Designed to Spread: A Generative Approach to Enhance Information Diffusion

Ziqing Qian*, Jiaying Lei*, Shengqi Dang, Nan Cao
Tongji University

---

Overview

Social media has fundamentally transformed how people access information and form social connections, with content expression playing a critical role in driving information diffusion. While prior research has focused largely on network structures and tipping point identification, it provides limited tools for automatically generating content tailored for virality within a specific audience.

To fill this gap, we propose the novel task of Diffusion-Oriented Content Generation (DOCG) and introduce an information enhancement algorithm for generating content optimized for diffusion. Our method includes:

• 📊 Influence Indicator — enables content-level diffusion assessment without requiring access to network topology
• ✏️ Information Editor — employs reinforcement learning to explore interpretable editing strategies, leveraging generative models to produce semantically faithful, audience-aware textual or visual content

Experiments on real-world social media datasets and a user study demonstrate that our approach significantly improves diffusion effectiveness while preserving the core semantics of the original content.

---

Paper

Designed to Spread: A Generative Approach to Enhance Information Diffusion
Ziqing Qian, Jiaying Lei, Shengqi Dang, Nan Cao
Proceedings of the AAAI Conference on Artificial Intelligence, 2026, 40(2): 944–952
📄 Paper | 🔗 DOI

---

Repository Structure

designed-to-spread/
├── Data/                    # Raw inputs & pipeline outputs (posts/accounts JSONL, features, models, …)
├── Data_preprocess/         # End-to-end preprocessing (PySpark + Python; entry: main.py)
├── Long_CLIP/               # Long-CLIP model code (used by image / CLIP features)
├── baselines/               # llm_baseline.py, in_context_learning.py
├── influence_indicator/     # train_pijc_model.py, Evaluation.ipynb
├── information_editor/      # text_editor.py, visual_editor.py, base_editor.py (imported by scripts/)
├── scripts/                 # Runnable entry points: text/visual editors, tweet→image preprocess
└── requirements.txt         # Python dependencies

Note: There is no preprocess.ipynb or run_pipeline.sh in this repo. Preprocessing is driven by Data_preprocess/main.py (and its modules). The only notebook is influence_indicator/Evaluation.ipynb (evaluation / analysis). Text and image editing are run via scripts/run_text_editor.py and scripts/run_visual_editor.py.

---

Getting Started

Step 1 — Prerequisites

Make sure you have the following installed:

• Python >= 3.8
• CUDA 12.1 (required for GPU acceleration; the dependencies are built against cu121)
• Conda (recommended for environment management)

Step 2 — Clone the Repository

git clone https://github.com/idvxlab/designed-to-spread.git
cd designed-to-spread

Step 3 — Create a Virtual Environment

It is strongly recommended to use a dedicated virtual environment to avoid dependency conflicts:

# Using conda (recommended)
conda create -n docg python=3.10
conda activate docg

# Or using venv
python -m venv venv
source venv/bin/activate   # On Windows: venv\Scripts\activate

Step 4 — Install Dependencies

pip install -r requirements.txt

Key dependencies include:

Package	Version	Purpose
torch	2.4.1	Core deep learning framework
transformers	4.46.3	Pre-trained language models
diffusers	0.36.0	Diffusion-based image generation
openai	2.2.0	LLM API calls
scikit-learn	1.3.2	Machine learning utilities
pyspark	3.5.8	Large-scale data processing

Note: If you do not have a CUDA 12.1 GPU, you may need to install a CPU-only version of PyTorch separately. See the PyTorch installation guide.

Step 5 — Prepare the Data & Run Preprocessing

1. Place (or symlink) your social data under Data/ as expected by Data_preprocess/extract_tables.py:

   • posts.jsonl (required), optionally posts_all.jsonl
   • accounts.jsonl (required for the user table)

   The released sample under Data/raw_data/ only contains ID lists and small *_example.jsonl snippets; for the full pipeline you need full posts.jsonl / accounts.jsonl (or pass explicit paths).

