MotifRole-Diff: Risk-Optimal Role-Aware Masking for Molecular Graphs
A new masked discrete diffusion model for generating molecular graphs, named MotifRole-Diff, has been introduced by researchers. This innovative approach replaces the conventional uniform corruption schedule with a strategy that is aware of roles. It assigns masking rates according to the empirically determined difficulty of denoising and the impact of various molecular token roles, all while maintaining the original model architecture, a clean sequence space, and a lossless decoder. A theorem defines the optimal allocation of a fixed masking budget among token roles, enhancing the role-weighted residual loss. This method overcomes the challenge of treating structurally diverse molecular components as equally difficult to reconstruct. The paper can be found on arXiv with the ID 2607.21634.
Key facts
- MotifRole-Diff is a role-aware corruption process for masked molecular graph diffusion.
- It allocates masking rates based on denoising difficulty and perturbation impact.
- The method preserves model architecture, clean sequence space, and lossless decoder.
- A theorem characterizes risk-optimal allocation of a fixed masking budget.
- The approach addresses uniform corruption's limitation in handling heterogeneous molecular components.
- The paper is on arXiv with ID 2607.21634.
Entities
Institutions
- arXiv