=Paper=
{{Paper
|id=Vol-1310/paper5
|storemode=property
|title=Web Experience Enhancer Based on a Fast Hierarchical Document Clustering Approach
|pdfUrl=https://ceur-ws.org/Vol-1310/paper5.pdf
|volume=Vol-1310
|dblpUrl=https://dblp.org/rec/conf/semweb/DospinescuPC14
}}
==Web Experience Enhancer Based on a Fast Hierarchical Document Clustering Approach==
https://ceur-ws.org/Vol-1310/paper5.pdf
Web Experience Enhancer Based on a Fast
Hierarchical Document Clustering Approach
Alexandru Ionut Dospinescu1 , Florin Pop1 , Valentin Cristea1
University Politehnica of Bucharest, Faculty of Automatic Control and Computers,
Computer Science Department, Romania
alexandru.dospinescu@cti.pub.ro, florin.pop@cs.pub.ro.
valentin.cristea@cs.pub.ro
Abstract. As the importance of the Internet has increased rapidly to
new heights especially with wide adoption of web services and advent of
cloud computing, the amount of sheer information that can be obtained
from this global information service center leads to major challenges
concerning finding relevant information for a casual user or a group of
users. On the other hand, even after finding useful documents, it is hard
to keep track of them as the amount of useful findings (bookmarks)
increases. Also, one has to manually bookmark each time a page is even
remotely interesting in order to make it easier to return to that page. In
this paper we formulate a new approach to tackling these issues through
a model that takes into account not just a single user’s experience and
intentions, but those of an entire group of users as well. We propose
an application in the shape of a browser plugin/extension powered by a
dedicated and highly responsive web server that can help the user keep
track of his web experience in a highly organized and easy to navigate
fashion while also enabling the user to share some of that experience. The
clustering of the data is done with a customized algorithm that allows
overlapping while showing improved speed and readability.
Keywords: web document clustering, closed frequent itemsets, social
web, user interaction, web application
1 Introduction
With the ongoing trends of web services and the advent of cloud computing,
the users are brought closer and closer to the Internet. Its importance in the
functioning of the civilized society has never been so staggering. Still, the ever
growing amount of information can overwhelm even the best web search engine
in finding relevant information for a given user, especially when faced with the
trends of applying Search Engine Optimization techniques such as stuffing web
sites with a high amount of low-quality links, with little or no topical connection
to the host site, in an attempt to gain a higher search engine results page.
A highly adopted solution to the problem of finding relevant information
relative to the intention of the user is search result clustering by which documents
from the results list returned are grouped and labeled with relevant tags [11].
Such is the case with the commercial solutions Carrot or Vivisimo [3]. Still, even
with such an organized structure it is possible that the information under a group
may be either too encompassing for what the user seeks. Also, this clustered
information retrieval is exclusively dependent upon the results returned by the
search engine which, in turn is highly dependent on the search query. For users
that lack the knowledge of the proper terminology for building the appropriate
queries to supply to the search engine, this is a great issue.
Even after finding useful documents, the vast amount of information can
create obstacles in keeping track of them. While bookmarks are a temporary
solution, as the size of bookmarks increases, they soon lose relevance. Also, there
is another limitation with bookmarks which is reflected in the user experience.
Each time a user finds a page to be remotely interesting, that user must manually
bookmark that page. This is highly unwanted and a more unobtrusive approach
would be preferred to collecting and labeling references by hand.
In order to alleviate some of the problems encountered by casual as well as
experienced users when searching and keeping track of relevant information we
propose an application that would keep track of a user’s web experience (visited
web pages) in a seamless way, would be able to present it to the user in a highly
organized and easy to navigate fashion and would enable users to share their
experience with other users that use the application. Also, it would provide a
way for a user to filter through his web experience.
Practically we suggest a browser plugin/extension (our first implemented ver-
sion is in the form of a Google Chrome extension) that is backed by a dedicated
web server running on a cloud platform that does that by meeting the following
requirements:
• the user’s experience is seamlessly recorded and stored in an efficient manner
as the user accesses various web documents;
• when the user wants to view this experience through the application, the web
server should be able to supply the extension with a hierarchical structure
representing his experience;
• a user should be able to filter through his web experiences based on time,
popularity or a feedback indicator;
• a user should be able to share experience as well as make use of others’
shared experiences;
• the application should be very responsive and scalable.
