SAGRS: Beyond Individual Taste—Leveraging Social Affinity for Group Recommendations

TL;DR

When picking a movie with friends, the "loudest" person or the strongest relationship often dictates the choice over individual ratings. This paper introduces SAGRS (Social Affinity-Based Group Recommender System), a framework that models these hidden social dynamics. By constructing an affinity graph and using a novel TF-IDF weighting for movie features, the system predicts what a group will enjoy by understanding who influences whom.

Background: The Group Dynamic Gap

Standard recommender systems (like Netflix or Amazon) are built for individuals. When they try to recommend to a group, they usually just "average" everyone's taste. This fails in reality:

• Position/Power: In a couple, one partner might consistently yield to the other's preference.
• Context: Some friends share a niche interest that none of them pursue individually.
• Sparsity: Most users haven't watched movies with every member of a new group.

The authors argue that Social Affinity—the strength and direction of a relationship—is a better predictor for group satisfaction than individual historical ratings.

Methodology: Mapping the Influence

The SAGRS framework operates through a sophisticated three-stage pipeline.

1. Feature Weighting (The TF-IDF Twist)

Not all movie features (Director, Genre, Actors) are equally important to every friendship. The authors adapt the TF-IDF (Term Frequency-Inverse Document Frequency) metric. Specifically, they use a squared TF component to reward features of movies that two users have watched together.

2. The Social Affinity Graph

The system builds a directed graph where nodes are users and edges represent the "Affinity."

• Direct Affinity: Calculated if User A and User B have watched movies together.
• Indirect Affinity (The Multi-hop Concept): If James has watched a movie with John, and John has watched one with Patricia, the system can calculate an affinity between James and Patricia via "two-hop" propagation.

Note: The formula above shows the summation of direct and indirect paths (H-hops) to estimate social influence.

3. Group Aggregation

Once the affinities are mapped, the system identifies the "influencers" within the specific group requesting a recommendation. It uses a Maximum Affinity method to ensure that the most significant relationships in the group are preserved rather than "diluted" by averaging everyone's scores.

Experimental Insights

The authors validated SAGRS using a synthetic dataset of 18 users with varying levels of relationship "certainty."

Table 2: Average affinity scores across different group certainties.

The results confirmed the system's ability to distinguish established social ties (Group A) from uncertain ones (Group C), showing a clear gradient in calculated affinity (0.584 for A-to-C vs 0.416 for C-to-A). This directed nature is crucial: it captures cases where one person's taste "leads" while others "follow."

Critical Perspective

Strengths

• Cold-Start Resilience: By using indirect connections in the affinity graph, the system can make recommendations for groups that have never met before.
• Intuitive Weighting: The use of squared TF-IDF correctly identifies that shared consumption is a massive signal of social alignment.

Limitations & Future Work

• Scale: The current study only tested 18 users. Real-world social networks (like Facebook, which the authors plan to integrate) present much noisier data.
• Computational Cost: Calculating all-pairs multi-hop paths can be expensive as the number of users grows.
• User Subjectivity: The "Maximum Affinity" strategy might lead to "dictatorship" where one person's taste overrides the group; balancing this with "Fairness" remains an open challenge.

Conclusion

SAGRS represents a shift from "Recommendation as Math" to "Recommendation as Social Science." By acknowledging that we are social animals whose choices are dictated by our peers, the system moves closer to a truly "human-centric" AI.