Predicting network members from partial contact records on social media: A machine learning approach.

Surveys conducted on social groups often generate incomplete information due to imperfect response rates. Drawing on Facebook data from a nationally representative sample of graduating college students in Taiwan, we examined the extent to which partial contact records predict which Facebook users belong to a specific class. We first used data from classes with low to middle response rates to train a model for classmate prediction. Based on data from classes with high or perfect response rates, we simulated data by using four different sampling methods with various response rates, and applied the trained model on simulated data to classmate prediction. With a minimal response rate of 40 percent, we achieved an accuracy rate of 90 percent and a true positive rate of 86 percent. Chronological order sampling had the best prediction performance, followed closely by popularity sampling, then by random sampling, and lastly by unpopularity sampling. • Sociological surveys often generate incomplete information. • A machine learning approach to identify members of a specific social group from incomplete social media data (Facebook). • Social media data from 40 % of the group members is sufficient to successfully identify 86 % of the group members. • Social media data from early respondents result in better models for membership prediction. • Social media data from popular members also result in better models for membership prediction. [ABSTRACT FROM AUTHOR]