← All work
published2023 – 2024

DGDNN

Decoupled graph diffusion for multi-stock movement prediction.

  • Graph diffusion
  • Financial ML
  • Representation learning

Problem

Stock relationships and temporal representations can become entangled when graph propagation and feature transformation are learned together.

Approach

Decouple graph diffusion from representation transformation to control how relational information is mixed.

Role

First author and implementation author

DGDNN separates propagation from transformation to make relational mixing more controllable in dynamic stock graphs.