Cross-Domain Event Participant Prediction with Public Event Description

Event participant prediction is an important problem when planning future events. Previous works have found that cold-start recommendation techniques can be used to solve the problem effectively. However, we consider the case of a newly started domain where training data for learning the recommendation model is limited. We propose to use a support domain data that has a similar user behavior model to help improve the training process. Our assumption does not involve linked users across domains, but uses public event description as the bridge. We show how such data can be combined with the target domain data under such assumptions. Experimental evaluation with real-world event participation datasets shows that adding a support domain data with our method does steadily improve prediction accuracy in the target domain.