A Novel Mixed-Method Approach to Missing Value Imputation: An Introduction to MissMixed

Handling missing data is a pervasive challenge in machine learning and statistical analysis, often leading to biased models and reduced accuracy if not properly addressed. Numerous imputation algorithms have been proposed, ranging from simple statistical methods to complex machine learning models. In this paper, we introduce MissMixed, an innovative approach to the imputation problem that combines flexibility with robust performance. Unlike conventional imputation algorithms, MissMixed enables the dynamic selection of imputation algorithms and corresponding hyperparameters at each iteration, allowing it to adapt to a dataset's unique structure and missingness patterns. While the default configuration employs Random Forest and Gradient Boosting Tree models, MissMixed's modular design accommodates various machine learning algorithms and parameter settings, similar to the customizable layers in deep learning architectures. To evaluate MissMixed, we conducted extensive experiments on five diverse datasets, including the Breast Cancer Wisconsin and Letter Recognition datasets, with different missing rates and under various missingness mechanisms. We tested MissMixed alongside some of well-established methods, such as MissForest, MICE, KNN, and Variational Autoencoders (VAE), as well as simpler statistical imputation methods. The results reveal that MissMixed consistently achieves high accuracy and, in most scenarios, outperforms other methods. Its adaptability across various conditions underscores MissMixed”s potential as a powerful, versatile imputation approach suited to real-world applications.