A Safe Approach to Shrink Email Sample Set while Keeping Balance between Spam and Normal

To deal with any possible cases for training anti-spam machine learning models, it is crucial to design a safe way to shrink the size of training sample set via reducing redundancies with minimal information loss for classification as well as make distribution of samples balanced. Presently, there is no such solution to do so. In this paper, we propose a safe approach to address these problems and improve the quality of training email sample pool (set) for getting high quality machine learning models for better anti-spam engine with non-biased high spam detection rates as well as low false positive rates.