[IEEE Access] RaC: Accelerating Opinion-Leader Mining via Phased Clustering and Behavioral Analysis

TL;DR

The RaC (Rank after Clustering) algorithm introduces a two-stage pipeline to identify social media influencers. Instead of calculating the influence of every user, it first uses topological clustering to find a "candidate pool" and then applies a deep behavioral ranking. This approach cuts computation time significantly while maintaining SOTA-adjacent accuracy.

Problem & Motivation: The High Cost of Influence

In the era of massive social datasets like Weibo or X (Twitter), identifying "Opinion Leaders"—users who dominate information dissemination—is computationally expensive. Existing methods face a dilemma:

• Score-driven models are accurate but treat every user as a candidate, leading to $O(N^2)$ or worse complexity.
• Heuristic models are fast but often use only topological "surface" data, missing the nuances of actual user interaction and trust.

The authors' core insight is that topology acts as a gatekeeper. Most users are "common users" who physically cannot be opinion leaders due to their position in the graph. By filtering them out early, we can afford to run expensive behavioral analysis on the remaining few.

Methodology: The Two-Phase RaC Framework

Phase 1: Topological Clustering

The goal here is to reduce the search space. The authors use an improved K-means++ algorithm based on four specific features:

1. Indegree: Total followers.
2. Two-hop Clustering Coefficient: A new metric calculating how "interconnected" a user's network is within two steps. High local clustering often limits information "leakage" to the wider network.
3. Betweenness: How often a user acts as a bridge between separate communities.
4. Center: A composite index ($IndegreeimesBetweenness/Clustering$).

Phase 2: Behavioral Ranking

Once the candidates are selected (e.g., 134 candidates from ~50k users), the system calculates User Leadership ($ULD$): $ULD(u)=UI(u)imesUA(u)imesCenter(u)$

• User Activeness ($UA$): Uses the Analytic Hierarchy Process (AHP) to weight forwarding, publishing, and evaluating.
• User Influence ($UI$): Not all followers are equal. RaC weights a follower's contribution by their "Attention Degree" ($A{D}_{vu}$), essentially measuring the specialized trust between a follower and the leader.

Experiments & Results

The authors tested RaC on a Weibo dataset containing nearly 50,000 users.

Performance vs. Efficiency

The "Gold Standard" in the study, UI_LR, provided the highest influence accuracy but at a massive temporal cost. RaC achieved performance levels nearly identical to UI_LR but ran 7 times faster.

Figure: Influence comparison based on the Degree-based Linear Threshold (DLT) model. RaC consistently outperforms standard ranking (AllUserRank) and pure clustering (ClusterRank).

Critical Runtime Advantage

As seen in the table below, the "Candidate" approach is a game-changer for production environments. Calculating user activeness for 134 candidates took only ~1 second, compared to over 368 seconds for the full network.

(Note: Topology calculation remains the bottleneck, but the subsequent analysis is virtually eliminated.)

Critical Analysis & Conclusion

Takeaway

RaC proves that Efficiency vs. Accuracy is not a zero-sum game in social network mining. By using a phased approach, we can retain the "Global View" of graph theory (Phase 1) and the "Local Detail" of user behavior (Phase 2).

Limitations & Future Work

• Dynamic Graphs: The current model uses a static snapshot. Social networks are evolving; a leader today might be irrelevant tomorrow.
• Topic Sensitivity: The current RaC is topic-agnostic. However, an opinion leader in "AI Technology" might have zero influence in "Macroeconomics." The authors suggest "Topic-specific RaC" as the next research frontier.

In conclusion, RaC offers a robust blueprint for real-time public opinion monitoring and commercial influencer discovery where scale and speed are paramount.