Joint Social and Content Strategy: Decoding the Multi-Dimensionality of Video Sharing

TL;DR

Social video consumption involves two distinct behaviors: Importing (bringing external content in) and Re-sharing (spreading existing content). This paper proposes a joint framework that bridges the gap between social influence and content analysis, effectively solving the cold-start problem by simulating how videos propagate through social tiers.

Problem & Motivation: The Sparsity Trap

Most recommendation engines rely on explicit user ratings. However, in Online Social Networks (OSN), user actions like "importing" or "re-sharing" a video are implicit. This leads to a massive Sparsity Problem:

1. Cold-Start Users: New users have no history.
2. Cold-Start Videos: User-generated videos often have near-zero initial views.
3. Missing Context: Traditional Collaborative Filtering (CF) doesn't "see" how a video travels from User A to User B across a social graph.

The authors argue that to recommend effectively, we must look at the Joint Social-Content Space, recognizing that the motivation for importing a video (content interest) differs fundamentally from re-sharing one (social influence).

Methodology: The Two-Step Evolution

The framework operates on three primary matrices: User-User (A), Content-Content (C), and User-Content (B).

Phase 1: Matrix Update via Propagation

Before recommending, the system "pre-fills" the empty spots in the User-Content matrix.

• Social Propagation: If your friends are sharing a video, the system predicts you might too.
• Content similarity: If you liked "Action Video A," you are likely to interact with "Action Video B." By simulating these connections over multiple "rounds," the system generates a much denser foundation for recommendation.

Phase 2: Building the Joint Relevance Space

The system maps users and videos into a lower-dimensional space using Representative Items.

• User Space: Users are grouped based on the "celebrities" (top-followed users) they follow.
• Content Space: Videos are grouped by tag-based similarities.

The final relevance score is a weighted sum: $RE(u,c)=\alpha \cdot (extUserSpaceRelevance)+(1-\alpha )\cdot (extContentSpaceRelevance)$

Experiments: Why "Why" Matters

The most striking finding is the role of the weight parameter $\alpha$.

• Importing ($\alpha =0.3$): Users act as "curators." They care more about the content's intrinsic quality.
• Re-sharing ($\alpha =0.7$): Users act as "distributors." They are heavily influenced by the social identity of the person who shared it first.

As shown in the charts, ignoring either the social or the content dimension results in a significant drop in accuracy. The "Joint" approach (the peak of the curves) consistently outranks single-modality baselines.

Deep Insight & Conclusion

This work highlights that social activity is not a monolith. By treating "Importing" and "Re-sharing" as mathematically distinct tasks, the authors provide a blueprint for more nuanced social media algorithms.

Limitations: The model relies on tag-based keyword analysis for content similarity, which can be noisy or sparse compared to modern deep-learning-based visual embeddings (e.g., CLIP).

Future Outlook: Integrating 3D tensors to include temporal and geographical data could further refine these predictions, especially for localized viral content.