SocialAU: Unmasking Real Influence Through Multilayer Tensor Networks

TL;DR

The "million follower fallacy" suggests that popularity isn't influence. This paper introduces SocialAU, a method that looks beyond the follower count. By constructing a three-layer network (Users, Items, Keywords) and using advanced multilinear algebra (tensor decomposition), it identifies users who are not only prominent in the social graph but are also authoritative voices on specific topics.

Problem & Motivation: The Limits of Topology

Most influencer detection tools are "blind" to content. They treat social networks as simple graphs where an edge only signifies a "follow" or "friend" request. However, real influence is topical. You might follow a celebrity for entertainment but listen to a specific tech reviewer for smartphone advice.

The authors argue that existing methods like PageRank or HITS fail because:

1. They ignore the semantic context of posts.
2. They don't distinguish between a user who is "loud" (posting a lot) and one who is "authoritative" (being cited/retweeted by others).

Methodology: The Three-Layer Tensor Approach

The genius of SocialAU lies in its representation of social data. Instead of a flat graph, it uses a Multilayer Network:

• User Layer: Reflects interpersonal relations (mentions, retweets).
• Item Layer: Contains the subjects of discussion (e.g., specific TV shows or products).
• Keyword Layer: Captures the sentiment and vocabulary used to describe items.

Mathematically Fusing Context and Topology

Information is modeled as a 3rd-order tensor $\mathcal{X}$, where an entry $x_{uik}$ exists if user $u$ talked about item $i$ using keyword $k$.

The authors extend the TOPHITS algorithm. While standard TOPHITS finds dominant patterns in the tensor, SocialAU injects HITS scores from the individual layers back into the iteration. This ensures that a "dominant user" isn't just someone who spam-tweets keywords, but someone who also holds a high authority score within the user-to-user interaction graph.

Experiments: Real-World Validation

The researchers tested SocialAU on two distinct datasets:

1. Twitter (TV Series): Analyzing over 20,000 tweets about shows like The Walking Dead and NCIS.
2. Yelp: Using reviews and categories as a proxy for social influence.

Key Result: Quality over Quantity

In the Twitter experiment, SocialAU identified the official @ncis_cbs account as the top influencer. Standard tensor methods (like TOPHITS) ranked it much lower because it didn't participate in as many "triples," failing to recognize its massive authority (1M+ followers and high retweet rates).

Scalability

One of the most impressive feats is the efficiency. Using the MATLAB Tensor Toolbox, the algorithm handles massive sparse tensors with $10^{15}$ potential nodes in under 10 seconds. The computation time scales linearly with the number of actual interactions (triples), making it viable for production-scale social listening.

Depth Insight: Resilience to Bots

A fascinating byproduct of this approach is Bot Detection. Bots and "fraudsters" often have high activity (triples) but low authority in the User Layer (no one meaningful retweets them). Because SocialAU balances these two factors, bots naturally sink to the bottom of the rankings, while authentic authoritative voices rise.

Conclusion

SocialAU moves social media analysis from "Who is big?" to "Who is authoritative?". By bridging the gap between graph topology and natural language content through tensor decomposition, it provides a sophisticated tool for marketers and researchers to find the real needles in the social media haystack.

Takeaway: Effective influencer marketing requires a multidimensional view. Don't just count the followers; count the authoritative echoes of their opinions across the keyword and item space.