Self-organization in social tagging systems

Chuang Liu, Chi Ho Yeung, Zi-Ke ZhangarXiv 2011 · arXiv

Abstract

Social tagging systems, or folksonomies, are fundamental to modern web organization. This paper investigates the human dynamics driving these systems using a generative model based on peer imitation. We introduce "tagivity" ($p_i$) as a measure of a user's propensity to add tags and a confidence parameter $eta$ representing individual stubbornness. Our findings reveal a sharp phase transition at $eta = 0.5$: systems with low user confidence ($eta < 0.5$) self-organize into stable "Active Tagging" states, whereas high confidence leads to "Inactive Tagging" and eventual system collapse. The model is validated against empirical data from Delicious and Flickr, demonstrating that imitation is a key driver of real-world tagging behavior.

Introduction

Folksonomies, such as those found on Flickr or Delicious, allow users to collaboratively categorize resources using freely chosen keywords. While the mathematical properties of the resulting tripartite graphs are well-documented, the behavioral mechanisms that determine how many tags a user assigns remain poorly understood. This research addresses whether individual tagging habits are formed in isolation or through the subconscious mimicry of community norms.

The Model: Tagivity and Imitation

The authors propose a model where each user $i$ possesses a "tagivity" $p_i$, defined as the probability of continuing to add tags to a resource. The dynamics are governed by three main components:

  1. Observation: Users monitor the average tagivity of the community.
  2. Imitation: Users randomly sample posts. If a sampled post has more tags than their current habit, they increase their tagivity; otherwise, they decrease it.
  3. Confidence ($\beta$): This parameter scales from 0 (total imitation) to 1 (total stubbornness).

The study compares two update mechanisms: Linear (fixed increments/decrements) and Multiplicative (scaling behavior).

Results: The Phase Transition

The most significant theoretical finding is the existence of a critical threshold at $\beta = 0.5$. As shown in Figure 4, the system's steady-state behavior changes abruptly at this point.

Phase transition of average tagivity $\langle p \rangle$ vs confidence $\beta$.

In the "Active Tagging" regime ($\beta < 0.5$), users are sufficiently open to influence that the community converges on a high level of activity. Conversely, in the "Inactive Tagging" regime ($\beta > 0.5$), user stubbornness prevents the formation of a collective norm, causing tagging activity to dwindle. Near the critical point, the system exhibits "critical slowing down," where the time required to reach a steady state increases dramatically.

Empirical Validation

To test the model's real-world applicability, the authors analyzed datasets from Delicious and Flickr. The goal was to see if the predicted distribution of "post lengths" (tags per resource) matched observed data.

Empirical post-length distributions for Delicious and Flickr.

As illustrated in Figure 5, the model—specifically using the linear update rule—shows excellent agreement with empirical data. Both Delicious and Flickr exhibit power-law decay in post lengths with exponents of 4.1 and 4.3, respectively. This suggests that real-world tagging systems naturally operate in the low-confidence, high-imitation regime ($\beta < 0.5$), which is necessary for maintaining a vibrant, tag-rich environment.

Conclusion

This research demonstrates that social tagging is a self-organizing process driven by imitation. The discovery of a phase transition at $\beta = 0.5$ provides a theoretical basis for understanding why some online communities thrive while others fail to generate significant metadata. These insights could be leveraged to improve recommender systems by accounting for user susceptibility to social proof.

Summary updated

BibTeX
@misc{liu2011self,
  author = {Chuang Liu and Chi Ho Yeung and Zi-Ke Zhang},
  title = {{Self-organization in social tagging systems}},
  year = {2011},
  eprint = {1102.3989},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1102.3989},
}
Figure