For the last requirement we have implemented a custom version of the Fre-
quent Itemset Hierarchical Clustering (FIHC) [5] algorithm. We present in this
paper the architecture and technologies used for its development, the advantages
and novelty by referring to others algorithms. We evaluate of the proposed ap-
plication by comparison of cluster quality and clustering speed between FIHC
and iFIHC (improved FIHC).
The rest of the paper is structured as follows. In Section 2 we will briefly
overview some of the most relevant related work. In Section 3 we will describe
the problem at hand and our solution and architecture. In Section 4 we will
present some of our experimental results and finally in Section 5 we will draw
our conclusions and future directions for our application.
2 Related work
In this section we will briefly overview document clustering while focusing on
FIHC and frequent itemset mining algorithms.
2.1 Document Clustering
Document clustering can be seen as being divided into: flat/partitioning, hierar-
chical, density-based, probabilistic, graph-based and model based. Among these
the most common approaches have been the hierarchical and partitioning ones.
Regarding the partitioning clustering algorithms, the ones that are mostly
used for web document clustering are flat medoid-based. This class of algorithms
results in partitions of the document collection called k-means. Each partition
will be represented by an object within it. The main scope of this class is to
find medoids that minimize the dissimilarities inside their cluster. Some of the
more promising algorithms from this class are K-Means, K-Medoids, CLARA,
CLARANS [8]. Though these algorithms are very resilient to outliers and are
faster than hierarchical algorithms, they do not allow overlapping which is some-
thing truly required for our approach.
Still, there is another type of partitioning cluster algorithms, namely fuzzy
algorithms. This class allows for overlapping. One of the most representative
such algorithm is FCM [2]. The basic idea behind it is that it tries to minimize
an objective function normally describing the distance between the medoids of
all the clusters an object might belong to. It starts with randomly selecting
initial medoids and calculating the membership of all the objects in the dataset
to those medoids. At each iteration, new medoids are computed based on the
current memberships and new memberships are then assigned. This algorithm
would have sufficed only if it didn’t require a predefined number of clusters which
cannot be known given the dynamic nature of our problem.
Concerning the class of hierarchical clustering algorithms one that produces
really good results is UPGMA [6]. It uses a measure determined by the cosine
similarity between clusters divided by the size of the clusters being evaluated.
But since it requires a precomputed similarity matrix, it is expensive and inflex-
ible when faced with an expanding collection of documents.
Another popular approach for this class of algorithms is taken by those based
on frequent itemset methods. This subclass of algorithms is based on clusters
being represented by labels made by itemsets that are very frequent inside the
representative cluster and very rarely in others. Such labeling provides important
information regarding the structure of the cluster contents.
One of the more scalable methods belonging to this class is FIHC that,
based on the frequent itemsets found by employing the Apriori algorithm [1],
organizes clusters into a topic hierarchy. Some of the features and advantages of
this approach are:
• reduced dimensionality of the vector space;
• a very high clustering accuracy and speed;
• does not require a predefined number of final clusters as input;
• easy to navigate with meaningful descriptions.
Overall, this algorithm can be divided into four independent phases:
1. Extracting the feature vectors - stop word removal, stemming and vectoriza-
tion;
2. Determining the global frequent itemsets - all sets of terms that are present in
at least a minimum global fraction of the document set are determined. The
terms are tested for global frequency by their presence and not by TF (term
frequency) or TFxIDF (term frequency x inverse document frequency);
3. Clustering - generates the initial clusters from the global frequent itemsets
and then makes the clusters disjoint based on a score function that represents
how well a document belongs to a certain cluster;
4. Constructing the cluster tree - entails the construction of a hierarchical clus-
ter tree. In this tree, every cluster has exactly one parent that has a more
general label than that of the child. This step also consists of tree pruning
through which similar clusters (siblings or children) are merged in order to
increase accuracy.
There are a couple of variations of this algorithm that are worth mentioning
namely TDC [20] and the one presented by Vikram et al. [17].
The major difference between TDC and FIHC is that TDC uses only closed
frequent termsets for building the initial clusters. The reasoning behind this
change is that closed termsets can dramatically reduce the size of the final
termsets while managing to retain completeness since all non-closed termsets
are always included in the closed ones. Also, another important change is that
the determination of the minimum support parameter used in obtaining the ini-
tial frequent termsets is determined automatically so that it is maximized while
covering the entire document set.
