UserRank: Quantifying Peer Influence in Online Health Communities

TL;DR

In online health communities like MedHelp, not all active users are truly influential. This paper introduces UserRank, a specialized algorithm that identifies health "influencers" by looking beyond simple connection counts. It blends content relevance (what you say) and response speed (how fast you say it) to create a weighted social graph, proving far more effective than Google’s PageRank in identifying the real leaders of medical forums.

Background: Beyond the "Power Law"

Most social networks follow a "Power Law" where a tiny fraction of users generates the vast majority of content. In health settings, these "e-patients" are vital—they provide emotional support, share treatment experiences, and even influence others' medical decisions. However, identifying the right influencers is critical: while they can propagate positive health habits, they can also inadvertently spread medical misinformation.

The Problem: The Blindness of Topology

Prior methods like PageRank treat a social network like a collection of webpages. If User A replies to User B, a link is formed. The more replies you get, the higher your "authority."

The authors argue this is flawed for medicine:

• Topic Drift: A reply might be a "spam" post or a complete change of subject.
• Temporal Value: In a health crisis (like the H1N1 flu), a response that comes three weeks late is far less influential than one that arrives in minutes.

Methodology: The UserRank Formula

The core innovation is the weighting of the social graph. Instead of a binary "link or no link," every interaction between users $v_i$ and $v_j$ is assigned a weight $W(v_i,v_j)$ based on two factors:

1. Message Similarity: Using Cosine Similarity to ensure the responder is actually addressing the original post's topic.
2. Response Immediacy: An exponential decay function that gives higher weight to faster replies.

The influence is then calculated iteratively, where a user's rank is the sum of the weighted influence passed to them by others.

The weight function combining temporal decay ($\alpha$ and $\beta$) and semantic similarity.

Experiments & Results

The researchers crawled three distinct MedHelp communities: Swine Flu (small), Smoking Addiction (medium), and Alcoholism (large). They compared UserRank and PageRank against a "Gold Standard" (GS) created by human experts who manually rated user influence based on message quality.

Using the Kendall’s tau ($au$) metric—where a lower score indicates a closer match to the human "Gold Standard"—the results were clear:

As the forum size and complexity grew, UserRank became significantly more accurate. In the Alcoholism forum, PageRank's performance degraded sharply (reaching a distance of 37), while UserRank remained robust (distance of 24).

Deep Insight: Why it Matters

The success of UserRank confirms an "Inductive Bias" specific to social support: Influence is earned through relevance and presence. In a medical context, an influential user isn't just someone who "talks a lot"; they are someone who provides timely, contextually appropriate advice that resonates with the seeker's needs.

Limitations & Future Work

The authors acknowledge that they haven't yet considered:

• Message Length: Longer, more detailed responses might indicate higher effort and influence.
• Sentiment Analysis: Understanding if the influence is positive or negative (e.g., a "troll" might be influential but harmful).

Conclusion

This work provides a vital tool for health organizations. By identifying these "super-users" through UserRank, organizations can effectively harness social networks to disseminate life-saving information more efficiently than ever before.