First author and implementation author · 2023–2024
More than one kind of relationship
Stock relations vary over time and can depend on the horizon at which they are measured. A single static relation graph cannot capture all of that structure.
Diffusion and retained representations
MGDPR constructs multiple stock graphs, refines relations through diffusion, and combines diffused features with retained representations. Parallel retention helps connect relational context to distinctive node features.
Reading the demonstration
Two hand-built relation graphs carry a synthetic scalar signal. Toggle either graph to inspect its contribution, then blend the diffused signal with the original features. This linear blend illustrates the separate branches; the paper’s learned parallel-retention layer uses projections and interactions rather than this scalar blend. No stock returns or predictions are shown.
Interactive toy experiment
Separate relations, retained features
Enable the relation graphs, then change how much of each original node feature is retained.
Gray: input · Light blue: relation A · Hatched: relation B · Dark blue: combined
2 relations enabled · Original-feature weight 0.35 · Node A output 0.629. Weight 1 retains every original feature.
How this toy works
Each relation averages a node with its neighbors. Let d be the mean of enabled relation outputs, or zero if both are off. The output is h = ρx + (1 − ρ)d.
The scalar blend illustrates the separate branches. The learned parallel-retention operator instead uses projections and interactions. Features and graphs are synthetic, not financial predictions.
Synthetic illustration · Not a trained-model output or a published benchmark.