Mining Social Tags to Predict Mashup Patterns

Khaled Goarany Gregory Kulczycki M. Brian Blake

Virginia Polytechnic Institute Virginia Polytechnic Institute University of Notre Dame
and State University and State University Notre Dame, IN, USA
Falls Church, VA, USA Falls Church, VA, USA
[email protected]
[email protected] [email protected]

ABSTRACT motivations for tagging range from content organization and

In the past few years, tagging has gained large momentum as a retrieval to social practices [35]. Recent research have studied the
user-driven approach for categorizing and indexing content on the use of tags in recommending Web resources [25][16][5][34]. In
Web. Mashups have recently joined the list of Web resources this line of research, tags are viewed as a rich, yet cheap, source of
targeted for social tagging. In the context of the social Web, a collective intelligence representing the community consensus on
mashup is a lightweight technique for integrating applications and what concept(s) best describe a Web resource.
data over the Web. Crafting new mashups is largely a subjective Recently, mashups have joined the list of Web resources targeted
process motivated by the users’ initial inspiration. In this paper, for social tagging. In the context of the social Web, a mashup has
we propose a tag-based approach for predicting mashup patterns, evolved as a user-centric lightweight approach for integrating
thus deriving inspiration for potential new mashups from the applications and data over the Web. The fast adoption of mashups
community’s consensus. Our approach applies association rule by online communities is motivated by its promise of enabling
mining techniques to discover relationships between APIs and nontechnical users to integrate existing applications in new and
mashups based on their annotated tags. We also advocate the innovative ways. This promise is further enhanced considering the
importance of the mined relationships as a valuable source for growing number of user-friendly mashup tools, such as Yahoo
recommending mashup candidates while mitigating for common Pipes [31], Intel Mash Maker [14] and IBM Mashup Center [13].
problems in recommender systems. We evaluate our methodology Furthermore, developers see mashups as a new approach for
through experimentation using real-life dataset. Our results show marketing their applications to a wider consumer base.
that our approach achieves high prediction accuracy and
outperforms a direct string matching approach that lacks the ProgrammableWeb [20] is a popular online community built
mining information. around user-generated mashups, where users collaborate in
posting, tagging and rating Web APIs and mashups. This social
Categories and Subject Descriptors aspect has attracted the attention of a number of researchers, and
H.2.8 [Database Management]: Database Applications – Data enabled the study of many interesting phenomena related to online
Mining; H.3.5 [Information Storage and Retrieval]: Online user behavior and mashup creation trends [27][33][29]. In our
Information Systems – Web-based services; H.5.3 [Information work, we utilize the ProgrammableWeb’s repository, to discover
Interfaces and Presentation]: Group and Organization Interfaces relationships between APIs and mashups based on their annotated
– Collaborative computing, Evaluation/methodology, Web-based tags, and use these relationships to predict interesting mashup
interaction. patterns. Relationships are discovered using association rule
mining techniques. We evaluate our methodology through
General Terms experimentation on real-life dataset crawled from
Experimentation, Performance. ProgrammableWeb. Our results demonstrate that our approach
achieves high pattern prediction accuracy with significant
Keywords performance gains with respect to a state-of-the-art direct string
matching approach. Our work is distinguished by three key
Social tags, Mashup, Web mining, User-generated content.
contributions:
1. INTRODUCTION (1) A mining approach for discovering mashup patterns
In the past few years, tagging has gained large momentum as a from user-generated tags - From a skeptic point of view,
user-driven approach for categorizing and indexing resources on tags are often considered as a form of uncontrolled
the Web. Thousands of photos on Flickr [10], videos on YouTube vocabulary that suffers from inconsistencies, typos,
[32], and bookmarks on Delicious [4] are being tagged daily. The proliferation of synonyms and unstructured organization. By
using empirical evaluation, we show that social tags can be
Permission to make digital or hard copies of all or part of this work for used effectively to detect patterns within online communities
personal or classroom use is granted without fee provided that copies are despite these characteristics. In our experimentation, we use
not made or distributed for profit or commercial advantage and that tags without any prior cleaning, stemming or consolidation
copies bear this notice and the full citation on the first page. To copy of terms. Our results show that accurate patterns can be
otherwise, or republish, to post on servers or to redistribute to lists, consistently discovered from raw tags by applying simple
requires prior specific permission and/or a fee. mining techniques that only consider tag pair relationships.
