Collaborative tagging has recently attracted the attention of both industry and academia due to the popularity of content-sharing systems such as CiteULike, del.icio.us, and Flickr. These systems give users the opportunity to add data items and to attach their own metadata (i.e., tags) to stored data. The result is an effective content management tool for individual users. Recent studies, however, suggest that, as tagging communities grow, the added content and the metadata become harder to manage due to increased content diversity. Thus, mechanisms that cope with increase of diversity are fundamental to improve the scalability and usability of collaborative tagging systems. This paper analyzes whether usage patterns can be harnessed to improve navigability in a growing knowledge space. To this end, it presents a characterization of two collaborative tagging communities that target the management of scientific literature: CiteULike and Bibsonomy. We explore three main directions: First, we analyze the tagging activity distribution across the user population. Second, we define new metrics for similarity in user interest and use these metrics to uncover the structure of the tagging communities we study. The properties of the structure we uncover suggest a clear segmentation of interests into a large number of individuals with unique preferences and a core set of users with interspersed interests. Finally, we offer preliminary results that suggest that the interest-based structure of the tagging community can be used to facilitate content retrieval and navigation as communities scale.
Two types of techniques, implicit and explicit, are traditionally
used to elicit user preferences in the Web context [
In a tagging community context, the tags themselves can be
interpreted as explicit metadata added by each user. Additionally,
observed tagging activity including the volume and frequency
with which items are added, the number of tagged items, or tag
vocabulary size can be harnessed to extract implicit information.
Due to the youth of collaborative tagging systems, relatively little
work has been done on tracking usage and exploring
contextualized user attention in these communities. However,
several studies present techniques and models for collecting and
managing user attention metadata in the wider web context
without exploring tagging features [
As a first step to modeling user attention in tagging communities,
it is necessary to characterize collaborative tagging behavior. In
this respect, Golder and Huberman [
The urn model can be formulated as follows: consider an urn that contains two colored balls. Iteratively, a ball is drawn at random from the urn and returned to the urn together with a new ball of the same color. If this process is repeated a number of times, the fraction of balls of a particular color stabilizes. The interesting aspect of this model is that if the process is restarted, this fraction converges to a different number. Golder and Huberman argue that this model captures the evolution of tag proportion observed in the del.icio.us data set. In studies related to contextualized user attention, this model may be valuable to predict future user tagging assignments which can be a useful input to recommendation mechanisms. Golder and Huberman’s study, however, is limited in scale: their results on tagging behavior dynamics rely on only four days of tracked activity.
Other authors follow different approaches to investigate the
characteristics of tagging systems. Schimtz [
By analyzing del.icio.us, Chi and Mytkoswicz [
Niwa et al. [
A collaborative tagging community allows users to tag items via a web site. Users interact with the website by searching for items, adding new items to the community, or tagging existent items. The tagging action performed by a user is generally referred as a tag assignment.
For example, in CiteULike and Bibsonomy, each user has a library, i.e., a set of links to scientific publications and books. Each item in the library is associated with a set of terms (tags) assigned by users. It is important to highlight that, in both CiteULike and Bibsonomy, the process of assigning tags to items is collaborative, in the sense that all users can inspect other users’ libraries and assigned tags. User can thus repeat tags used by others to mark a particular item. This is unlike other communities (e.g., Flickr) where each user has a fine-grained access control to define who has permissions to see the content and apply tags to it. In CiteULike and Bibsonomy users have two options to add items to their libraries: 1. Browse the content of popular scientific literature portals (e.g. ACM Portal, IEEE Explorer, arXiv.org), to add publications to their own library, and 2. Search for items present in other users' libraries and add them to their own library.
While posting an item, a user can mark it with terms (i.e., tags) that can be used for future retrieval. The collaborative nature of tagging relies on the fact that users potentially share interests and use similar items and tags. Thus, while the tagging activity of one user may be self-centered the set of tags used may facilitate the job of other users in finding content of interest.
We represent a collaborative tagging community by the tuple: C=(U,I,T,A), where U represents the set of users, I is the set of items, T is the tagging vocabulary, and A the set of tag assignments.
