DifRec: Rethinking Recommendations Through the Lens of Social Diffusion

TL;DR

Most recommender systems waste valuable screen real estate by showing you posts your friends were going to share with you anyway. DifRec (Social-Diffusion-Aware Recommender) solves this by predicting the path of "re-posts" across a social graph. By withholding redundant items, it frees up slots for new, relevant content, boosting total user engagement by up to 190%.

Problem & Motivation: The "Wasted Slot" Paradox

In platforms like Twitter, X, or Tumblr, users have two primary ways of discovering content:

1. The Recommender System: Explicitly suggested items.
2. The Social Feed: Items re-posted or shared by people they follow.

Traditional systems (Content-based or Social-filtering) only look at relevance. If a post is highly relevant to both User A and their follower User B, the system recommends it to both. However, if User A re-posts it, User B sees it twice. Since recommendation slots (K) are limited, this is a massive inefficiency. The authors argue that we should treat these slots as a scarce resource and maximize the unique relevant content a user encounters.

Methodology: The Neighborhoodness Graph

The core innovation of DifRec lies in how it models "conflict" between social sharing and direct recommendation.

1. The Diffusion Model

First, the system calculates a Reliable Node Set ($R_i$) for every user. Using a probabilistic graph where edges represent the likelihood of a re-post ($w_i$), the system uses Monte Carlo sampling to find which followers are likely to receive a message from User $i$.

2. The Recommendation Problem as MWIS

The authors define a Neighborhoodness Graph ($G_n$). In this graph, an edge $(i,j)$ means that if User $i$ gets a post, User $j$ will likely see it via diffusion. The goal is to maximize: ${\sum }_{i=1}^N{\sum }_{t\in S}{E}_{it}{x}_{it}$ subject to the constraint that if an edge exists between $i$ and $j$, they cannot both be recommended the same post $t$ ($x_{it}+x_{jt}\le 1$).

Figure 1: Comparison between (a) Traditional Recommendation vs (b) Diffusion-aware Recommendation.

3. Heuristic Algorithms

Since solving the Maximum-Weight Independent Set (MWIS) problem is NP-hard, DifRec uses a greedy approach. It sorts users by their "Engagement Potential" (the sum of relevance for them and all their reachable followers) divided by their degree in the neighborhoodness graph, ensuring we pick high-impact "seed" users for recommendations.

Experiments: Real-World Impact on Tumblr

The authors tested DifRec using a massive dataset of 1.5 million Tumblr users.

SOTA Comparison

DifRec was compared against the Real Tumblr recommender, a standard Content-based system, and a Social-aware system.

DifRec maintained a consistent lead, providing roughly 2.8x to 3x higher engagement scores than the production baseline.

Figure 2: Percentage of users achieving high engagement thresholds across different models.

Critical Insight & Conclusion

DifRec proves that a recommender system shouldn't just be "smart" about what a user likes; it must be "aware" of the platform's external dynamics.

Takeaways:

• Diversity via Diffusion: By avoiding social redundancy, DifRec naturally increases the diversity of content a user sees.
• Efficiency: It maximizes the value of every pixel on a mobile screen by ensuring no two slots are wasted on the same content path.
• Limitations: The model assumes a static social graph and constant re-post probabilities, which may shift in real-time. Future work involving dynamic Graph Neural Networks could make these "Reliable Sets" even more accurate.

Ultimately, DifRec moves us from Individual Relevance to Network-Wide Utility.