Proceedings of the 2nd International Workshop on Search and Mining
User-generated Contents (SMUC 2010), October 30, 2010, Toronto,
Canada. Copyright 2010 ACM.
(2) A mashup recommendation methodology that mitigates • Personalization; mashups extracting user-specific data
for the cold start problem - The mined tag relationships through APIs.
constitute a valuable source of mashup candidate • Focused View; mashups returning a subset of data from
recommendations for APIs that have no prior usage or an API based on a query.
community rating information. We hence provide a solution • Real-Time Monitoring; mashups monitoring real-time
for the cold start problem [24] typically experienced by updates in websites through their APIs.
recommender systems, where no recommendations can be
generated for newly added items due to their lack of rating Research in mashup recommendations is rooted in the automatic
and usage history. discovery of Web service compositions. Work in this field mostly
(3) A new heuristic for mashup recommendations to balance follows a semantic-based approach that relies on knowledge
the long tail effect - Another characteristic of online representation and ontology-based techniques. A survey can be
communities is the emergence of the long tail effect [18], found in [22]. These techniques aim at representing services’
where few items are constantly reused by the community functionalities and attributes using formal ontologies.
leading to continuous increase in their popularity. The Compositions are then discovered by applying systematic
majority of items however are marginalized because of their reasoning and inference methodologies on the ontological
limited usage and hence grow even less popular. This holds concepts. The Resource Definition Framework (RDF) [23] and
true for the ProgrammableWeb community, as stated in [27] Web Ontology Language (OWL) [17] are two standards used in
and confirmed by our own analysis. By resorting to tags, we semantically annotating services. While this approach offers a
can create a ripple effect in this community, where less formal and precise way of describing a service offering, building
popular APIs are recommended based on their tag similarity ontologies and annotating services with formal concepts remains
to the more popular APIs. This can benefit both API largely a manual process, guided by domain experts, and hence
developers by marketing their less known APIs, as well as complex and time consuming. This constitutes a bottleneck in
consumers by exposing them to potentially interesting APIs practical application of semantic techniques, hindering their
that they didn’t know of before. widespread adoption by service providers. Moreover, it has been
argued that the insufficient involvement of users in the
While we apply our approach to APIs and mashups, the use of construction of domain ontologies has rendered them non-
tags in recommendations is a general concept applicable to other intuitive from the users’ perspective [12]. Work in [2], [21], and
types of Web resources, where all the aforementioned benefits are [7] propose alternatives to mitigate for the semantic techniques
of equal importance. shortcomings, by relying on syntactical-based matching of
services. For example, in [2] an approach is proposed for
The remaining sections are organized as follows. Section 2 is a matching services based on the analysis of human naming
survey of related work. In Section 3, we present our solution tendencies within Web Service Description Language (WSDL)
methodology, including a detailed description of our mining specifications [28]. The approach studies the dependency
approach and its potential applications. The results and analysis of between the message names of services and their mashability. We
our experimentation are presented in Section 4. We conclude the use a similar approach to [2] as a baseline in our evaluation.
paper and outline our plans for future work in Section 5.
In recent years, mashups have evolved as a light-weight user-
2. RELATED WORK centric form of compositions created by combining Web APIs. In
ProgrammableWeb has attracted the attention of many researchers the Web context, an API can refer to RESTful services [19], RSS
as a popular online repository of mashups. In [27], the authors use feeds, Javascript widgets, or WSDL-based Web services, among
ProgrammableWeb to study the API and tag usage patterns. They other forms of Web accessible functionalities. The user-centric
employ clustering techniques to discover relationships between nature of mashups has encouraged its adoption by online
user categories and the usage of both the API and tags. Their work communities such as PorgrammableWeb, where users collaborate
concludes that users with geographical proximity tend to use the in posting and tagging both APIs and mashups.
