Learning Dynamic User Interactions for Online Forum Commenting Prediction

Predicting whether a user would interact with a particular message or topic in online social services is essential for applications related to recommendation systems. Two perspectives are commonly adopted when tackling this problem: the user's interest in the post content and the social relationship between posters and commenters. However, explicit social relationships might not be available in online forums, e.g., Stack Overflow. This makes it challenging for predictive models to understand the social connections between forum users. In this paper, we propose a novel framework to solve the comment prediction problem in online forums. Specifically, the framework incorporates attention mechanisms to capture users' interests in the post contents and a stacked graph convolutional network to perceive users' implicit social relationships from their past temporal interactions. The multi-modal features learned by the framework would be combined together through a fusion layer and used for the final prediction. We verify the effectiveness of our framework from multiple perspectives using real forum datasets. Experimental results show that our framework could achieve better performance than existing state-of-the-art methods.