Navigational Efficiency of Broad vs. Narrow Folksonomies Denis Helic Christian Körner Michael Granitzer Knowledge Management Knowledge Management Chair of Media Informatics Institute Institute University of Passau Graz University of Technology Graz University of Technology Passau, Germany Graz, Austria Graz, Austria Michael.Granitzer@uni- [email protected] [email protected] passau.de Markus Strohmaier Christoph Trattner Knowledge Management Institute for Information Institute and Know-Center Systems and Computer Media Graz University of Technology Graz University of Technology Graz, Austria Graz, Austria [email protected] [email protected] ABSTRACT Keywords Although many social tagging systems share a common tri- Navigation, Folksonomy, Keywords, Tags partite graph structure, the collaborative processes that are generating these structures can di ffer significantly. For ex- 1. INTRODUCTION ample, while resources on Delicious are usually tagged by all In social tagging systems, users organize information us- users who bookmark the web page cnn.com , photos on Flickr ing so-called tags – a set of freely chosen words or concepts are usually tagged just by a single user who uploads the – to annotate various resources such as web pages on Deli- photo. In the literature, this distinction has been described cious, photos on Flickr, or scientific articles on BibSonomy. as a distinction between broad vs. narrow folksonomies . In addition to using tagging systems for personal organiza- This paper sets out to explore navigational di fferences be- tion of information, users can also socially share their an- tween broad and narrow folksonomies in social hypertextual notations with each other. The information structure that systems. We study both kinds of folksonomies on a dataset emerges through such processes has been typically described provided by Mendeley - a collaborative platform where users 1 as “folksonomies ” ( fol k-generated ta xonomies ). Usually, can annotate and organize scientific articles with tags. Our such folksonomies are represented as tripartite graphs with experiments suggest that broad folksonomies are more use- hyper edges. These structures contain three finite, disjoint ful for navigation, and that the collaborative processes that sets which are 1) a set of users u ∈ U, 2) a set of resources are generating folksonomies matter qualitatively. Our find- r ∈ R and 3) a set of tags t ∈ T annotating resources ings are relevant for system designers and engineers aiming R. A folksonomy as a whole is defined as the annotations to improve the navigability of social tagging systems. F ⊆ U × T × R (cf. [26]). A bookmark or post refers to a single resource r and all corresponding tags t of a user u. Categories and Subject Descriptors Although this tripartite structure of folksonomies can be mapped onto a broad range of di fferent systems in hetero- H.5.4 [ Information Interfaces and Presentation ]: Hy- geneous domains (such as Delicious, Flickr, Mendeley and pertext/ Hypermedia—Navigation; H.5.3 [ Information In- others), the collaborative processes that are generating these terfaces and Presentation ]: [Group and Organization In- structures can di ffer significantly . For example: While re- terfaces—Collaborative computing] sources on Delicious are usually tagged by a larger group of users (e.g. by everybody who has bookmarked the web page cnn.com ), photos on Flickr are usually tagged just by a sin- General Terms gle user (e.g. just by the user who has uploaded the photo). In past discussions, this distinction has been described as a Experimentation, Measurement, Algorithms 2 distinction between broad vs. narrow folksonomies . Thus, while broad folksonomies are structures that have been generated as a result of aggregating data from many people tagging the same resource , narrow folksonomies are Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are structures that have been generated as a result of aggre- not made or distributed for profit or commercial advantage and that copies gating data from single users tagging their own resources . bear this notice and the full citation on the first page. To copy otherwise, to Although both kinds of folksonomies can be mapped onto republish, to post on servers or to redistribute to lists, requires prior specific 1 http://www.vanderwal.net/folksonomy.html permission and/or a fee. 2 HT’12, June 25–28, 2012, Milwaukee, Wisconsin, USA. http://personalinfocloud.com/2005/02/explaining_ Copyright 2012 ACM 978-1-4503-1335-3/12/06 ...