In [17], the hierarchical clustering algorithm works similar to TDC only that
they use generalized closed frequent itemsets and use Wikipedia as an ontology
for improving the document representation. Though this gives better clustering
results it is slower than the score based approach.
2.2 Frequent Itemset Mining
Instead of using Apriori to mine frequent itemsets, for a faster implementation
of FIHC a better solution seems to be using a closed frequent itemsets mining
algorithm such as CloStream[18] that can also handle stream data (a sequence
of transactions that arrives in a timely manner).
CloStream is a frequent itemset mining algorithm that efficiently addresses
some of the challenges posed by stream data such as speed of processing a
new transaction, the support for incremental mining. Moreover, CloStream only
mines closed frequent itemsets which have been shown to be more relevant for
analysis [15] and they do not lead to the loss of information [7]. A closed itemset
represents a frequent itemset for which there is no frequent itemset that is a
superset of the original itemset having the same support.
In order to efficiently store mining information from previous transactions
which can then be used to update the closed frequent itemsets as a new trans-
action is added or removed, the algorithm uses two data structures stored in
memory, namely a ClosedT able and a CidList, but also a temporary hash table
generated and used on an update over the transactions. The ClosedT able con-
tains information about the found closed itemsets. Each entry consists of three
fields: Cid - uniquely identifies a closed itemset, CI - contains the itemset and SC
- records the support of that closed itemset. The CidList is used to maintain an
association between items and closed itemsets where they are present. It consists
of two fields namely Item field which identifies the item and a CidSet field which
contains a set of the id’s of all the associated itemset from the ClosedTable. Also,
the algorithm uses two methods for handling what happens when a transaction
is added or removed [19]: CloStream+ and CloStream-.
While there is a great advantage with CloStream since it maintains the closed
frequent itemsets without rescanning any of the original transactions, the size of
the ClosedT able increases exponentialy as it stores every itemset derived from
the document set (not just frequent or closed ones).
As a different approach DCI Closed [10] is a non-incremental closed frequent
itemset miner that boasts a reduced memory footprint and excellent execution
times compared to its peers. The algorithm is based on a divide-et-impera ap-
proach that makes use of a bitwise vertical representation of the database. Due to
its efficient utilization of memory space, it can handle very low support thresh-
olds well. After two passes over the dataset, it finds frequent singletons and
generates the vertical bitmap of those singletons relative to the transactions
they appear in so that for each item singletoni ∈ setOf AllF requentSingletons
a binary array is created where a bit bitj is only if singletoni is present in
transaction j. Because of this, every closed frequent itemset found can easily
be associated with all the transactions that contain it. The detection of closed
frequent itemsets is done by closure climbing in that the algorithm tries to find
unique order preserving generators, determine their closure and then try to find
other generators as supersets of that closure.
3 Problem Description and Proposed Solution
We seek to improve a casual user’s informational gain from navigating the web by
making use of that user’s experience coupled with that of its peers. Practically, we
are interested in designing an application that can keep track of the documents
the user visits in a seamless way, can organize that information and present it to
the user in a helpful and easy to navigate fashion while also enabling that user
to share his experience and use the shared experience of others.
In a way, we seek to stimulate cultural evolution between users while also
improving for a user the efficiency of using his past experience.
Another aspect to consider as part of our problem is a way to automatically
differentiate without receiving explicit input from the user to whether a visited
web document contains information with any relevance or not. For example
linking pages, or spam pages should not be considered important.
3.1 Problem Requirements
Since there is an increasing trend of migrating from local space to location in-
dependent space such as it is the case with Storage-as-a-Service services such as
DropBox or OneDrive, our application should be able to be location independent
in that every user would have a profile and for that profile all information which
the user wants to have organized or be shareable should be stored online. This
would also alleviate the need for storage and processing resources from the user
to the application’s system. Moreover, it would enable the possibility of sharing
user experiences since all the information would be centralized.
Considering we are interested in being able to give the user’s the best of
their web experiences, the application needs to be able to filter out less impor-
tant experiences. This implies having a feedback measure associated with each
document in the document set that make up a particular experience, which can
be automatically determined by the application.
