Impute Missing Entries with Uncertainty (U-VAE)

This repository is the official implementation of 'Impute Missing Entries with Uncertainty' (AAAI, 2026).

NOTE: This repository supports WandB MLOps platform!

Overview

Dataset

Download and add the datasets into data folder to reproduce our experimental results.

Reproducibility

Arguments

• --dataset: dataset options (anuran, banknote, breast, concrete, default, kings, letter, loan, redwine, shoppers, whitewine)
• --missing_type: how to generate missing (MCAR, MAR, MNARL, MNARQ)
• --missing_rate: missingness rate (default: 0.3)
• --M: the number of multiple imputation (default: 100)

Training

python main.py --dataset <dataset> --missing_type <missing_type> --missing_rate <missing_rate> 

Imputation & Evaluation

RQ1. Does U-VAE achieve state-of-the-art performance in single imputation tasks?

python imputer.py --dataset <dataset> --missing_type <missing_type> --missing_rate <missing_rate> 

RQ2. Can U-VAE support statistically valid multiple imputation by capturing uncertainty in the imputed values?

python imputer.py --dataset <dataset> --missing_type <missing_type> --missing_rate <missing_rate> --M <M>

RQ3. How robust is U-VAE to varying missingness rates and patterns in sensitivity analyses?

python imputer.py --dataset <dataset> --missing_type <missing_type> --missing_rate <missing_rate> 

Directory and codes

.
+-- data
+-- assets 
+-- datasets
|       +-- preprocess.py
|       +-- raw_data.py
+-- modules 
|       +-- evaluation.py
|       +-- evaluation_multiple.py
|       +-- metric_congeniality.py
|       +-- metric_fidelity.py
|       +-- metric_utility.py
|       +-- missing.py
|       +-- model.py
|       +-- train.py
|       +-- utility.py
+-- main.py
+-- imputer.py
+-- U-VAE_supp.pdf
+-- U-VAE.png
+-- README.md

Citation