings of the ACM Web Conference 2025 (WWW ’25), 576–589. https://doi.org/10.1145/3696410.3714618

The repository is organized into three main components:

1. Data processing – from raw Reddit dumps to user trajectories and inferred policies.
2. Classification experiments – training and evaluating policy- and content-based classifiers.
3. Visualization – scripts to reproduce the figures in the paper.

1. Getting Started

• Python >= 3.10
• Install dependencies

pip install -r requirements.txt

2. Repository Structure

At a high level, the repository is organized as:

.
├── reddit_processing/      # Reddit dumps → cleaned user trajectories
├── policy_inference/       # IRL / GAIL inference
├── experiments/            # Classification experiments and evaluation
├── visualization/          # Plotting and figure generation
├── requirements.txt        # Python dependencies
└── README.md               # This file

You do not need to run all components if you only want to reproduce specific stages (e.g., classification on pre-computed policies), but the full pipeline is:

1. reddit_processing/
2. policy_inference/
3. experiments/
4. visualization/

3. Data Processing

Reddit Processing and Trajectory Construction

Scripts for processing raw Reddit data and constructing user trajectories are located in raw_dump_processing/.

Inputs:

• Monthly Reddit comment and submission dumps from the Pushshift API (compressed JSON files).
• An additional file listing the organic users used in this work.

Important (Ethics): The list of organic users cannot be shared publicly due to ethics and privacy constraints. To reproduce the full dataset, you will need access to a comparable set of organic users from Reddit and must obtain appropriate ethics approval for your institution.

To reproduce our data processing pipeline, run each of the scripts inside raw_dump_processing/ in order of the number at the beginning of each script.

Each script reads the outputs of the previous step and writes intermediate results to data/processed/ (or the path specified in the script). The final outputs are:

• A set of user trajectories (trolls and organic users) with state–action sequences.
• Metadata used downstream for policy inference and classification.

4. Policy Inference

Scripts for inferring user policies from trajectories are located in policy_inference/.

Supported policy representations include:

• Empirical policies (state–action frequency estimates).
• MaxEnt Deep IRL policies (maximum-entropy deep inverse reinforcement learning).
• GAIL (Generative Adversarial Imitation Learning) policies.

These scripts assume that all preprocessing steps in raw_dump_processing/ have been successfully completed.

5. Classification Experiments

Classification experiments are located under experiments/. These scripts:

• Load the inferred policies (policy-based representations).
• Load content-based representations (e.g. text embeddings).
• Train and evaluate classifiers (e.g. random forest, gradient boosting).

6. Visualization

Figure-generation scripts are located in visualization/. They mainly take the result files from experiments/ and other raw data and produce the plots shown in the paper.

7. Citation

If you use this codebase or build upon our methods, please cite:

@article{xxx2025beyondcontent,
  title        = {Beyond Content: Behavioral Policies Reveal Actors in Information Operations},
  author       = {XXX},
  journal      = {TBD (Under Review)},
  year         = {2025}
}

@inproceedings{yuan2025behavioralhomophily,
  title        = {Behavioral Homophily in Social Media via Inverse Reinforcement Learning: A Reddit Case Study},
  author       = {Yuan, Lanqin and Schneider, Philipp J. and Rizoiu, Marian-Andrei},
  booktitle    = {Proceedings of the ACM Web Conference 2025 (WWW '25)},
  pages        = {576--589},
  year         = {2025},
  publisher    = {ACM},
  address      = {New York, NY, USA},
  doi          = {10.1145/3696410.3714618}
}

8. Contact

For questions about the code or the paper, please contact:

• XXX