Still, the biggest requirement of our application revolves around the way it
handles web document clusterization. First of all, the application should be able
to extract the valid text content from, but not limited to, any given HTML
or PDF document; support for formats such as DOC, DOCX, RTF and other
could be included in a future version. Then it should be able to generate a
summary for that text which would be used for really fast results. So based on
the user preference of allotted time for generating the cluster hierarchies, either
the original text or the summary will be used. Secondly, it would be ideal if
the application would be able to find relevant mining information related to the
terms in each document in a timely fashion. Thirdly, the clustering algorithm
should not expect any user supplied parameters such as number of clusters or
minimum support.
Concerning the algorithm, it should also be fast, scalable to the size of the
document set and as accurate as possible while also allowing overlapping. The
reason we want to allow overlapping is because some documents should inher-
ently belong to multiple clusters. For example “frozen” may belong to phase
transitions, movies, music, digital animation etc. Lastly, the algorithm should
be able to supply meaningful and readable labels for its clusters. Stemmed terms
like those used in FIHC or TDC are not entirely qualified as readable.
3.2 Proposed Solution
For our initial version of the application we present a Google Chrome extension
that allows the user to create an account and authenticate, thus associating a
profile with his web experience. The user can manage via the extension his web
experiences as sessions. For example, a session might be called linguistics and
the user would make it the current session while navigating the web in search
of relevant documents related to language and language theory. Then he might
switch to a session called casual which would track documents related to casual
browsing such as meme pages.
The tracking of the documents is done seamlessly while the extension is
active, by sending a notification to the server with the session name and URL
of the web documents loaded by the user. When the user wishes to see his web
experiences organized he opens the search page of the extension and he can select
the session for which he wants his experiences as well as various filters for that
experience such as time interval or based on feedback measurements.
The application should also be able to facilitate the sharing of web expe-
riences between willing users. This is basically done by having the option of
creating or subscribing to an existing community. While the user browses under
a community, all his experiences are shared with that community.
Architecture Our application has two components: the client application which
is the Google Chrome extension and the web server powering the extension.
Fig. 1. Web Experience Enhancer Application Architecture
The client application is in the form of a Google Chrome extension that
makes use of the Chrome API to track the web documents the user accesses as
well as the time he spends on such a document in a browsing session.
The extension is implemented entirely with HTML5, JavaScript and CSS3.
This brings about various advantages compared to developing a native applica-
tion such as:
• perfect integration with the browser technologies;
• the ability to create a rich and easy to use interface comparable to Flash like
interfaces;
• the ability to use multiple threads thorough web workers so as to maintain
the browser responsiveness while generating the output from the clustering
results;
• can easily parse JSON messages with minimum overhead which is most useful
for the communication between our client and server components.
Regarding the web server, since we wish for our application to be highly
accessible in that it will be location and user machine independent (so it would
not matter if the user is using a thin client or not) while also being elastic
enough to provide excellent performance and make use of resources as efficiently
as possible, seems that using a cloud platform for deploying our application’s
web server is a reasonable choice, especially since this component is the real core
of the application. The web server’s importance can be easily explained given the
fact that it handles almost everything (creation and storage of user profiles, user
sessions, communities and web experiences the database, clustering, extraction
to text content, document pre-processing, summarization and pattern mining).
Having considered these, we have chosen Heroku [12] and MongoLab [9] as our
initial deployment solutions. Regarding the technologies used for implementing
the web server, we decided in using NodeJS coupled with a few Java processes
and MongoDB as a database solution (see Figure 1).
Concerning NodeJS, it is a different approach to the stateless request / re-
sponse paradigm that allows for the developing of real-time web applications
employing push technology over web sockets (it makes possible for two-way con-
nections). Also it is innovative in the sense that uses a non-blocking, event-
driven I/O model which makes it possible to remain lightweight and efficient
when dealing with data-intensive applications and it is capable of dealing with
a high number of simultaneous connections with high throughput (highly scal-
able). Also, since it is based on JS, we can adopt JS across the entire stack which
means unifying the development language as well as the data exchange format
since both NodeJS and client JS can natively parse JSON messages.
The reason for also employing Java is due to NodeJS’s weakness relating
to CPU intensive tasks which could impair the very advantages NodeJS can
provide. That is why NodeJS will execute Java processes for each of the CPU
bound tasks such as clustering, closed frequent termset mining, summarization
and text processing.
Regarding MongoDB, we choose it since it is one of the leading NoSQL
databases designed for handling big data that features among others document-
oriented storage with documents stored in BSON format (can be seen as a giant
array of JSON objects), schema free (which is a great advantage in iterative
development), replication, auto-sharding, map/reduce support.