2. Run the eight-step preprocessing pipeline from the repo root (uses PySpark in step 1):

cd Data_preprocess
python main.py
# Optional: custom Data dir or JSONL paths
# python main.py --base_dir /path/to/Data --posts_path ... --accounts_path ...

main.py runs, in order: extract_tables → extract_network → split_dataset → generate_samples → filter_texts → compute_message_features → compute_user_features → compute_features_and_labels.
You can also import and run any single module’s main() for debugging. Intermediate artifacts are written under Data/ (e.g. features, splits, outputs/ for .pt training packs).

Step 6 — Train the Influence Indicator (PIJC)

Train the cascade / diffusion predictor on the preprocessed .pt samples:

cd influence_indicator
python train_pijc_model.py

By default this reads Data/outputs/outputs_train and Data/outputs/outputs_test, and saves checkpoints under Data/outputs/model/ (e.g. model_epoch_100.pth referenced elsewhere in the repo).

Optional — Jupyter evaluation: open and run influence_indicator/Evaluation.ipynb.
(If cells fail to import Long-CLIP weights, note the repo directory is Long_CLIP/ (underscore); align sys.path and model_path in the notebook with that folder and your weight file, e.g. longclip-B.pt.)

Step 7 — Run the Information Editor (RL)

Core logic lives in information_editor/ (text_editor.py, visual_editor.py). Entry points are under scripts/ from the repository root:

Text editing

python scripts/run_text_editor.py
# See defaults and overrides:
python scripts/run_text_editor.py --help

Image / visual editing (expects CSVs with tweet_id, tweet_text, image_prompt, plus images under --orig_image_folder):

python scripts/run_visual_editor.py
python scripts/run_visual_editor.py --help

Defaults assume a trained PIJC model at Data/outputs/model/model_epoch_100.pth, user features at Data/features/user_features_dict.pt, and tweet splits under Data/tweet2image/ (see script --help).

Tweet → image preprocessing & train/test split (uses LLM_API_KEY / LLM_BASE_URL from a .env file; generates prompts, images, and train_tweets.csv / test_tweets.csv):

python scripts/tweet2image_preprocess_and_split.py
# Edit the paths in the `if __name__ == "__main__"` block if your CSV/output locations differ.

Step 8 — Run Baselines (Optional)

Two Python scripts (no notebooks):

# LLM-style baseline
python baselines/llm_baseline.py --help

# In-context learning baseline
python baselines/in_context_learning.py --help

Use --text and/or --image to select modalities; see argparse defaults for CSVs, model paths, and output dirs.

Step 9 — Pipeline Summary

There is no single scripts/run_pipeline.sh. A typical order is:

1. Data_preprocess/main.py → features & Data/outputs/ tensors
2. influence_indicator/train_pijc_model.py → Data/outputs/model/*.pth
3. (Optional) scripts/tweet2image_preprocess_and_split.py → Data/tweet2image/ for the image track
4. scripts/run_text_editor.py / scripts/run_visual_editor.py (and/or baselines/*.py)

Configure paths via each script’s CLI flags or the hard-coded defaults in tweet2image_preprocess_and_split.py’s main block where applicable.

---

Citation

If you find this work useful in your research, please consider citing:

@inproceedings{qian2026designed,
  title     = {Designed to Spread: A Generative Approach to Enhance Information Diffusion},
  author    = {Qian, Ziqing and Lei, Jiaying and Dang, Shengqi and Cao, Nan},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  volume    = {40},
  number    = {2},
  pages     = {944--952},
  year      = {2026},
  doi       = {10.1609/aaai.v40i2.37063},
  url       = {https://ojs.aaai.org/index.php/AAAI/article/view/37063}
}

---

Contact

If you have any questions, feel free to open an issue or contact us at 2411920@tongji.edu.cn.

---

Acknowledgments

This project is from the Intelligent Big Data Visualization Lab (iDVX Lab) at Tongji University. The projects released here are associated with publications from the lab.