Introduction
How do groundbreaking ideas actually take over a field? While innovation is the lifeblood of scientific progress, it doesn't spread instantly. On average, it takes nearly 39 years for a major innovation to reach widespread adoption. Researchers have long sought to understand why some ideas "catch fire" while others languish in obscurity.
This study analyzes the diffusion of seven classic machine learning algorithms to quantify the tension between two primary communication channels:
- Broadcasting: Information spreading from a single source to many (e.g., reading a paper in an open-access journal).
- Virality: Information spreading through interpersonal cascades (e.g., a mentor teaching a student a new technique).
Methodology: The Diffusion Tree
To measure these channels, the researchers proposed a novel way to reconstruct the diffusion trajectory of an idea by building a Diffusion Tree. They used co-authorship as a proxy for virality.
Under this model, if an author cites an algorithm and none of their previous co-authors have ever cited it, that author is classified as a Broadcasting Member. Conversely, if an author cites an algorithm after one of their co-authors has already used it, they are a Virality Member. By mapping these relationships over time, the study calculated the scale (total adopters), depth (generations of collaborators), and speed of diffusion.
Datasets and Scope
The study focused on seven "superstar" algorithms in computer science with clear, traceable citation histories. Data was primarily sourced from the Microsoft Academic Graph (MAG), which proved significantly more comprehensive for this task than traditional databases like Web of Science or Scopus.
As shown in the figure below, the MAG database provided a much larger pool of data for analysis, particularly when looking at the "citations of citations" which are crucial for mapping deep diffusion trees.
Key Insights: Broadcasting vs. Virality
- Broadcasting is the Spark, Virality is the Fuel: Broadcasting is more powerful in the early stages of an innovation's life. However, as a paper ages and becomes harder to find in the "noise" of new publications, virality takes over to sustain long-term growth.
- Virality is Faster and Stickier: Peer-to-peer diffusion is significantly faster than independent discovery. Learning from a collaborator reduces the perceived technical barrier of a new method. Furthermore, researchers who learn an algorithm through peers are more likely to use it repeatedly in future work.
- The "Soil" Matters for Depth: Deep diffusion (cascades reaching 10+ layers) almost exclusively happens within the innovation's home domain (Computer Science). When an idea jumps to a distant field (like History or Art), the cascades remain shallow, usually only 1-4 layers deep.
Results and the "10/5 Rule"
The researchers identified a consistent growth pattern across the studied algorithms, which they termed the 10/5 Rule: roughly 10% of adopters in any given layer become "active" (passing the idea to others), and each active member typically recruits 5 new adopters for the next layer.
While all seven algorithms followed the classic "S-curve" of technology adoption, newer algorithms like Word2Vec and GloVe reached their "take-off" stage much faster than older ones like LSA, suggesting an acceleration in the speed of scientific communication.
Limitations and Future Outlook
While co-authorship is a strong signal, it does not capture informal conversations at conferences or social media interactions. Additionally, the "10/5" rule was derived from computer science; further research is needed to see if these dynamics hold in fields like biology or the social sciences. Future work will involve content-based citation analysis to distinguish between those who use an algorithm as a tool versus those who seek to improve the algorithm itself.
Summary updated