First author and implementation author · 2023–2025
Relational and individual structure
Stocks carry individual temporal features as well as changing relationships to other stocks. Repeatedly interleaving transformations with graph propagation can make those roles hard to separate.
Decoupled representation learning
DGDNN builds dynamic graphs using entropy-driven edge generation, learns dependencies through graph diffusion, and separates the representation-learning stage from relational propagation. The extended publication develops the same research direction.
Reading the demonstration
Start with six signed features, apply a fixed nonlinear encoder once, then diffuse its output over a ring. Controls vary the encoder gain and the number of diffusion steps. This is an operational illustration of decoupling, not the trained encoder, learned graph, or full DGDNN architecture.
Interactive toy experiment
Encode once. Then diffuse.
Set the encoder gain and step through graph propagation. The encoded features stay fixed while their mixture changes.
Gray: raw input · Light blue: encoded once · Dark blue: diffused representation
Encoder gain 1.5 · 2 diffusion steps · Node A: -0.90 → -0.874 → -0.233. Zero steps preserves the encoded features.
How this toy works
Encode hᵢ = tanh(gxᵢ) once. Each step uses hᵢ ← 0.5hᵢ + 0.25hᵢ₋₁ + 0.25hᵢ₊₁ on a six-node ring.
This fixed encoder and graph isolate decoupling. They do not reproduce entropy-driven graph construction, learned features, or the full DGDNN operator.
Synthetic illustration · Not a trained-model output or a published benchmark.