I built ModerationDEQ, a Python project titled “Simulating Online Community Health with Deep Equilibrium Models,” to study long-term online-community behavior under moderation policies. It is a policy-analysis model, not a classifier or next-step predictor: I aim to learn what steady state a community reaches when a policy stays in place.
I represent community evolution as z_{t+1}=f(z_t, {policy}) and seek the equilibrium z*=f(z*, {policy}) rather than unrolling the dynamics indefinitely. The system flow starts from an initial community state, applies a neural dynamics function, and uses a fixed-point solver to obtain the long-term equilibrium. Training uses implicit differentiation, so gradients flow through the equilibrium rather than through time steps. I use Deep Equilibrium Networks to represent infinite-depth behavior, feedback loops, and stable fixed points without explicit recurrence.
- I represent community state with four interpretable dimensions, each bounded in [0, 1]: Content Quality (signal-to-noise ratio), Toxicity (harmful behavior level), Moderation Pressure (reports or workload), and Engagement (participation level). Policies have two parameters: Strictness (moderation aggressiveness) and Threshold (tolerance for toxicity). I embed policy parameters non-linearly to allow expressive policy effects.
- The README reports stable equilibria with residuals around 1e-5, policy-dependent regime shifts, toxicity–engagement trade-offs, robust convergence from different initial states, and no trivial collapse. The listed experiments include a synthetic dataset inspired by real community behavior, equilibrium-residual tracking, policy-sweep visualizations, tests across multiple initial-state basins, a DEQ-versus-MLP baseline comparison, and ablations of stability versus realism. The captured README gives no dataset size, architecture or training details, solver choice, evaluation protocol, baseline or ablation results, or supporting measurements beyond the approximate residual.
python -m venv venv
venv\Scripts\Activate.ps1
pip install torch numpy pandas matplotlib
python data/generate_dataset.py
python train.py
python -m analysis.visualize_equilibria