As more and more people rely on social platforms such as Facebook for holding social activities among friends, this paper addressed the need of guest recommendation to ease the task of selecting proper guests from a large number of friends for invitation to kinds of specific types of activities. For the problem, this paper proposed the Guest Invitation to Hosts (GIH) algorithm to learn the dominant factors behind the guest invitations from historical data. Concerning that there are usually multiple factors dominating the guest invitation and the factors differ from the types of social activities, GIH consists of two learning mechanisms: (1) the learning of discrimination between activity types, by the topic words and descriptions of activities; and (2) the learning of latent dominant factors for guest invitation (to specific types of activities), by the Non-negative Matrix Factorization and Top-k Frequent Pattern mining techniques. Evaluated on the Facebook data, GIH can reach an accuracy of at least 80% in the guest recommendation, and outperform naïve approaches in terms of precision, recall, F1 score and accuracy. The case studies showed that the dominant factors (rules) identified by GIH comply with the human intuition. The latent dominant factors (rules) identified by GIH for guest recommendation can further be used as references for advanced studies in social science.