Still, the really great thing about using NodeJS with MongoDB and JS for
the client component is that JSON is the message exchange format all around
which alleviates the need for data conversion and constraint validation.
Filters and feedback measure The application allows for the retrieval of a
hierarchical cluster created only from documents of the user session that satisfy
some given filters. The filters we proposed for such cases are temporal, related
to popularity (the number of visits for a document) and user feedback value.
The latter is a measure of positive feedback from the user interacting with
the document over several visits. According to [14], pages that get allocated 2
or more minutes are usually ”good pages” and pages that are closed in just a
few seconds are ”bad”, with pages usually being useful if more than 30 seconds
pass (see Figure 2).
Fig. 2. Probability of leaving a page in relation to elapsed time [14].
As such the extension will call the server once when a new visit has occurred,
and if the user stays more than 30 seconds on the page, it will call the server
again so as to register a positive feedback for the page. The feedback value itself
is a binary value where 0 means not received positive.
When a clustering request occurs the feedback of a document is calculated
according to the following function:
P
vi ∈Vj fvi
F eedback(docj ) = (1)
NVj
where fvi is the feedback value for a visit vi of docj , Vj represents all the visits
for docj and Nvj is the total number of visits for that document.
Processes and services used by the web server Besides the main node
process which runs the server and listens for incoming requests, the application
makes use of four Java processes and a commercial web service for extracting
text content from an HTML document. In our first implementation we made use
of the free web service provided by AlchemyAPI which specializes in analyzing
unstructured content. Regarding the four emplyed Java processes these are:
1. PDF Text Extractor - makes use of the PDFBox library from Apache [16];
2. Text Summarizer - extracts a sentence based summary using a specialized
algorithm such as LexRank [4];
3. Text Processor - extracts an “improved” feature vector from the document;
4. Clustering Processor - responsible for closed frequent termset ming and hi-
erarchical document clustering.
3.3 iFIHC - improved Frequent Itemset Hierarchical Clustering
Improved preprocessing of the text documents The preprocessing stage
of any document clustering algorithm is extremely important and can profoundly
affect the result of the clustering. All preprocessing and selection steps in doc-
ument clustering try to eliminate “noise” from the texts. Basically they try to
minimize the existence of particles of text that that do not carry useful informa-
tion while trying to reduce the dimensionality of the input data by finding and
eliminating informational redundancy. This is usually done though stop word
removal, stemming and vectorization.
Though these steps are standard with most document clustering algorithms,
we suggest modifying the part related to stemming with a combination of lemma-
tization and stemming especially since our objective is to obtain a friendly and
easy to navigate hierarchical tree where each node has as label a series of terms.
Though stemming and lemmatization are both cases of word normalization
which try to reduce inflectional and related forms of a word to a common base
form (cars => car), they differ quite a lot. Stemming uses a heuristic for chopping
off the end of the words and often includes the removal of derivational affixes. On
the other hand, lemmatization makes use of a dictionary and the morphological
analysis of the words in hope of returning the dictionary form of a word which
is also known as a lemma.
Stemming reduces words that are related derivationally while lemmatization
only reduces the inflectional forms of a given lemma. In other words stemming
would reduce something like “evolved” and “ponies” to “evolv” and “poni” while
lemmatization would result in evolve and pony (see Figure 3).
Fig. 3. Difference between lemmatization, stemming and combining them
Still, using a lemmatizer usually doesn’t result in much dimensionality reduc-
tion since it would keep the verb ”cook” and the nominalized noun ”cooking”
distinct whereas they could be reduced to the same base form namely ”cook”
which is the shortest. That is why we propose to reduce all the words to their
lemma form while associating them with their lemma stemmed form. This way
we can reduce words to a unique valid root form. The lemmatization is done
based on the WordNet dictionary for the English language [13].
Custom FIHC There were a lot of drawbacks with using the clustering steps in
the original algorithm considering our application requirements (some degree of
overlapping, accuracy, improved readability, improved speed, etc). That is why
we propose using a modified version we call iFIHC that uses DCI Closed instead
of Apriori and our combined stemmer and lemmatizer approach.
Cluster Construction. The first part is similar to FIHC with a slight difference
in that the documents that do not meet any filters supplied by the user such as
popularity, visiting time or feedback value are removed along with any closed
termset that remains without any associated documents.