$10.00. and_.html the tripartite structure of folksonomies, it is reasonable to [12]. The overall outcome of our investigations allows us expect that they di ffer with regard to their overall network to shed light on the di fferences between broad vs. narrow characteristics and topology, form and function. In this pa- folksonomies in theoretical but also in practical navigation per we will argue that without thorough investigations of settings (by considering UI constraints). For our simulations the di fferent characteristics of di fferent kinds of folksonomies we use a dataset that currently includes about 150 million (e.g. broad vs. narrow), our understanding of the poten- scientific articles and has a community of about 1,5 million tials and limitations of social tagging systems will be lim- of users who tag articles in an unconstrained manner. ited. Therefore, understanding the usefulness and utility Our results suggest that both broad (tag-based) and nar- of di fferent kinds of folksonomies for di fferent tasks - such row (keyword-based) folksonomies support e fficient naviga- as navigation, emergent semantics or information retrieval - tion in theory. However, taking some practical limitations represents a problem of both theoretical and practical im- of typical user interfaces into account, we find that broad portance. folksonomies outperform narrow folksonomies significantly Similar classifications of metadata have been analyzed in on our dataset. other application areas such as learning objects metadata. In summary, this paper reports on the following findings In their analysis in [29] the authors distinguish between “au- based on our dataset: thoritative” metadata that is provided by o fficial data de- • Narrow folksonomies create less e ffective navigational struc- scriptors, e.g. learning object authors and“non-authoritative” tures than broad folksonomies when real-world user inter- metadata which emerges through the usage of learning ob- face constraints are considered. jects in di fferent contexts, e.g. it is created by a user com- • Our analysis suggests that navigational e ffectiveness of munity. In our terminology “authoritative” metadata corre- tags comes from the di fferent viewpoints of readers pro- sponds to narrow folksonomies and“non-authoritative”meta- vided through tagging resources. data to broad folksonomies. The authors argued in their • Broad folksonomies provide substantially higher quality of study that there are significant di fferences in the utility navigational structures than narrow ones. We speculate of di fferent types of metadata. For example, they demon- that with growing numbers of tags in broad folksonomies, strated that the “non-authoritative” metadata is crucial for their navigational advantage becomes even greater. More effective discovery and reuse of learning objects in di fferent research on this question is warranted though. contexts. In this paper, we aim to systematically compare di ffer- The remainder of this paper is organized as follows. In ences between broad and narrow folksonomies on a large Section 2, we discuss related work. In Section 3 we shortly tagging system (Mendeley). Mendeley is a collaborative present our simulation model for user navigation. In Sec- platform for scientists where users can annotate and orga- tion 4, we outline our experimental setup and in Section 5 nize scientific articles with tags. Because Mendeley not only we present our experimental results. In Section 6 we dis- captures data about the set of tags assigned by users, but cuss the results and provide a possible explanation for the also about the set of keywords assigned by the authors of observed di fference in navigational e fficiency. articles (extracted from library and metadata information), we can generate both broad and narrow folksonomies for the 2. RELATED WORK same set of resources (i.e. scientific articles) at the same Related work in this field of research can be split up into time. This means that we can generate broad folksonomies two di fferent parts: folksonomies , and navigation and hier- based on the tags users assigned to scientific articles, and archies in networks . we can generate narrow folksonomies for the same set of re- Folksonomies: In the past, folksonomies have been stud- sources based on the keywords that authors assigned to their ied from at least two di fferent perspectives – from an onto- papers. logical and an information retrieval perspective. From the In this work, we will compare the usefulness of broad vs. ontological perspective, our community analyzed emergent narrow folksonomies for a given task : navigation. We start semantic structures. For example [2, 14, 24] propose algo- by applying hierarchical clustering algorithms (such as the rithms for constructing semantically sound tag hierarchies algorithm by [2] and others) to create hierarchies of tags and from social tagging data. A detailed analysis of approaches keywords