same APIs and tags and favor the same mashups. By studying the
In our work, we consider social tags as a hybrid solution between
entropy of mashups created in ProgrammableWeb, they also
the semantic and syntactical approaches used with Web service
confirm that, generally, users tend to reuse popular mashup
compositions. First, tags describe the functionalities of APIs and
patterns. The authors of [33] perform surveys to discover
mashups from a user perspective, and hence their descriptions
expectations of non-programmers from mashups, and the level of
stem from practical usage. Also, the tags are also collectively
expertise needed to generate them. Their work leverages the
provided by the community, which disseminates the creation
reasoning that mashups constitute a valuable tool for empowering
effort. Moreover, tags are a generic form of descriptors for Web
non-developers on the Web. One of their findings is that the
resources, and hence can be used to annotate any Web API
mashup usefulness perceived by the end users is more important
regardless of its underlying protocol or standards. While tags
than the expected difficulty of creating it. Their results are
lack the consistency and rigor of a formal ontology, we
particularly useful for designers of Web mashup tools. In [29], the
demonstrate through our experimentation that they are still of
authors categorize mashups surveyed from ProgrammableWeb
great benefit for discovering relationships between different APIs
under the following categories, which we refer to later in this
and for predicting interesting mashups patterns.
paper:
Among the research employing social tags for Web services
• Aggregation; mashups aggregating data through multiple
discovery is [9], where a tag-based clustering technique is
APIs.
proposed for establishing similarity between services based on
• Alternate UI; mashups offering alternative interfaces for similarity between their tag clouds. Their work considers the tag
interacting with data from one or more APIs.
cloud as a lightweight form of ontology that can be used to index association rule is an implication of the form X ⇒ Y where X ⊆ I,
services and consequently retrieve services that match a user Y ⊆ I and X ∩ Y = ϕ. The rule X ⇒ Y has support s if s% of
query. A similar approach is proposed in [3], where tags are used transactions in DB contain X ∪ Y. The association rule holds in the
to discover and compose services through AI techniques. In this set of transactions DB with confidence c if c% of transactions in
research, the rational is that users are more likely to use DB that contain X also contains Y.
vocabulary from the tag cloud than from a formal ontology. We
share the same rational, but we employ an empirical evaluation The problem of mining association rules is to find all rules whose
using real-life data to show the usefulness of mining tags in support and confidence are higher than a given minimum support
discovering mashup candidates. and confidence thresholds. For an association rule X ⇒ Y, we can
calculate:
Tag-based recommendation has also been explored in [25], which
evaluates different algorithms for predicting users’ preferences for support(X ⇒ Y) = support(X ∪ Y) , and
movies based on their preferences for the movies’ associated tags. support(X ∪ Y)
This work builds on research in [16], which proposes a cluster- confidence(X ⇒ Y) =
based algorithm for recommending web-pages based on the pages support(X)
users have tagged, and [5], which creates tag-based user profiles
In our work, we model ProgrammableWeb as a set of mashups M,
to use in recommendations. In [34], a collaborative filtering
and a set of tags T. Each mashup in M is comprised of one or
approach is proposed that discover similarity between users based
more tags from T. These tags either annotate the mashup, or one
on the semantic distance among their tags. In our work, we apply
of its Web APIs. In this context, mashups correspond to the
association rule mining to discover relationships between tags,
transactions, and tags correspond to the items.
and use these relationships in predicting mashup patterns.
Recommending mashup candidates and API usage scenarios is By applying association rule mining, we aim to discover tag
among the potential applications of our approach. association rules on the form ti ⇒ tj where {ti , tj} ⊆ T. To achieve
this goal, we apply the following steps:
Mashup recommendation is another area of research relevant to
this work. A mashup recommender tool, called MashupAdvisor, is 1- For each mashup, we generate all possible tag pairs {ti ,
proposed in [8], which provides design-time assistance to mashup tj}, where:
creators. To generate recommendations, the MashupAdvisor a. ti and tj are two tags annotated to two
keeps a repository of mashups and uses it to calculate co- different APIs used by the mashup, or
occurrence patterns between each pair of concepts. Concepts are b. ti is annotated to one of the mashup APIs, and tj
extracted from APIs inputs and outputs. Similarity between is annotated to the mashup itself.
concepts is established based on domain independent thesaurus 2- We construct a knowledge base of all generated tag pairs.
and domain dependent ontologies. Compared to MashupAdvisor, Each pair is assigned a count, representing the number of
our proposed approach relies on user-assigned tags, instead of API mashups where the pair is encountered.
inputs and outputs, in generating recommendations. We believe 3- These counts are used to calculate support and confidence
that tags represent a more comprehensive set of concepts metrics for each pair.