Concerning the disjoint step our custom implementation, instead of using the
same Score function that needed the cluster support parameter we decided to
use the equation from [17], namely:
X docj × tf (tk , docj )
Score(Ci , docj ) = , (2)
length(Ci )
tk ∈Ci
where tf is an augmented term frequency used for preventing a bias towards
larger documents and is calculated based on the function:
0.5 × f req(tk , docj )
tf (tk , docj ) = 0.5 + . (3)
max{f req(w, docj ), w ∈ docj }
Based on the values of the new Score function, we only keep the duplicates
that are above or equal to the mean of all scores for docj . After removing the
less relevant duplicate documents and all empty clusters, we try to improve the
quality of the cluster labels by removing all irrelevant terms.
We resort to this in order to further eliminate noise since the closed frequent
itemsets are mined by presence and not TF or TFxIDF. As such for every cluster
we calculate the mean of the augmented term frequency values and eliminate all
those below it. This is done by repeatedly eliminating the terms below the mean,
until the difference between the maximum value and the mean is smaller than
δ standard deviations. In our experiments setting δ to 1.2 yielded good results.
This step could also be applied before duplicate removal as well.
Tree Building. Our implementation works in a top-down fashion starting from
the first level with clusters containing the smallest number of terms and adding
new nodes as the length of the cluster label increases.
A very important difference between this phase and the one in the original
FIHC is that when for a cluster we find a parent, a copy of that cluster will be
assigned to the parent. This is done so we really have overlapping not just for
documents but also for entire clusters. After the cluster tree has been constructed
each cluster obtains the document indexes of the children. These indexes will be
used with our experiments concerning the quality of the clustering processes.
4 Experimental results
For testing the performances of the algorithms used, we ran them on local plat-
form based on an Intel i5 430M processor with 8GB of RAM and a 64-bit Win-
dows 7 operating system.
Regarding the minimum support for mining the closed frequent termsets, we
choose a minimum support of 6% for our tests.
In evaluating the cluster quality we used as comparison the values obtained
by [17] since they are the most recent. As such we have tested the algorithm out
on the same Re0, Wap, and Ohscal datasets1 .
Regarding the way we can evaluate clustering quality for a hierarchical docu-
ment clustering algorithm that allows overlapping, we use the same measurement
functions as for the original FIHC [5], more specifically by overall F-measure
given by the following equations:
nij
Recall(Ki , Cj ) = (4)
|Ki |
nij
P recision(Ki , Cj ) = (5)
|Cj |
2 × Recall(Ki , Cj ) × P recision(Ki , Cj )
F (Ki , Cj ) = (6)
Recall(Ki , Cj ) + P recision(Ki , Cj )
X |Ki |
F (C) = max {F (Ki , Cj )} (7)
|D| Cj ∈C
Ki ∈K
where nij is the number of documents assigned to cluster Cj that are also present
in the natural class Ki . The function F (C) represents the overall F-measure of
the final cluster tree C, as the weighted sum of each of the maximum F-measures
for each natural class and cluster.
As it can be seen, the iFIHC algorithm behaves far better than FIHC in
terms of accuracy while improving the readability of the end result. For testing
the execution speed we experimented with a scaled up Re0 dataset and we only
evaluated the clustering and tree building steps (see Table 1).
It seems iFIHC scales much better and is a lot faster than the original FIHC
which is what was needed by our application (see Figure 4). This is mainly due
to the reduced dimensionality of the initial cluster set generated from the closed
frequent term sets as well as due to the truncation and elimination of clusters
and cluster terms during the cluster construction phase.
1
http://glaros.dtc.umn.edu/gkhome/cluto/cluto/download
Dataset FIHC iFIHC
Re0 0.529 0.586
Ohscal 0.325 0.461
Wap 0.391 0.516
Table 1. Cluster quality comparison.
Fig. 4. Clustering speed comparison between FIHC and iFIHC.
5 Conclusions
Given the high need to filter and manage all the user relevant information avail-
able on the web we proposed a useful application in the form of a browser exten-
sion that can help the user by keeping track of his web experiences, organizing
them in an efficient manner and filtering them out according to user specified fil-
ters. Moreover the application allows for the enabling of the users to share their
web experiences stimulating collaboration where and each of their experiences
with the web documents are seamlessly rated by the extension accordingly in
relation to the interest they show.
We have provided a viable architecture for the application that takes ad-
vantage of cloud computing and available web services in an efficient way (the
application is location and machine independent) and uses a custom approach
to hierarchical document clustering.