The set of tag assignments is denoted by A = {(u, t, p) | u ∈ U, t ∈ T, p∈ I}. From this definition of tag assignments, we can derive the definition of an individual user, item and tag, as follows: A user uk ∈ U is denoted by a pair uk = (Ik, Tk), where Ik is the set of items user k has ever tagged. Thus, an item p ∈ Ik if and only if ∃ (uk,t, p) ∈ A, for any t ∈ Tk. Similarly, Tk is the set of tags user uk applied before, where t ∈ Tk if and only if ∃ (uk, t, p) ∈ A.
An item pi ∈ I is denoted by pi = (Ui , Ti ), where Ui is the set of users who tagged this item, and Ti is set of tags this item has received.
A tag tj ∈ T is denoted by tj = (Uj , Ij ), where Uj is the set of users who used the tag tj before, and Ik is the set of items annotated with the tag tj.
Both tagging communities we analyze: CiteULike [
The data sets analyzed in this article were provided by the administrators of the respective web sites. Thus, the data represents a global snapshot of each system within the period determined by the timestamps in the traces we have obtained (Table 1). It is important to point out that the Bibsonomy data set has timestamps starting at 1995, which we considered a bug. Moreover, Bibsonomy has two separate datasets, scientific literature and URL bookmarks. We concentrated our analysis on the scientific literature part of the data.
In the original CiteULike data set, the most popular tag is “bibteximport” while the second most popular tag is “no-tag”, automatically assigned when a user does not assign any tag to a new item. The popularity of these two tags indicates that a large part of users use CiteULike as a tool to convert their list of citations to BibTex format, and that users tend not to tag items at the time they post a new item to their individual libraries. Clearly, this is relevant information for system designers who might want to invest effort in improving the features of most interest. Also, in CiteULike one user posted and tagged more than 3,000 items within approximately 5 minutes (according to the timestamps in the data set). Obviously, this behavior is due to an automatic mechanism. Our objective is to concentrate only on those users who are using the system interactively to bookmark and share articles. Consequently, for the analysis that follows, we have the “robot” user (i.e., a user with 3,000 items tagged within 5 minutes) and users who used only the tags bibtex-import and/or no-tag. The total number of users removed from CiteULike represents approximately 14% of the original data set, while the users removed from Bibsonomy are around 0.6% of the original data set. Table 1 summarizes the characteristics of each data set after the data cleaning operation.
To gain an understanding on the usage patterns in these two communities, we start by evaluating the activity levels along several metrics: the number of items per user, number of tagging assignments performed, and number of tags used. The question answered in this section is the following: Q1: How is the tagging activity distributed among users? We aim to quantify the volume of user interaction with the system, either by adding new content to the community, or by tagging an existing item. Intuitively, one would expect that a few users are very active while the majority rarely interacts with the community.
Determining how often users perform tag assignments is
important to help designing systems that track user attention. For
example, in a context where activity information is used to
recommend new items based on tag similarity, it can be necessary
to compute the similarity level at the same rate as the rate with
which new information is added into the system. Figure 1 presents
the user rank according to the number of tag assignments
performed during the time frame of our data set. In the results that
follow, we present the data points observed together with a curve
that provides a good model to the observed data (i.e. Hoerl
function [
A second metric for tagging activity is the size of user libraries. Figure 2 plots user library size for users ranked in decreasing order according to the size of their libraries for CiteULike and Bibsonomy, respectively. This shows the size of the set of items a particular user pays attention to. The results confirm that the users in these two systems are heterogeneous in terms of activity intensity, as it has already been indicated by the tag assignment activity.
