Abstract
Social instant messaging services (SMS) such as WhatsApp, Snapchat, and WeChat have significantly changed communication patterns. In these platforms, social groups are central, yet the dynamic mechanisms underlying their evolution—specifically, who is invited to join—remain under-explored. We formalize the problem of predicting potential group invitees and propose a Multi-Level Factor Graph Model (ML-FGM) to capture the fundamental factors determining invitation probability. Using a massive dataset from WeChat, we demonstrate that our model significantly outperforms state-of-the-art baselines.
Introduction
Unlike public social networks, groups in SMS platforms like WeChat are characterized by definite membership, privacy, and limited capacity. Joining a group typically requires an invitation from an existing member. Predicting the next member is not merely a matter of finding similar users but requires understanding the social triggers of an invitation. We define the "invitee prediction problem": given the current state of a group and its "fringe" (friends of current members), who will join in the next time step?
The Multi-Level Factor Graph Model (ML-FGM)
To address this problem, we developed the ML-FGM, a probabilistic graphical model that integrates three layers of influence:
- Group Level (Group Constraint): Groups exhibit non-linear growth, expanding rapidly after creation and slowing as they mature. To linearize this for the model, we employ an "elastic time interval" strategy using a cubic root function.
- Peer Level (Peer Correlation): This captures the "fellowship" effect, where the invitation of one user is correlated with the invitation of their friends who are also in the group's fringe.
- Individual Level (Individual Attributes): This focuses on structural features like the number of friends a candidate has in the group and the presence of "adjacent triads."
Dynamics of Group Evolution
Our empirical analysis of WeChat data reveals that group growth is highly non-linear. As shown in Figure 1, the proportion of final group size reached increases sharply in the initial hours and minutes, while the probability of any specific fringe node being invited decreases over time. This decay in invitation probability necessitates the use of elastic time intervals to accurately model the "urgency" of invitations during different lifecycle stages.
Experimental Results
The study utilized a massive dataset from WeChat, tracking nearly 1 million invitation records over a year. We compared ML-FGM against standard classifiers (SVM, C4.5) and traditional link prediction algorithms (Adamic/Adar, Preferential Attachment).
- Accuracy: ML-FGM achieved an F1-measure of 0.411, significantly higher than the best baseline (SVM at 0.351).
- Precision: The model reached a Precision of 0.639, indicating high reliability in its top predictions.
- Ablation Study: Removing the "Group Factor" or "Peer Factor" resulted in a significant drop in performance, confirming that individual attributes alone are insufficient for predicting social invitations.
Conclusion
Predicting group invitees requires a multi-faceted approach that accounts for group-level growth dynamics and peer-level correlations. The ML-FGM provides a robust framework for this task. Future work could incorporate chat content or user demographics to further refine predictions and explore the distinction between receiving an invitation and choosing to join.
Summary updated