Abstract
Based on the fourth-wave Beijing College Students Panel Survey (BCSPS), this study aims to provide accurate estimation of the percentage of the potential sexual minorities among the Beijing college students by using machine learning methods. Specifically, we employ random forest (RF), an ensemble learning approach for classification and regression, to predict the sexual orientation of those who were not willing to disclose his/her inherent sexual identity. To overcome the imbalance problem arising from far different numerical proportion of sexual minority and majority members, we adopt the repeated random sub-sampling for training set by partitioning those who expressed heterosexual orientation into different number of splits and further combining each split with those who expressed sexual minority orientation. The prediction from 24-split random forest suggests that youths in Beijing with sexual minority orientation amount to 5.71%, almost two times that of the original estimation 3.03%. The results are robust to alternative learning methods and covariate sets. Besides, it is also suggested that random forest outperforms other learning algorithms, including AdaBoost, Naïve Bayes, support vector machine (SVM), and logistic regression, in dealing with missing data, by showing higher accuracy, F1 score, and area under curve (AUC) value.