Personalized Prediction of Trust Links in Social Networks (Student Abstract)

Alexandre Parmentier, Robin CohenAAAI 2020 · DOI

Introduction

Modern recommender systems often leverage "trust-aware" algorithms, operating on the principle that users are likely to share preferences with those they trust. However, these systems face two primary challenges: data sparsity in explicit friendship networks and the "one-size-fits-all" assumption that trust is formed identically across all users. This paper introduces Personalized Multi-Faceted Trust Modeling (PMFTM), a framework designed to predict trust links by clustering users and learning group-specific trust predictors.

Methodology

The PMFTM approach consists of a three-stage pipeline:

  1. Clustering: Users are partitioned into groups based on similarity. The authors evaluated two metrics: Social Circle Overlap (Jaccard Similarity) and Review Score Correlation (Pearson Correlation).
  2. Trust Link Prediction: Instead of a global model, a distinct Logistic Regression classifier is trained for each cluster using 18 multi-faceted features, including user popularity and feedback ratios.
  3. Recommendation Evaluation: Predicted trust links are integrated into the TrustMF recommender system to predict user ratings on a test set.

Experimental Results

The study utilized a subset of the Yelp dataset, comparing the personalized model (PMFTM) against generic predicted trust (MFTM) and explicit friendship data (RealFriends). The performance of the recommendation engine is highly sensitive to the "social regulation" parameter, which weights the influence of trust links.

Figure 1: Mean Squared Error (MSE) across different social regulation values. The personalized model (PCCCluser_PCCPredict) achieves the lowest error, significantly outperforming the RealFriends baseline.

As shown in Figure 1, the personalized model (PCCCluser_PCCPredict) reaches a minimum MSE of approximately 1.28 at a social regulation value of roughly 0.09. This represents a substantial improvement over the RealFriends baseline, which remains relatively stagnant around 1.34-1.36.

Figure 2: Mean Absolute Error (MAE) comparison. Similar to MSE, the personalized approach demonstrates superior accuracy as social regulation increases.

Figure 2 illustrates a similar trend for Mean Absolute Error (MAE). The personalized model eventually surpasses both the generic prediction and the real friendship baseline, reaching its optimal performance at a higher social regulation value (approximately 0.13).

Key Insights

  • Subjectivity of Trust: The success of PMFTM confirms that different user groups weigh evidence differently when forming trust. By allowing the model to learn different weight vectors for different clusters, the system captures the diverse ways people form trust.
  • Predicted Trust vs. Explicit Friends: A pivotal finding is that predicted trust links often serve as a more powerful signal for recommendations than actual social media "friendships," which may not reflect shared tastes.

Conclusion

Personalizing trust prediction through user clustering significantly enhances the accuracy of trust-aware recommender systems. Future work will explore more sophisticated clustering algorithms and a broader array of domain-specific features to further refine these predictive models.

Summary updated

BibTeX
@inproceedings{parmentier2020personalized,
  author = {Alexandre Parmentier and Robin Cohen},
  title = {{Personalized Prediction of Trust Links in Social Networks (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i10.7218},
}
Figure