Missing Data Imputation: A Systematic Comparison of Traditional, Machine Learning and Deep Learning Approaches

The most important step in data processing is handling missing data. Missing data introduces bias and degrades machine learning's model performance. Traditional imputation techniques, such as Mode Imputation and K-Nearest Neighbors with One-Hot Encoding, provide computationally efficient solutions but often fail to capture complex feature dependencies. Machine learning-based methods, including Bayesian Networks, CatBoost, and LightGBM imputation, enhance accuracy by leveraging feature interactions and probabilistic modeling. Deep learning-based approaches, particularly Generative Adversarial Imputation Networks (GAIN), Recurrent Neural Networks (RNN) and Transformer-based models (BERT-like Imputation), offer state-of-the-art performance by learning data distributions and capturing long-range dependencies. This paper presents a review of missing data imputation methods based on traditional, machine learning and deep learning approaches. A comparative analysis evaluates their effectiveness based on F1-score, precision, and recall, demonstrating that deep learning techniques, particularly GAIN and Transformer-based models, outperform others but require substantial computational resources. While these advanced methods yield high-quality imputations, challenges such as scalability, interpretability, and data dependency remain. This study provides a road map for researchers and practitioners to select appropriate imputation techniques, balancing accuracy, computational efficiency, and scalability for real-world applications.