Estimating the proportion of sexual minorities through surveys is challenging, as many respondents conceal their orientation, leading to systematic underreporting. Using data from the Beijing College Students Panel Survey (BCSPS), this chapter applies supervised machine learning to predict the orientation of respondents unwilling to disclose their identity. To mitigate class imbalance, heterosexual respondents were randomly split into subsets and each subset was paired with minority samples for model training. The prediction from random forest suggests that 5.71% of Beijing youths identify as sexual minorities, nearly twice the original 3.03%. The results demonstrate the value of machine learning for improving prediction accuracy and handling missing data in quantitative social science.