OPINN: Physics-Informed Neural Opinion Dynamics

A Python implementation of the architecture from "Advancing Opinion Dynamics Modeling with Neural Diffusion-Convection-Reaction Equation", with a Reddit data pipeline for real-world opinion tracking.

The model combines graph neural networks with physics-informed ODEs to forecast how opinions evolve across a social network. A Diffusion-Convection-Reaction (DCR) framework captures three forces shaping opinion change: diffusion (consensus via neighbor averaging), convection (directional drift), and reaction (nonlinear individual dynamics). These are integrated using classical ODE solvers (RK4/Euler) over a learned latent space.

.
├── opinn.py            # Core OPINN model (ODE solvers, DCR modules, encoder/decoder, training)
├── reddit_opinn.py     # Reddit scraping + sentiment analysis + OPINN pipeline
├── data/               # Scraped datasets (timeseries, adjacency, metadata CSVs)
├── results/            # Trained weight matrices, loss summaries, forecast plots
└── weights/            # Pre-trained weights for consensus/polarization scenarios

1. Scrape a subreddit for posts/comments matching a topic keyword
2. Score each text with DistilBERT sentiment (mapped to [-1, 1])
3. Build a user interaction graph (reply edges + thread co-participation)
4. Bin sentiment into a timeseries per user, forward-fill gaps
5. Train the OPINN model on the resulting [N, 1, T] tensor + adjacency matrix
6. Forecast future opinion trajectories and compare against a train-mean baseline

python -m venv .venv
source .venv/bin/activate
pip install torch numpy pandas matplotlib requests transformers

python reddit_opinn.py --subreddit technology --topic "AI" \
    --n_users 30 --days 30 --epochs 80

Key arguments:

from opinn import OPINN, run_experiment, evaluate

result = run_experiment(
    dataset_name="my_dataset",
    X_seq=X_seq,       # [N, 1, T] opinion timeseries
    adj=adj,            # [N, N] adjacency matrix
    hidden_dim=32,
    context_len=10,
    horizon=5,
    n_epochs=100,
)

model = result["model"]
metrics = evaluate(model, X_test, context_len=10, horizon=5)
print(f"MAE={metrics['MAE']:.4f}  RMSE={metrics['RMSE']:.4f}")

• results/<name>_opinion_dynamics.png -- Forecast plot with actual vs predicted mean opinion, percentile bands, and train/test split
• results/<name>_W_D.csv, W_C.csv, W_V.csv -- Learned diffusion, convection, and gating weight matrices
• results/<name>_omega.csv, delta.csv -- DCR mixing coefficients
• results/summary.csv -- Frobenius norms of all weight matrices across experiments
• data/<name>_timeseries.csv -- Per-user sentiment timeseries
• data/<name>_adjacency.csv -- User interaction graph
• data/<name>_metadata.csv -- Raw text, timestamps, and sentiment scores

• Reddit scraping uses public .json endpoints (no API key needed) with a 2-second rate limit
• Sentiment analysis runs on CPU via distilbert-base-uncased-finetuned-sst-2-english
• The weights/ directory contains pre-trained weights for consensus and polarization scenarios, along with cross-prediction distances