The correlation between a user’s library size and her vocabulary is
important to understand whether the diversity of the vocabulary
used by each user grows with the number of items in her personal
library. We observe that, in both communities the users’ library
and vocabulary sizes are strongly correlated for CiteULike
(R2 = 0.98, n = 5954) and less strongly, but still positively
correlated, for Bibsonomy (R2 = 0.80, n = 654). Although such
correlation may seem intuitive, since users with a more diverse set
of items would need more tags to describe them, this behavior is
different from that observed by Golder and Huberman in
del.icio.us [
A third finding is that tagging activity distributions are not well
modeled by a Zipf-like distribution. Instead, a Hoerl model [
f(x) = abxxc
Similar to the Zipf distribution, the Hoerl function has been used
to model a large number of natural phenomena. The most relevant
to collaborative tagging is the use of Hoerl function to describe
the distribution of bio-diversity across a geographic region
[
Although a Hoerl function is a good fit for the activity distributions, this does not directly imply that diversity of user libraries or vocabularies represents a phenomenon which is similar to those presented by studies on biodiversity. Nevertheless, the Hoerl function does provide a good model for collaborative tagging activity and it can be useful to study user diversity in collaborative tagging systems in the future. To summarize: in the communities we study, the intensity of user activity is distributed over multiple orders of magnitude, it is well modeled using the Hoerl function and, unlike in other communities, there is a strong correlation in activity in terms of items set and vocabulary sizes.
While the analysis above is important for an overall usage profile
evaluation of each community, it provides little information about
user interests. Assessing the commonality in user interests is
important for identifying user groups that may form around
content of common interest. Thus, a natural set of questions that
we aim to answer in this section are:
Q2: Is the tagging community segmented into several
subcommunities with different interests? Do users cluster around
particular items and tags?
To address these questions, we define the interest-sharing graph
after the intuition of data-sharing graphs introduced by Iamnitchi
et al. [
More formally, consider a graph G = (U, E) where nodes are users and edges represent the existence of shared interests or activity similarity between users. The rest of this study explores three possible definitions for user interest or activity similarity. All these definitions employ a threshold t for the percentage of items or tags shared between two users: 1) The User-Item similarity definition considers two users’ interests similar if the ratio between the sizes of the intersection and the size of the union of their item libraries is larger than a threshold t. This is expressed by Equation 2. 2) The User-Tag definition is similar to the definition above but considers the vocabularies of the two users rather than their libraries.
ekj ∈ E ⇔
3) Unlike the User-Item definition in Equation 2 above, the Directed User-Item considers two users’ interests similar if the ratio between the intersection of their item libraries and the size of one user library is larger than a threshold t. The idea is to explore the role played by users with large libraries via the introduction of direction to the edges in the graph. In our analysis of real tag assignment traces from the two tagging communities, even with low values for the sharing ratio threshold t, the final graph contains a large number of isolated nodes. Indeed, by setting the threshold as low as one single item (i.e., two users are connected if they share at least one item); we find that, in CiteULike, 2,672 users (44.87%) are not connected to any other user. This suggests that a large population of users has individual preferences. increases (Note that we exclude isolated nodes from this count of connected graph components).
The plots in Figure 4 show that the number of connected components increases up to a certain value of our similarity threshold. After a certain value of t, the number of connected components in the graph starts decreasing, since more and more connected components will contain only one node and will thus be excluded. The critical threshold value is different for each user similarity definition. Figure 4 presents, for the three similarity metrics defined above, the number of connected components for both CiteULike and Bibsonomy, for thresholds t varying from 1% to 99%. These results show that regardless of the graph definition the number of connected components follow a similar trend as the threshold The initial increase in the number of connected components can be explained by the fact that, as the threshold increases, large components split to form new islands. Since these islands form naturally based on user similarity this result is encouraging since it offers the potential to cluster users according to their interests. As t continues to increase the definition of similarity becomes too strict and leads to more and more isolated nodes.
