Navigating Data Corruption in Machine Learning: Balancing Quality, Quantity, and Imputation Strategies

Data corruption, including missing and noisy entries, is a common challenge in real-world machine learning. This paper examines its impact and mitigation strategies through two experimental setups: supervised NLP tasks (NLP-SL) and deep reinforcement learning for traffic signal control (Signal-RL). This study analyzes how varying corruption levels affect model performance, evaluate imputation strategies, and assess whether expanding datasets can counteract corruption effects. The results indicate that performance degradation follows a diminishing-return pattern, well modeled by an exponential function. Noisy data harm performance more than missing data, especially in sequential tasks like Signal-RL where errors may compound. Imputation helps recover missing data but can introduce noise, with its effectiveness depending on corruption severity and imputation accuracy. This study identifies clear boundaries between when imputation is beneficial versus harmful, and classifies tasks as either noise-sensitive or noise-insensitive. Larger datasets reduce corruption effects but offer diminishing gains at high corruption levels. These insights guide the design of robust systems, emphasizing smart data collection, imputation decisions, and preprocessing strategies in noisy environments.