Measuring Influence & Information Dissemination in a Fashion Influencer Network

Overview

A team project (with Ka Woon Ho, Xia Mo, and Haotian Shi) modeling a network of Instagram fashion influencers as a graph, to study who is actually most "influential" and how quickly information spreads from different starting points in the network.

Motivation

Influence on social media isn't just follower count — some accounts bridge otherwise-disconnected communities, while others sit inside tightly-connected clusters. We wanted to model a real influencer network and test whether the most popular accounts are actually the fastest way to spread information, or whether structural position in the network matters more.

Approach

We modeled the network as a bidirectional graph, with each vertex representing an influencer's account (drawn from a public Instagram fashion-influencer dataset, downsampled to the top 500 most-followed accounts for tractable experiments). Rather than modeling explicit follow relationships — which the available dataset didn't capture — we connected influencers by a similarity score based on the difference in their follower, followee, and post counts, on the assumption that similarly-sized accounts are more likely to share information with each other directly. Each influencer is connected to its k most similar neighbors.

We computed two influence metrics per account — popularity (follower count) and degree (number of edges) — normalized and ranked with a merge sort. To simulate information spread, we ran a BFS-based diffusion model from four different starting points: the highest-popularity influencer, the lowest-popularity influencer, the highest-degree influencer, and a medium-popularity influencer, recording what percentage of the network had been "informed" at each time step.

Implementation

The graph and its algorithms were built from scratch: a data_loader module parses the CSV dataset into vertices with follower/followee/post metadata; build_influence_edges() computes pairwise similarity scores and wires up each vertex's k nearest neighbors; popularity_scores/degree_scores compute and normalize the two influence metrics; a BFS shortest_path() method finds shortest paths between any two accounts; and a recursive spread_over_time simulation models diffusion layer-by-layer until the whole reachable network is informed. Results are visualized with Matplotlib (distributions, diffusion curves) and both static and interactive (Plotly) NetworkX graphs.

Results

• Starting from the highest-degree influencer reached the entire network in 63 steps — fast and steady.
• Starting from the least popular influencer was nearly identical (64 steps), since it sat close to the highest-degree hub.
• Starting from the most popular influencer was the slowest (66 steps): it started almost flat, because highly popular accounts turned out to have very few edges — they aren't similar to many other accounts.
• Starting from a medium-popularity influencer was the fastest of all (49 steps) — it began slowly but accelerated sharply once the diffusion reached the network's well-connected core.
• The shortest path between the single least- and most-popular influencer in the 500-node network was 65 steps.

The headline finding: raw popularity is a poor predictor of how fast an account can spread information. Degree (structural connectivity) mattered far more than follower count.

Limitations

We originally planned to build the graph from real follow relationships, but couldn't obtain a dataset containing them, so edges are based on similarity in account statistics instead. That's a reasonable proxy for "likely to share information," but it doesn't capture actual one-to-many follower dissemination (e.g., an influencer's story being seen and reshared by their followers).

Ethical Considerations

If popularity isn't what drives fast information spread, the same mechanism that helps information travel quickly can just as easily accelerate misinformation — and a moderately-popular, well-connected account may be a more effective vector for it than a single celebrity account. Combined with engagement-based ranking algorithms on real platforms, this is worth taking seriously.

Future Improvements

• Model how platform recommendation algorithms (e.g. engagement-based ranking) interact with organic diffusion — a decision tree could approximate that boosting behavior
• TODO: add source code link