describing the API functionalities and hence our
recommendations would potentially enable the discovery of a For example, consider the “Local Twitter Trends” mashup, listed
wider range of interesting APIs. Our experiments show that we on ProgrammableWeb. Local Twitter Trends displays the most
can generate recommendations without the need of constructing a posted terms on Twitter per location, by integrating data from
domain ontology, which is generally perceived as an impractical Twitter [26] and Google Maps [11] APIs. In ProgrammableWeb,
requirement. Intel Mash Maker [14] is another tool that enables Google Maps is annotated with the tag “mapping”, and twitter is
users to customize the content of browsed websites by adding annotated with “microblogging”, and “messaging”. Local Twitter
functionalities, visualizations or data from other websites. The Trends is annotated with “microblogging”, “mapping”, “local”,
recommendations provided by the tool are based on the content of and “trends”. By following steps 1.a and 1.b described earlier,
the Web resource, in their case the currently browsed website, we generate the following set of all possible tag pairs for this
rather than on mining the tags annotating the resource as in our mashup:
case. { {mapping, microblogging},
{mapping, messaging},
3. TECHNICAL APPROACH {mapping, trends},
Our proposed solution is based on the hypothesis that mining {mapping, local},
social tags can be used as a powerful tool for predicting mashup {microblogging, trends},
patterns. In this section we describe in details our methodology in {microblogging, local},
mining tags, and predicting mashup patterns. It also provides an {messaging, trends},
overview of the applications and benefits we envision for our {messaging, local} }
approach.
By repeating this process for each mashup crawled from
3.1 Mining Mashup Tags ProgrammableWeb, we construct a knowledge base from all
The association rule mining problem is formally defined in [1] as extracted tag pairs. This knowledge base enables us to calculate
follows: Let I = {i1,i2…in} be a set of items. An itemset X with k the support and confidence for each pair by tracking its
items from I is called a k-itemset. Let DB be a set of transactions, repeatability. A frequent tag pair; a pair whose support and
where each transaction A is an itemset such that A ⊆ I. Given an confidence are above specified thresholds, represents a discovered
itemset X ⊆ I, a transaction A contains X if and only if X ⊆ A. An association rule.
3.2 Predicting Mashup Patterns
As can be seen in Section 3.1, the annotated tags essentially
represent all the concepts related to the mashup. These concepts
may describe the functionalities of the mashed up APIs (e.g.
“mapping”, “microblogging”), the new functionalities introduced

API Count
by the mashup (e.g. “trends”), or any other attributes the users
find useful in describing the mashup (e.g. “local”).
In our model, each association rule discovered between these
concepts is an indication of how frequently users integrate the
concepts into mashups. We hence consider each rule as a mashup
pattern. In the previous example, mapping ⇒ trends is considered
a mashup pattern if support(mapping ⇒ trends) and
confidence(mapping ⇒ trends) are above specified thresholds. Mashup Count

This mining approach can be significantly beneficial in a number

Figure 1. The long tail distribution of API usage in
of applications. First, we can use the discovered patterns in
ProgrammableWeb
recommending mashup candidates within a set of APIs, based on
their tags. For each two APIs in the set, we can generate a list of By using our mining approach, an API is picked up in
tag pairs from cross-multiplying all the tags annotating the two recommendations based on the popularity of each of its tags. In
APIs. By consulting the knowledge base of association rules, we our model, the popularity of a tag is a function of the popularity of
can assign a mashability score to the two APIs, based on the all the APIs it annotates, and thus the less popular APIs benefit
collective support and confidence values of their generated tag from their tag similarity to the more popular ones. As mentioned
pairs. Accordingly, for each API in the set, we can recommend a earlier, we envision this to be of significant value to both API
list of APIs to use in mashups, ranked by their mashability score. developers and consumers.
