FILLER: A New Feature Imputation Method Using Latent Space Search
FILLER is a feature imputation method that searches the two-dimensional latent space of a generative model to fill missing values in incomplete datasets. The generative model, G-NeuroDAVIS, is trained on fully observed data, and FILLER uses it to impute missing entries in corrupted test samples. The method includes a mathematical proof of convergence for its iterative search. Evaluated on image datasets under random and structured missingness patterns, FILLER addresses challenges in balancing scalability and structural consistency in real-world machine learning applications.
Key facts
- FILLER proposes feature imputation via latent location exploration and retrieval.
- The generative model used is G-NeuroDAVIS.
- FILLER searches a two-dimensional latent space.
- The method includes a mathematical proof of convergence.
- Evaluated on image datasets under random and structured missingness.
- Addresses challenges in balancing scalability and structural consistency.
- The generative model is trained on fully observed data.
- FILLER imputes missing values in corrupted test samples.
Entities
—