Innovation adoption: Broadcasting versus virality

Yujia Zhai, Ying Ding, Hezhao Zhang

TL;DR

This paper introduces a novel method to analyze how innovations spread through "broadcasting" (one-to-many) and "virality" (one-to-one, cascading) channels, using academic co-authorship as a proxy for viral spread. They find both channels are crucial, with broadcasting dominating early adoption and virality becoming more influential and faster in later stages, leading to repeated adoption by those in viral networks.

What Problem Does This Paper Address?

Understanding how new ideas and technologies spread is fundamental to innovation and societal progress. While diffusion has been studied, the distinct roles and interplay of different communication channels remain unclear. This paper focuses on two key channels: broadcasting, where information is disseminated widely from a single source (like publishing an article), and virality, where adoption spreads through interpersonal connections and social contagion (like word-of-mouth or collaboration). The authors aim to quantify the diffusion power of these channels and understand how they influence adoption across different stages of an innovation's lifecycle.

Methods

The core of this research is building a diffusion tree to visualize and analyze the spread of innovations. They use academic co-authorship as a proxy for viral diffusion, treating citations of influential algorithms as adoption events.

Here's a breakdown of their approach:

  • Diffusion Tree Construction: For each algorithm (innovation), they construct a co-author vector and define adoption times based on publication dates of papers citing the algorithm. They then build a directed tree where edges represent the flow of adoption, either directly from the source (broadcasting) or from an existing adopter to a co-author (virality).
  • Measuring Diffusion: They analyze the diffusion trees along three dimensions:
  • Scale: The number of adopters (children of a node).
  • Depth: The number of layers in the tree, indicating how far an innovation has spread through cascading adoptions.
  • Speed: The time interval between a child node adopting and its parent, measuring how quickly the innovation spreads.
  • Channel Identification: Nodes directly connected to the original algorithm are classified as "broadcasting members," while those further down the tree are "virality members."

Datasets

The study uses citation data for seven classic algorithms:

  • LSA (Latent Semantic Analysis)
  • SVM (Support Vector Machines)
  • NMF (Non-negative Matrix Factorization)
  • PLSA (Probabilistic Latent Semantic Analysis)
  • LDA (Latent Dirichlet Allocation)
  • LDA sample (a 10% sample of LDA citations)
  • Word2Vec
  • Glove (Global Vectors for Word Representation)

The primary data source for citation information is the Microsoft Academic Graph (MAG), chosen for its comprehensive coverage. They also compared MAG with Web of Science and Scopus for LDA citations to validate their choice.

Key Insights and Ideas

  • Broadcasting vs. Virality: Different Roles, Equal Power: While both channels are crucial for innovation adoption, they play distinct roles. Broadcasting is powerful for initial awareness, while virality drives deeper, cascading adoption and repeated use.
  • Virality's Growing Influence: Broadcasting dominates the early stages of adoption, but virality gradually catches up and can even surpass it in later stages. This suggests that while initial exposure is important, sustained adoption and integration into research practice rely heavily on interpersonal networks.
  • Faster Adoption in Viral Networks: Individuals adopting through viral channels tend to adopt innovations more quickly and repeatedly. This is attributed to the trust and efficiency of learning from peers and collaborators who have already adopted and integrated the innovation.
  • Cross-Domain Diffusion: Cross-domain diffusion (innovation spreading to different fields) primarily occurs through the broadcasting channel and the initial stages of viral diffusion. As diffusion deepens, it tends to stay within related domains.

Key Results

  • Diffusion Scale: Broadcasting members form an initial dense ring around the innovation, while virality members extend outwards in numerous branches, indicating a broad initial reach followed by deeper, more specialized spread.
  • Diffusion Depth: Most diffusion trees reach a depth of 10-12 layers, with younger innovations like Word2Vec and Glove having shallower trees (7-8 layers). The depth growth has slowed in recent years, possibly due to innovation saturation, limited academic communities, or the emergence of newer innovations.
  • Diffusion Speed: Virality is significantly faster than broadcasting for initial adoption. As diffusion deepens, the speed generally increases, especially when both distributors and receivers are from related domains.
  • Repeated Adoption: Authors who adopt through viral channels are more likely to cite the innovation multiple times, integrating it more deeply into their research practice compared to those who adopted via broadcasting.

Limitations and Future Work

The authors acknowledge several limitations:

  • Data Coverage: The MAG dataset, while extensive, does not cover all scientific publications.
  • Innovation Scope: The study focuses solely on computer algorithms, which may limit the generalizability of findings to other types of innovations.
  • Co-authorship Proxy: Using co-authorship as the sole proxy for virality might be insufficient, as informal communication without co-authoring also contributes to knowledge sharing.

Future work will involve:

  • Exploring innovation adoption through content-based citation analysis to restore a more comprehensive diffusion process.
  • Investigating how innovations diffuse across different fields and break down inter-domain barriers.

Summary updated

BibTeX
@article{zhai2021innovation,
  author = {Yujia Zhai and Ying Ding and Hezhao Zhang},
  title = {{Innovation adoption: Broadcasting versus virality}},
  journal = {J. Assoc. Inf. Sci. Technol.},
  year = {2021},
  doi = {10.1002/asi.24420},
}
Figure