AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-22
Quantum-Inspired Optimization Process for Data Imputation
Missing data can bias downstream inference in clinical datasets, so the authors propose a quantum-inspired imputation framework combining PCA with quantum rotation-based reconstruction and evaluate it on the UCI Pima Indians Diabetes dataset, where structural zeros are treated as missing. Rotation angles are selected by gradient-free optimizers (Simulated Annealing, COBYLA, and Differential Evolution), and missing values are reconstructed under distribution-aware bounds such as ±2 standard deviations to avoid unrealistic collapse toward the central tendency. Quantum-inspired variants achieve substantially better distributional fidelity under Wasserstein W1 than modern imputers — the best setting cuts average W1 by about 66% versus SoftImpute, 84% versus MissForest, 90% versus VAE, and 60% versus Mean imputation — while preserving variance and standard deviation closer to the observed data.