Explicit vs. Implicit Tagging for User Modeling

Enrique Frías-Martínez, Manuel Cebrián, J. Moises Pascual, Nuria OliverPMPC@UMAP 2009 · OpenReview

Introduction

The advent of Web 2.0 established explicit tagging—the manual assignment of keywords to digital content—as a primary mechanism for information organization. While powerful, explicit tagging is often limited by user inconsistency and the cognitive load of manual labeling. This research investigates "implicit tagging," a method that derives user interests from behavioral data, such as mobile phone usage. By comparing the statistical properties of implicit and explicit systems, we can determine if behavioral patterns mirror the well-studied dynamics of social tagging, enabling the application of established web-based user modeling techniques to physical-world interactions.

Methodology and Data Sources

The study conducts a comparative analysis between two distinct tagging environments:

  • Explicit System (Bibsonomy): A social bookmarking dataset comprising 816,197 entries. After filtering, the analysis focused on 2,467 unique users and 69,902 unique tags.
  • Implicit System (SME): A behavioral dataset from a major mobile carrier involving approximately 3,000 users. Tags were automatically generated by concatenating business attributes from call logs (e.g., Hotel-4Stars-Madrid).

Statistical distributions were analyzed using Clauset’s power-law fitting algorithm to evaluate tag frequency, the number of tags per user, and the resulting interest-based social networks.

Comparative Statistical Analysis

Tag Frequency Distributions

A fundamental characteristic of tagging systems is the "long tail" distribution, where a small number of tags are used extensively while the majority appear infrequently. Despite the infinite vocabulary available in explicit systems, users naturally gravitate toward popular terms, a behavior known as preferential attachment. The analysis revealed that both systems follow a power-law distribution with remarkably similar scaling parameters ($\alpha \approx 1.75$ for explicit and $\alpha \approx 1.8$ for implicit).

Figure 1: Tag frequency distributions for explicit (Bibsonomy) and implicit (SME) tagging systems. Both follow a power-law distribution, indicating a "rich-get-richer" effect in both manual and behavioral tagging.

User Tagging Complexity

The distribution of tags per user highlights differences in user behavior between the two systems. The implicit SME system follows a straightforward power law ($\alpha = 2.93$). However, the explicit Bibsonomy system exhibits greater complexity; while the "tail" of the distribution follows a power law ($\alpha = 2.13$), the "head" (representing common user behavior) is more accurately described by a lognormal distribution.

Figure 2: Distribution of tags per user. While the implicit SME system follows a simple power law, the explicit Bibsonomy system exhibits a more complex behavior best captured by a lognormal fit for the head and a power law for the tail.

Network Architecture and Role Inversion

The study also examined the structure of social networks formed by shared interests. A notable "inversion of roles" was observed: in the explicit web network, users act as the primary connectors between tags, whereas in the implicit mobile network, the tags themselves (the interest nodes) serve as the central hubs connecting users. Despite this structural difference, both interest-based networks share a similar architecture, with degree distributions best fitted by power laws ($\alpha = 2.37$ for Bibsonomy and $\alpha = 1.83$ for SME).

Conclusion and Future Directions

The findings suggest that implicit behavioral data captures user interests with statistical signatures nearly identical to those of manual tagging systems. This validates the use of behavioral logs for automated user modeling. Future research will explore the impact of vocabulary size on these distributions and investigate the applicability of Double Pareto log-normal models for more precise data fitting. These insights will support the development of mobile applications that leverage implicit models to provide personalized services and recommendations.

Summary updated

BibTeX
@inproceedings{friasmartinez2009explicit,
  author = {Enrique Frías-Martínez and Manuel Cebrián and J. Moises Pascual and Nuria Oliver},
  title = {{Explicit vs. Implicit Tagging for User Modeling}},
  booktitle = {PMPC@UMAP},
  year = {2009},
  url = {https://openreview.net/forum?id=WFnFkkr6NS},
}
Figure