Two observations about the results in Figure 4 can be noted: first,
as the threshold increases, the number of components decreases
faster for the User-Item graph (Equation 2) than for the
DirectedUser-Item graph (Equation 4). This illustrates the effect of using
an asymmetrical definition for shared interests. The idea explored
here is similar to the friendship graph, where connections
between two nodes are not reciprocal (e.g., A might consider B to
be his friend, but, at the same time, B is indifferent to A) [
Second, there are more isolated nodes (i.e., zero-degree nodes) in the User-Item and Directed-User-Item graphs than in the User-Tag graph (defined by Equation 3). This indicates that users tend to have larger overlaps among their vocabularies than among their libraries. One reason for this observed interest sharing pattern may be the fact that users browse and consume items from other users' libraries without necessarily adding those items to their personal libraries in the system. An approach to verify this observation is to perform an analysis of user browsing histories to determine how often users download items from others’ libraries without adding them to their personal set of items. From a system design perspective, knowing that users are more likely to share tags than data items may be useful for designing item recommendation heuristics based on vocabulary overlap.
All the similarity definitions above generally divide the original graph into one giant component, several tiny components, and a large number of isolated nodes. Figure 5 presents the total number of nodes in the components with at least two nodes and the number of nodes in the largest connected component for thresholds varying from 1% to 99% for the three similarity measures defined above.
The results presented in this section demonstrate that using a similarity metric and the resulting interest-sharing graph it is possible to segment the user population according to manifested interest. Based on this intuition, we conjecture that it is possible to build tag/item recommendation mechanisms that exploit usage patterns, i.e., the shared interests among users. The next section offers a preliminary analysis of this hypothesis.
Chi and Mytcowicz [
We have verified that the diversity of the data present in the two communities we study grows over time. Figure 6 presents the evolution of entropy, as the metric to evaluate item diversity, for the CiteULike data since November 2004.
As a reminder, we note that the entropy of a set of items I, where |I| = N, is defined as:
N H (I ) = −∑ pi ⋅ log( pi ) i=1 (5) where pi is the popularity of an item i.
The entropy of a set may increase in two circumstances: First, as the randomness in the item set increases (i.e., the popularity distribution gets closer to a uniform random distribution) and, second, as the set becomes larger. In practical terms, in a collaborative tagging community, the increase in entropy of an item set means that the user needs to filter out more items to find the one she is interested in. Similarly, high entropy makes it harder to find a tag that describes an item well. Conversely, lower entropy makes it potentially easier for a user to reach an item of interest. Thus, the question to be answered in this section is the following: Q3: Can the interest-sharing graph be used to reduce the entropy perceived by users when navigating through the system? Our two-part answer is briefly presented below and detailed in the rest of this section. First, we demonstrate that the interest-sharing graph can be used to reduce the entropy perceived by users. To this end we define a user’s neighborhood as its set of neighbors in the sharing graph and show that this construction can be used to present users with an item set with low entropy.
Second, we offer preliminary results that suggest that this segmentation of the user population based on neighborhoods in the interest-sharing graph has a good predictive power: the items consumed by a user’s neighbors predict well the future consumption pattern of that user. Thus, this offers a path to build recommendations systems based on the interest-sharing graph. The rest of this section presents the above two-step exploration in detail. We first define a user’s ‘neighborhood entropy’ as the entropy of the union of the item sets of that user’s neighbors in the interest-sharing graph. To demonstrate that our group selection technique is effective in reducing entropy, Figure 7 compares the average neighborhood entropy in the interestsharing graph with two types of other constructions. All results are reported with 95% confidence intervals.
Firstly, we compare for CiteULike, the average neighborhood entropy in the largest connected component of an interest sharing graph (Average Entropy – Figure 7), which is built by using the real tagging activity trace, to the total entropy in the system (almost double at 11.6 as presented in Figure 6). Similarly, the neighborhood entropy in the largest connected component is compared to the total entropy in the same component (Largest Component – Figure 7). Secondly, we compare with two random graph constructions: the entropy of a random graph, which is built by selecting random neighbors from the set of users in the largest connected component (Random Component – Figure 7); the second random graph is built by selecting random from the entire user set (Random Graph - Figure 7).
We observe that the average ‘neighborhood entropy’ in the interest sharing graph is lower for almost all thresholds analyzed: First, the ‘neighborhood entropy’ in the interest sharing graph is between one half and one tenth of the entropy of the entire dataset. Additionally, for all thresholds analyzed except for the 1% threshold that does not offer enough discrimination the ‘neighborhood entropy’ is significantly (24% to 400%) lower than that of similar random constructions.
