A Unified Probabilistic Model of User Activities and Relations on Social Networking Sites

Xiaofeng Yu, Junqing Xie, Shuai WangIJCAI 2015

Introduction

In social media ecosystems, a fundamental "chicken-and-egg" problem exists: do users follow others because of shared interests, or do they develop those interests because of the people they follow? This phenomenon, termed Bidirectional Mutual Interaction (BMI), suggests that user activities and social relations are deeply intertwined. Traditionally, research has treated these as separate domains, often relying on static heuristics like homophily—the idea that "birds of a feather flock together." However, social influence is dynamic and varies significantly between individuals.

To address this, researchers from HP Labs China developed JAR (Joint Activity and Relation), a unified probabilistic framework designed to simultaneously predict user activities and discover social relationships by quantitatively measuring social influence.

The JAR Model

The JAR model is a generative probabilistic framework that moves beyond binary friendship statuses. It treats social influence as a learnable probability, $P(s|u)$, representing the likelihood that a user $u$ will be influenced by the opinions of user $s$.

The JAR probabilistic graphical model illustrating the dependencies between user preferences, social influence, and activities.

Key Components

The model incorporates several latent variables:

  • Latent Topics ($t$): Hidden themes representing a user's personal preferences.
  • Social Influence ($s$): A variable that captures the strength of influence between users, allowing the model to distinguish between "tastemakers" and "followers."
  • Unified Inference: By calculating the joint probability of activities and relationships, JAR balances personal preference against social influence using specific weighting factors ($\alpha$ and $\beta$).

To optimize the model, the authors introduced a Generalized Expectation Maximization (GEM) algorithm. By employing "deterministic annealing"—a cooling process that helps avoid local optima—the GEM approach ensures a more stable, global solution while accelerating the model-fitting process.

Experimental Evaluation

The model was rigorously tested using a real-world dataset from Twitter, comprising over 5,000 users, 22,000 friendship links, and nearly 300,000 activity instances (tweets, comments, likes, and reshares).

Key Results

The JAR model consistently outperformed baseline methods, such as Logistic Regression (LR) and Friendship-Interest Propagation (FIP), in both activity prediction and relationship discovery.

F-measure performance of the JAR model compared to baseline methods across different topic sizes.
  • Activity Prediction: JAR achieved an F1-measure of 83.82%, a significant improvement over the 68.39% achieved by standard Logistic Regression.
  • Relation Discovery: The model reached an F1-measure of 89.81%, compared to 73.86% for decoupled baselines.
  • Optimal Weighting: The researchers discovered a "sweet spot" for autonomy, where the best predictions occurred when personal preference was weighted at 70-80% and social influence at 20-30%.

Conclusion

The JAR model demonstrates that social influence is a quantifiable, dynamic variable that can be learned from behavioral evidence. By modeling activities and relations together, the framework captures the feedback loop inherent in social networks more effectively than decoupled approaches. Future work will focus on scaling the model to larger datasets, exploring semi-supervised learning scenarios, and incorporating temporal dynamics to account for the evolution of social networks over time.

Summary updated

BibTeX
@inproceedings{yu2015unified,
  author = {Xiaofeng Yu and Junqing Xie and Shuai Wang},
  title = {{A Unified Probabilistic Model of User Activities and Relations on Social Networking Sites}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}
Figure