Moreover, since our mining takes into account mashup tags, our 4. EXPERIMENTATION AND
recommendations are inherently not limited to mashup candidates,
but also include possible API usage scenarios. These scenarios EVALUATION
capture potential customizations for the original functionalities
provided by the APIs. For example, in 3.1 the mashup tag
4.1 Dataset
“trends” suggests a usage for the “microblogging” functionality In our experimentation, we employ a dataset consisting of 3647
of Twitter. This is particularly beneficial for the case of one-API mashups and 1574 tags, all crawled from ProgrammableWeb on
mashups, i.e. mashups that only customize the functionality of a January 5th, 2010. Out of the crawled mashups, 2047 are
single API. In ProgrammableWeb, this category currently randomly selected to build a knowledge base, 800 are selected as
comprises more than half of the listed mashups. a training set, and the remaining 800 are used as a test set. Based
on the model described earlier, each mashup is related to one or
We envision two major advantages in basing recommendations on more tags, which annotate the API(s) involved in the mashup, or
tags instead of APIs. Typically, recommender systems experience annotate the mashup itself.
a problem in generating recommendations for newly added items
that have no usage history, also known as the cold start problem 4.2 Experiments Details
[24]. A new API in ProgrammableWeb is assigned a set of tags, Our baseline experiment (experiment 0), investigates the use of
both on creation and later by the community users. We are hence tags similarity in predicting mashup patterns, in the absence of
able, using our approach, to generate instant recommendations for historical information on tags mashability. In a similar approach
new APIs based on their tags, if prior mined tag rules exist in the to that described in [2], this experiment employs the Levenshtein
knowledge base. This is particularly useful due to the fact that, algorithm [15] to measure similarity. The algorithm is used to
within the same community, users tend to reuse tags [27]. calculate the edit distance between two tags t1 and t2, editDist(t1,
t2), as the number of deletions, insertions, or substitutions required
The other advantage of tag-based recommendations is the to transform t1 to t2. As an example, editDist(“blog”, “book”) is
potential to reduce the long tail phenomenon detected in equal to 2.
ProgrammableWeb. Figure 1 depicts the distribution of a
randomly selected subset of APIs from ProgrammableWeb and Since the edit distance as described depends on the length of input
their corresponding mashup counts. As can be seen in the figure, tags, we normalize the results by dividing the distance by the
the majority of the APIs contribute to 3 or fewer mashups each, length of the longest tag. As we are interested in measuring the
while very few APIs in the community, those considered as the similarity between the tags, rather than their difference, we use the
most popular, are continuously reused in most of the listed following derived similarity metric:
mashups, despite the existence of other APIs offering similar
functionality. With the presence of this trend in editDist ( t1 , t 2 )
Sim( t1 , t 2 ) = 1 −
ProgrammableWeb, a mashup recommender system can easily fall Max( Length( t1 ), Length( t 2 ))
into the cycle of continuously recommending the more popular
APIs, based on their ever-growing usage, at the expense of the In our second experiment (experiment 1), we evaluate the benefit
increasingly marginalized majority. of using the knowledge base, described in details in section 3, in
predicting mashup patterns. The knowledge base contains a set of
9017 rules, representing 2-itemset tag association rules derived correlation between tag similarity and tag mashup ratio;
from 2047 mashups. percentage of mashups where a tag occurs, is found to be 0.05.
This indicates that tag similarity can be considered as a poor
In both experiments, mashup predictions are evaluated using criterion for discovering patterns.
precision and recall metrics. Precision and recall are calculated as
follows:
Precision Recall
tp
Precision =
tp + fp

tp
Recall = (b)
tp + fn

where tp is number of true positives, fp is number of false

positives, and fn is number of false negatives returned by the
experiment. In experiment 0, precision and recall are measured at
different tag similarity thresholds, while in the experiment 1 we
vary association rules support and confidence thresholds.

4.3 Results and Analysis

The following graphs display the results from experiments 0 and Support Threshold
1.
Figure 3. Precision and recall achieved by experiment 1 at
different support thresholds
Precision Recall
Figure 3 depicts the results obtained from experiment 1, where we
predict mashup patterns based on a knowledge base of mined
association rules. Precision and recall of patterns prediction are
plotted against different tag pair support thresholds. As the
support threshold increases, an increase in precision is observed,
accompanied with a decrease in recall. The precision and recall
dew point, indicated by point (b) in the figure, is at 55%; a 45%
increase over the dew point of experiment 0.
The tag mining approach realizes a 0.8174 Spearman correlation
(a) between calculated support values for tag pairs and tag mashup
ratio, indicating a powerful pattern prediction mechanism.