To support our hypothesis that the interest-sharing graph is a good basis to develop recommendation systems, we analyze how efficient the neighbor’s item set in predicting future user attention over items. To this end, we evaluate the hit ratio: the proportion of items a user adds to her library at time T+1 that are already in her neighbor’ libraries at time T.
To evaluate the hit ratio, we considered the interest-sharing graph based on the User-Item similarity metric with 1% sharing ratio threshold. Preliminary results show that depending on the granularity considered (that is the length of our forecasting period: interval between T and T+1) the hit rate is as high as 20% for one hour granularity and decays to a low of 5% for a one-month forecast granularity. This indicates that a user’s neighborhood is a possible source of information to predict near future user attention and its predictive effectiveness decreases for longer time intervals.
The findings presented on this section can be summarized as follows: First, we verify that the entropy increases as the CiteULike community grows overtime; Second, we verify that the interest-sharing graph can be used to reduce the entropy of item sets presented to users; and finally, we find that a user’s ‘neighborhood’ in the interest-sharing graph is a possible source of information to predict future user actions. While the preliminary data we present here is encouraging we continue tour exploration for a complete analysis for diverse interest-sharing graphs based on multiple similarity metrics and thresholds.
This work presents a characterization of two collaborative tagging systems, CiteULike and Bibsonomy, as a first step to help extract implicit information about user attention in collaborative tagging systems.
First, we analyze the distribution of tagging activity, i.e., the distribution of the volume of items, tags, and tagging actions related to each user’ activity in the tagging community. We find that the activity distribution is highly heterogeneous along all these multiple axes: a few active users contribute with a large number of tag assignments and maintain a large number of items and tags, while the majority of users have a modest tagging activity.
Additionally, users with large libraries tend to have a large vocabulary of tags. While this may seem intuitive, this is not the norm across all tagging systems. In del.icio.us, for example user’s library size and number of tags used are uncorrelated. Second, we define the interest-sharing graph and investigate several definitions for interest similarity based on user activity in terms of items and vocabularies employed. Our main findings can be summarized as follows: 1.
Both communities present a large population of isolated users (zero-degree nodes in the interest-sharing graph). This indicates that there are a large number of users with unique preferences. On the other hand, by introducing direction in the graph of shared interests, it is possible to reduce the number of isolated nodes. The final main directed connected component contains approximately twice more nodes than the undirected one.
The structural analysis reveals the existence of a significant number of small sub-communities of interests totally separated from each other.
The structure of the interest-sharing graph can be used to reduce the diversity of the items a user is exposed to. To quantify this reduction we compare the entropy of the ‘neighborhood item set’ with that of the item set of the entire CiteULike/Bibsonomy item set, that of the main connected component, and that of various constructions of random item sets of similar sizes.
Finally, we provide preliminary evidence that suggests that user’s activity can be predicted by considering the union of the item sets of a node’s neighbors in the interest sharing graph. We conjecture that this property can be used to build efficient, online recommendation systems for tagging communities.
This work inspires a fresh set of questions on collaborative tagging communities and contextualized attention.
A possible approach to implicitly extract user attention is to
answer the following question: What are the patterns of
information propagation in collaborative tagging systems? An
answer to this question could build on previous work on
information diffusion in the blogspace [
A second intriguing issue to explore is the following How
malicious behavior affects a tagging system and whether it be
automatically detected? Search results that are manipulated by
tagging misbehavior can have an impact on usage in a
collaborative tagging community [
Finally, exploring the structure of user interest overtime can help
devising models for the formation and evolution of the user
similarity graphs, similarly to the study by Kumar et al. [
The authors would like to thank Richard Cameron for providing the CiteULike data set; Christoph Schmitz for providing the Bibsonomy data set; professor Lee Iverson for insightful discussions on early stages of this work, and Armin Bahramshahry, Samer Al Kiswany and Nazareno Andrade for their valuable comments. The graph analysis was executed in parallel using OurGrid (http://www.ourgrid.org).