C2P: Bridging the Gap in Multilayer Social Networks with Joint Recommendations

TL;DR

Modern mobile social networks are rarely flat; they consist of multiple layers of interaction (e.g., replying vs. mentioning on Twitter). This paper presents C2P (Cross-layer 2-hop Path), a novel algorithm that identifies potential friends by traversing "hybrid" paths across these layers. The result is a recommendation engine that is up to 32% more accurate than single-layer methods while maintaining high diversity and low computational overhead.

Problem & Motivation: The Silo Effect in Social Discovery

In an academic network, are you more likely to connect with someone because you co-authored a paper, or because you share a niche research interest? The answer is often both. However, most recommendation algorithms suffer from a "silo effect," analyzing only one type of connection at a time. This results in:

• Information Loss: Ignoring potential connections visible only in secondary layers.
• Poor Accuracy: Failing to capture the multifaceted nature of human friend-seeking behavior.
• Algorithmic Bias: Over-recommending popular "celebrity" nodes while ignoring relevant niche peers.

The authors argue that a user's demand is a product of their presence across all layers. For instance, a user active in "Layer A" might actually be looking for experts in "Layer B" to complement their profile.

Methodology: The Power of the "Hybrid" Hop

The core innovation of this paper is the Cross-layer 2-hop path. Instead of looking for a "friend of a friend" in the same layer (e.g., Author-Paper-Author), C2P looks for neighbors reachable through different layers (e.g., Author-Paper-Relay-Topic-Author).

Mathematical Intuition

The algorithm assigns weights to nodes based on the product of edge weights along these cross-layer paths. A key breakthrough is the User Demand Model, which balances a node's popularity (weighted degree) against the source user's specific activity levels.

The probability of recommending node $u$ to user $v$ is modeled as: $\mathbb{P}(R(v)=u)\propto \frac{{\tilde{D}}_{v1}{\tilde{D}}_{u2}+{\tilde{D}}_{v2}{\tilde{D}}_{u1}}{2|E_2|{\tilde{D}}_{v1}+2|E_1|{\tilde{D}}_{v2}}$ This formula ensures that the recommendation is sensitive to the relative importance of each layer for that specific user.

Experiments & Results: SOTA Performance

The researchers tested C2P against several baselines (FOF-L1, FOF-L2, and a simple Mixture model) using synthetic data and massive real-world datasets, including a Twitter dataset and the Microsoft Academic Graph (millions of nodes).

1. Accuracy Gain

C2P consistently outperformed baselines. In the Computer Vision (CV) dataset, the accuracy gain reached a staggering 32%.

2. Diversity & Scalability

Unlike many high-accuracy algorithms that fall into the "popularity trap," C2P maintains high Diversity. By utilizing cross-layer randomness, it avoids recommending the same few "super-nodes" to everyone. Furthermore, the algorithm is highly efficient—the complexity per recommendation stays at $O(M\log M)$ (where $M$ is the number of local neighbors), making it viable for platforms with millions of users.

Critical Analysis & Conclusion

Takeaway

The C2P algorithm serves as a bridge for MSNs that are naturally multilayered. It proves that joint information is not just "more data," but a qualitatively different way to understand social proximity.

Limitations & Future Work

• Layer Scalability: The current proof focuses largely on 2-layer systems. Extending this to $N$-layers without exponentially increasing the path length (which would dilute recommendation quality) remains a challenge.
• Interdependent Networks: The paper assumes nodes are shared across layers, but edges are independent. Investigating interdependent layers (where an edge in one layer forces an edge in another) is a promising next step.

In conclusion, C2P offers a mathematically rigorous yet practically efficient solution for the next generation of social recommendation engines, turning the complexity of multilayer networks into a distinct advantage for user discovery.