Moreover, further analysis of the results proves that this approach
has enabled us to predict tag mashability across all different
mashup categories from [29].
Tag Similarity Threshold
As can be observed in Figure 3, the tag mining approach generally
Figure 2. Precision and recall achieved by experiment 0 at achieves higher precision values than those achieved by using tag
different tag similarity thresholds similarity. This can be particularly useful for mashup
recommendation systems that return a predetermined number of
Figure 2 illustrates the effect of varying the tag similarity
mashup candidates based on predicted tag mashability. In such a
threshold on the precision and recall of patterns prediction. As
case, we would not be interested in recalling all possible
shown by the figure, the higher the similarity threshold, the higher
recommendations, but rather interested in ensuring the highest
the precision, but the lower the recall. Precision and Recall dew
precision possible for the returned results. Additionally, the
point, indicated by point (a) in the figure, is at 10%. The highest
calculated support values, used in predications, can also be used in
achieved precision is 47%, at a 100% similarity threshold, which
ranking the list of returned recommendations.
occurs when both input tags are identical. Through further
analysis of the results, it is found that the “Aggregation” mashup Next, we study the effect of complementing the support threshold
category, described in [29], dominates the returned predictions at by a confidence threshold, calculated from the mined tag pairs.
this threshold, particularly for mashups aggregating data from Figure 4 displays the precision and recall achieved for three
APIs offering similar functionality. Examples include “Only exemplary confidence thresholds (0, 0.02, and 0.04) at a fixed
Apple Stories”, a mashup aggregating Apple Inc. news queried support threshold of 0.001. As illustrated, increasing the
from Digg [6] and Delicious APIs, both tagged with “bookmark”, confidence threshold leads to a slight increase in precision.
and “Google vs Yahoo Maps”, a mashup displaying locations However, this increase comes at the expense of a significant
side-by-side from Google Maps and Yahoo Maps [30] APIs, both decrease in recall. This shows that confidence can be more
tagged with “mapping”. beneficial as an additional restrictive criterion to filter out false
predictions, in the cases where recall is of less importance. Again,
The correlation results indicate that the tag similarity approach
generally falls short in predicting mashup patterns. The Spearman
a typical application would be recommendation systems of fixed services. Proceedings of the 30th international conference
result size. on Very large data bases. VLDB'04 (2004), 372-383.
[8] Elmeleegy, H., Ivan, A. et al. 2008. Mashup advisor: a
recommendation tool for Mashup development.
Precision Recall Proceedings of the IEEE International Conference on Web
Services. ICWS'08 (2008), 337–344.
[9] Fernandez, A., Hayes, C. et al. 2008. Closing the Service
0.617
0.563
0.592 Discovery Gap by Collaborative Tagging and Clustering
0.536
Techniques. Proceedings of the 7th International Semantic
0.393
Web Conference. ISWC'08 (2008), 115-128.
[10] Flickr. http://www.flickr.com/.
0.248 [11] Google Maps. http://maps.google.com/.
[12] Hepp, M. 2007. Possible Ontologies—How Reality
Constrains the Development of Relevant Ontologies. IEEE
Internet Computing. (2007), 90–96.
[13] IBM Mashup Center. http://www-
0 0.02 0.04
01.ibm.com/software/info/mashup-center/.
Confidence Threshold [14] Intel Mash Maker. http://mashmaker.intel.com/web/.
[15] NIST Levenshtein distance.
Figure 4. Precision and recall for different confidence
http://www.itl.nist.gov/div897/sqg/dads/HTML/Levenshtein
thresholds at a fixed support threshold.
.html.
5. CONCLUSION AND FUTURE WORK [16] Niwa, S., Doi, T. et al. 2006. Web Page Recommender
In this paper, we demonstrate using empirical evaluation that we System based on Folksonomy Mining. Proceedings of the
can apply association rule mining to social tags to learn and 3rd International Conference on Information Technology:
predict mashup patterns. Our work proves through New Generations. ITNG '06 (2006), 388-393.
experimentation that social tags can be of significant benefit [17] OWL Web Ontology Language Overview.
despite their often touted unfavorable characteristics. Our results http://www.w3.org/TR/owl-features/.
show that, while only considering 2-itemsets of raw tags that did [18] Park, Y. and Tuzhilin, A. 2008. The long tail of
not undergo any prior processing, we are able to achieve high recommender systems and how to leverage it. Proceedings
accuracy in predicting patterns. of the 2008 ACM conference on Recommender systems
(2008), 11-18.
