Patterns in Interactive Tagging Networks

Yuto Yamaguchi, Mitsuo Yoshida, Christos Faloutsos, Hiroyuki Kitagawa

Abstract

How do users behave if they can tag each other in social networks? In this paper, we answer this question by studying the interactive tagging network constructed by Twitter lists. Twitter lists can be regarded as the tagging process; a user (i.e., tagger) creates a list with a name (i.e., tag) and adds other users (i.e., tagged users) into the list. This tagging network is by nature different from the resource tagging networks (e.g., Flickr and Delicious) because users on this network can tag each other. We address the following research questions: (RQ1) What is the common patterns and the difference between the interactive tagging network and the resource tagging networks? (RQ2) Do users tag each other on the interactive tagging network? And if so, to what extent? (RQ3) What is the difference between the two types of relationships on Twitter: who-tags-whom and who-follows-whom? By quantitatively studying million-scale networks, we found the pervasive patterns across the different tagging networks, and the interactive patterns within the interactive tagging network. This study sheds light on the underlying characteristics of the interactive tagging network, which is relevant to the social scientists and the system designers of the tagging systems.

Introduction: From Resources to People

Most research on social tagging has historically focused on "resource tagging" networks like Flickr (photos) or Delicious (bookmarks). In these systems, the relationship is bipartite: users apply tags to digital resources, but those resources do not interact back.

Twitter lists represent a fundamental shift toward an interactive tagging network. When a user adds another to a list named "Tech-Leaders," they are effectively tagging that person. Because both the tagger and the tagged are users, this creates a unipartite graph where social behaviors—such as reciprocity (tagging someone back) and following—become central dynamics.

Conceptual comparison between (a) resource tagging networks (bipartite) and (b) interactive tagging networks (unipartite).

Methods and Datasets

The authors modeled tagging as a directed multigraph $G_T = (U, R, T; E)$, where users ($U$) tag resources ($R$) with specific tags ($T$). In the case of Twitter, $U$ and $R$ are the same set of entities. The study utilized four million-scale datasets:

  1. Twitter (Tagging): ~7M nodes and 20M edges from Twitter lists.
  2. Twitter (Following): ~7M nodes and 432M edges representing the "who-follows-whom" graph.
  3. Flickr & Delicious: 10% samples used as baselines for traditional resource tagging.

Pervasive Patterns and Multiplicity

A key finding is that while all tagging networks follow pervasive power-law patterns regarding user behavior, their "multiplicity"—the number of times a user tags the same resource—differs significantly.

Multiplicity distributions across Twitter, Flickr, and Delicious, highlighting the plateau in resource tagging networks.

As shown in the figure above, Flickr and Delicious exhibit a "plateau" in their multiplicity frequency, indicating a mix of "describers" (who use many tags for one resource) and "categorizers." Twitter users, however, are almost exclusively categorizers, typically tagging a user once to sort them into a specific bucket.

Social Dynamics: Reciprocity and the Follow Gap

In interactive tagging, the probability of reciprocity (being tagged back) increases with multiplicity, but only up to a "reciprocity peak" (around 20 tags), after which it drops sharply. This "cliff" helps distinguish genuine social interaction from automated or spam behavior.

Furthermore, the study revealed a "Tag-Follow Gap." Approximately 40% of tagging relationships on Twitter do not exist in the follow graph. This suggests that tagging provides a unique layer of metadata—contextual categorization—that the social "follow" graph alone cannot capture.

Conclusion

This research demonstrates that while the fundamental "physics" of tagging (power laws) remains consistent across domains, interactive networks introduce unique social layers. By distinguishing between "Friendship tags" (high reciprocity) and "Subscription tags" (low reciprocity), system designers can better understand information flow and user intent in modern social systems.

Summary updated

Abstract

How do users behave if they can tag each other in social networks?In this paper, we answer this question by studying the interactive tagging network constructed by Twitter lists. Twitter lists can be regarded as the tagging process; a user (i.e., tagger) creates a list with a name (i.e., tag) and adds other users (i.e., tagged users) into the list. This tagging network is by nature different from the resource tagging networks (e.g., Flickr and Delicious) because users on this network can tag each other. We address the following research questions: (RQ1) What is the common patterns and the difference between the interactive tagging network and the resource tagging networks? (RQ2) Do users tag each other on the interactive tagging network? And if so, to what extent? (RQ3) What is the difference between the two types of relationships on Twitter: who-tags-whom and who-follows-whom? By quantitatively studying million-scale networks, we found the pervasive patterns across the different tagging networks, and the interactive patterns within the interactive tagging network. This study sheds light on the underlying characteristics of the interactive tagging network, which is relevant to the social scientists and the system designers of the tagging systems.
BibTeX
@inproceedings{yamaguchi2021patterns,
  author = {Yuto Yamaguchi and Mitsuo Yoshida and Christos Faloutsos and Hiroyuki Kitagawa},
  title = {{Patterns in Interactive Tagging Networks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/icwsm.v9i1.14616},
}
Figure