Genetic Algorithm-Enhanced CMILK for Robust Thermal Facial Data Imputation

Missing data are ubiquitous in real-world datasets and can substantially degrade the performance of statistical analyses and deep learning models when essential features are absent. Imputation methods address this problem by replacing missing entries with plausible values that preserve the underlying data structure. In this work, we evaluate CMILK (Correlation-based Multiple Imputation with Localk-neighbour matching) for imputing missing thermal facial landmark temperatures and present a Genetic Algorithm (GA)–based procedure to optimise its hyperparameters. Comparative studies verified the effectiveness of GA, while ablation studies further validate the effectiveness of the GA-driven optimisation and its impact on convergence behaviour. GA-optimised CMILK achieves consistent improvements over the baseline: mean absolute error (MAE) decreases from 0.5724 to 0.5479 (a 4.3% relative reduction) and root mean square error (RMSE) from 0.7657 to 0.7371 (a 3.7% relative reduction). Hypothesis testing confirms that the observed reduction in RMSE after GA optimisation is statistically significant, supported by consistent parametric and nonparametric analyses with a large effect size and positive confidence interval. The optimisation also widens CMILK’s lead over the next-best method to 5.0% in MAE and 7.0% in RMSE, while producing a modest runtime improvement of 12.1% versus the baseline. These results show that automated hyperparameter search using GA produces measurable gains in accuracy and efficiency, improving CMILK’s reliability for predictive modelling and robust analysis of thermal facial-landmark temperature data.