Using Optimization Algorithms for Effective Missing-Data Imputation: A Case Study of Tabular Data Derived from Video Surveillance

Data are crucial components of machine learning and deep learning in real-world applications. However, when collecting data from actual systems, we often encounter issues with missing information, which can harm accuracy and lead to biased results. In the context of video surveillance, missing data may arise due to obstructions, varying camera angles, or technical issues, resulting in incomplete information about the observed scene. This paper introduces a method for handling missing data in tabular formats, specifically focusing on video surveillance. The core idea is to fill in the missing values for a specific feature using values from other related features rather than relying on all available features, which we refer to as the imputation approach based on informative features. The paper presents three sets of experiments. The first set uses synthetic datasets to compare four optimization algorithms—Particle Swarm Optimization (PSO), Grey Wolf Optimization (GWO), Whale Optimization Algorithm (WOA), and the Sine–Cosine Algorithm (SCA)—to determine which one best identifies features related to the target feature. The second set works with real-world datasets, while the third focuses on video-surveillance datasets. Each experiment compares the proposed method, utilizing the best optimizer from the first set, against leading imputation methods. The experiments evaluate different types of data and various missing-data rates, ensuring that randomness does not introduce bias. In the first experiment, using only synthetic data, the results indicate that the WOA-based approach outperforms PSO, GWO, and SCA optimization algorithms. The second experiment used real datasets, while the third used tabular data extracted from a video-surveillance system. Both experiments show that our WOA-based imputation method produces promising results, outperforming other state-of-the-art imputation methods.