When Can We Ignore Missing Data in Model Training?

Imputing missing data is typically expensive, and as a result, people seek to avoid it when possible. To address this issue, we introduce a method that determines when data cleaning is unnecessary for machine learning (ML). If a model can minimize the loss function regardless of the missing data's actual values, then data cleaning is not required. We offer efficient algorithms for checking this condition in multiple ML problems, and by analyzing the algorithms, we show that data cleaning is unnecessary when dealing with irrelevant and redundant data. Our preliminary experiments demonstrate that our algorithms can significantly reduce cleaning costs compared to a benchmark method, without incurring much computational overhead in many cases.