Abstract
Opinion leaders are authoritative figures with significant influence in Bulletin Board Systems (BBS). Their linguistic behavior profoundly impacts the thoughts and actions of other internet users. This paper proposes the Opinion Leader PageRank Algorithm (OLRA), a topic-field-based approach to identifying these leaders. OLRA incorporates closeness degree factors and sentiment analysis, defining these as weights for links between users. Using data from the Tianya forum, the algorithm was compared against Interest-based PageRank, online time, and experience-based algorithms. Results indicate that OLRA more effectively identifies opinion leaders in internet forums.
Introduction
In the ecosystem of internet forums and Bulletin Board Systems (BBS), a small minority of users—opinion leaders—wield disproportionate influence over public consensus. Identifying these individuals is crucial for understanding social trends and managing public opinion. However, traditional identification methods often fall short. Simple metrics like "online time" or "number of posts" are easily manipulated by "Internet Water Armies" (paid ghostwriters). Furthermore, many existing algorithms treat all interactions as equal, ignoring the sentiment of replies and the topic-specific nature of influence. An expert in international news may have little sway in a discussion about cinema.
Methodology: The OLRA Framework
To address these limitations, we propose the Opinion Leader PageRank Algorithm (OLRA). This multi-layered approach refines the classic PageRank formula by integrating three key factors:
- Topic Clustering: Using fuzzy clustering and keyword extraction (via the ICTCLAS system), forum posts are categorized into specific topics. This ensures that influence is measured within a relevant context rather than as a generic popularity score.
- Emotional Attribute Analysis: Instead of merely counting replies, OLRA performs sentiment analysis on interactions. Replies are assigned values: positive (+1), neutral (0), or negative (-1). An author’s authority is weighted by the average emotional tendency of their followers, effectively penalizing "notoriety" or trolling.
- Closeness Degree: Borrowing from graph theory, the algorithm calculates a user's centrality. A high closeness degree indicates a user can disseminate information across the network more efficiently through fewer intermediaries.
These factors are integrated into a modified PageRank formula where the "weight" of a link between users is a combination of follower sentiment and the author's structural closeness.
Experimental Dataset
The study utilized real-world data crawled from the Tianya forum, a prominent Chinese online community. We focused on the "Livelihood" module, collecting posts from July 20th to July 30th, 2014. This period provided a rich dataset of trending real-world topics, such as the "MH17 crash" and "Chinese officialdom anti-corruption."
Results and Discussion
The researchers compared OLRA against three benchmarks: Interest-based PageRank, Online Time, and Experience-based algorithms. Key findings include:
- Topic-Specific Influence: Data confirmed that influence does not always transfer across domains. A top leader in one topic (e.g., MH17) might rank significantly lower in another (e.g., anti-corruption).
- Superior Identification: OLRA was more effective at identifying the top 10 opinion leaders for specific topics. While Interest-based PageRank and OLRA performed similarly for the "Top 10," OLRA showed a significant lead in identifying the "Top 7" influencers, suggesting higher sensitivity to the most elite tier of leaders.
- Sentiment as a Filter: By incorporating sentiment, the algorithm successfully filters out controversial figures who receive high engagement but lack genuine leadership or positive influence.
- In-degree Limitations: The results demonstrated that "in-degree" (simple reply counts) has negligible effect on identification accuracy compared to the nuanced weights of sentiment and closeness.
Conclusion and Future Work
While OLRA proves effective, it faces scaling challenges. As forum data grows, the computational complexity of calculating closeness and fuzzy clustering increases. Future work will focus on optimizing the algorithm for larger datasets and exploring distributed computing platforms like Hadoop to enable real-time identification of opinion leaders across massive social networks.
Summary updated