Reproduce “MissHDD: Hybrid Deterministic Diffusion for Hetrogeneous Incomplete Data Imputation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Youran Zhou, Mohamed Reda Bouadjenek, Sunil Aryal
MissHDD introduces a hybrid deterministic diffusion framework for heterogeneous incomplete tabular data, employing a two-channel architecture that handles numerical and categorical features distinctly. This approach achieves consistently lower imputation errors and maintains downstream task performance, while delivering up to a 5x inference speed-up and enhancing imputation stability compared to existing stochastic diffusion models.
Key insights
- Respecting the distinct statistical manifolds of heterogeneous data types through a specialized two-channel diffusion architecture prevents semantic drift and information collapse, improving imputation quality.
- Adopting deterministic diffusion (DDIM) for continuous variables drastically reduces inference time and eliminates output variance, making the imputation process more practical and reliable for tabular data.
- Training with a self-masking strategy on incomplete data effectively exposes the model to diverse missingness patterns, enhancing its robustness and generalizability to real-world scenarios (MCAR, MAR, MNAR).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.