We envision a number of applications for our proposed [19] Pautasso, C., Zimmermann, O. et al. 2008. RESTful Web
methodology, including a tag-based recommendation system for Services vs. “Big” Web Services: Making the Right
mashup candidates, and API usage scenarios. Our methodology Architectural Decision. Proceedings of the 17th
overcomes the cold start problem and the long tail phenomenon International Conference on World Wide Web (2008),
exhibited by other recommender systems that rely on historical 805–814.
usage information. [20] ProgrammableWeb. http://www.programmableweb.com/.
[21] Pu, K., Hristidis, V. et al. 2006. Syntactic rule based
Our future work includes mining larger itemsets consisting of
approach to web service composition. Proceedings of the
three or more tags, to discover more complex relationships. We
22nd International Conference on Data Engineering.
are also planning to evaluate the use of different tag-based
ICDE'06 (2006), 31–31.
clustering techniques in predicting mashup categories.
[22] Rao, J. and Su, X. 2005. A Survey of Automated Web
6. REFERENCES Service Composition Methods. Proceedings of the 1st
[1] Agrawal, R., Imieliński, T. et al. 1993. Mining association International Workshop on Semantic Web Services and
rules between sets of items in large databases. SIGMOD Web Process Composition. SWSWPC'04 (2005), 43-54.
Rec. 22, 2 (1993), 207-216. [23] RDF - Semantic Web Standards. http://www.w3.org/RDF/.
[2] Blake, M.B. and Nowlan, M.F. 2008. Predicting service [24] Schein, A.I., Popescul, A. et al. 2002. Methods and Metrics
mashup candidates using enhanced syntactical message for Cold-Start Recommendations. Proceedings of The 25th
management. Proceedings of the IEEE International Annual International ACM SIGIR Conference on Research
Conference on Services Computing. SCC'08 (2008). And Development In Information Retrieval. (2002), 253-
[3] Bouillet, E., Feblowitz, M. et al. 2008. A folksonomy- 260.
based model of web services for discovery and automatic [25] Sen, S., Vig, J. et al. 2009. Tagommenders: connecting
composition. Proceedings of the IEEE International users to items through tags. Proceedings of the 18th
Conference on Services Computing. SCC'08 (2008). international conference on World wide web. WWW'09
[4] Delicious. http://delicious.com/. (2009), 671-680.
[5] Diederich, J. and Iofciu, T. 2006. Finding communities of [26] Twitter. http://twitter.com/.
practice from user profiles based on folksonomies. [27] Wang, J., Chen, H. et al. 2009. Mining user behavior
Proceedings of the 1st International Workshop on Building pattern in mashup community. Proceedings of the 10th
Technology Enhanced Learning solutions for Communities IEEE international conference on Information Reuse &
of Practice. TEL-CoPs'06 (2006). Integration. IRI'09 (2009), 126–131.
[6] Digg. http://digg.com/. [28] Web Services Description Language (WSDL) Version 2.0.
[7] Dong, X., Halevy, A. et al. 2004. Similarity search for web http://www.w3.org/TR/wsdl20-primer/.
[29] Wong, J. and Hong, J. 2008. What do we "mashup" when Centric Computing. VL/HCC'08 (2008), 31–38.
we make mashups? Proceedings of the 4th International [34] Zhao, S., Du, N. et al. 2008. Improved recommendation
Workshop on End-user Software Engineering (2008), 35- based on collaborative tagging behaviors. Proceedings of
39. the 13th international conference on Intelligent user
[30] Yahoo! Maps. http://maps.yahoo.com/. interfaces (Gran Canaria, Spain, 2008), 413-416.
[31] Yahoo! Pipes. http://pipes.yahoo.com/pipes/. [35] Zollers, A. 2007. Emerging motivations for tagging:
[32] YouTube. http://www.youtube.com/. expression, performance, and activism. Proceedings of
[33] Zang, N. and Rosson, M.B. 2008. What’s in a mashup? and 16th International World Wide Web Conference. WWW'07.
why? studying the perceptions of web-active end users. (2007).
IEEE Symposium on Visual Languages and Human-