Our approach improves the readability of the cluster labels by combining
stemming and lemmatization and improves the overall clustering process so that:
• there is no need for a predefined number of clusters;
• the clustering process is really fast and scalable compared to previous ap-
proaches while maintaining accuracy;
• allows a larger degree of relevant overlapping;
• provides more meaningful labels with less redundant terms.
In order to achieve the qualities mentioned we modified the original Frequent
Itemset Hierarchical Clustering algorithm by replacing Apriori with an efficient
and suitable closed frequent itemset mining algorithm such as DCI Closed and
by implementing an improved version of the clustering steps called iFIHC that
is comparatively faster and at least as accurate as the original.
As future prospects, we will focus on incremental clustering, the usage of ses-
sions, in particular search queries that lead to the visited web pages as additional
information, the usage of a reference ontology, and to match pages with it, for
a better understanding of the content. The application can be further extended
in the sense of web experience sharing as well as in generating even more useful
and readable labels for the clusters by replacing them with the greatest common
set. Practical application of the work will be analyzed and extended from a third
party perspective (e.g., e-commerce vendors, search engine providers, etc.).
Acknowledgment
The research presented in this paper is supported by the following projects:
“KEYSTONE - semantic KEYword-based Search on sTructured data sOurcEs
(Cost Action IC1302)”; CyberWater grant of the Romanian National Authority
for Scientific Research, CNDI-UEFISCDI, project number 47/2012; clueFarm:
Information system based on cloud services accessible through mobile devices,
to increase product quality and business development farms - PN-II-PT-PCCA-
2013-4-0870; MobiWay: Mobility Beyond Individualism: an Integrated Platform
for Intelligent Transportation Systems of Tomorrow - PN-II-PT-PCCA-2013-4-
0321.
References
1. Agrawal, R., Srikant, R.: Fast algorithm for mining association rules. In J. B.
Bocca, M. Jarke, and C. Zaniolo, editors, Proc. 20th Int. Conf. Very Large Data
Bases, VLDB, pp. 487499. Morgan Kaufmann (1994)
2. Bezdek, J.C.: Pattern Recognition with Fuzzy Objective function Algorithms.
Plenum Press, New York (1981)
3. Carpineto, C., Osinski, S., Romano, G., Weiss, D.: A survey of Web Clustering
Engines. ACM Computing Surveys, Vol. 41, No. 3, Article 17 (2009)
4. Erkan, G., Radev, D.R.: LexRank: graph-based lexical centrality as salience in
text summarization. J. Artif. Int. Res. 22, pp. 457-479, 1 December (2004)
5. Fung, B.C.M., Wang, K., Ester, E.: Hierarchical document clustering using fre-
quent itemsets. International Conference On Data Mining (2003)
6. Jain, A.K., Dubes, R.C.: Algorithms for clustering data. Prentice Hall (1988)
7. Ji, L., Tan, K., Tung, A.: Compressed Hierarchical Mining of Frequent Closed
Patterns from Dense Data Sets. IEEE Trans. on Knowledge and Data Engineer-
ing 19(9), pp. 11751187 (2007)
8. Kwale, F.M.: A Critical Review of K Means Text Clustering Algorithms. In-
ternational Journal of Advanced Research in Computer Science, Vol. 4, No. 9,
July-August (2013)
9. Leonard, A.: Migrate MongoDB Database to Cloud. In Pro Hibernate and Mon-
goDB, pp. 283-296. Apress, (2013).
10. Lucchese, C., Orlando, S., Perego, R.: DCI Closed: A Fast and Memory Ef-
ficient Algorithm to Mine Frequent Closed Itemsets. Proceedings of the IEEE
ICDM Workshop on Frequent Itemset Mining Implementations, Brighton, UK,
November 1 (2004)
11. Marin, A; Pop, F.: Intelligent Web-History Based on a Hybrid Clustering Algo-
rithm for Future-Internet Systems, Symbolic and Numeric Algorithms for Sci-
entific Computing (SYNASC), 2011 13th International Symposium on, pp. 145-
152, 26-29 Sept. (2011)
12. Middleton, N. and Schneeman, R.: Heroku: Up and Running (1st ed.). O’Reilly
Media, Inc, (2013)
13. Miller, G. A.: WordNet: a lexical database for English. Communications of the
ACM, 38(11), 39-41, (1995).
14. Nielsen, J.: How Long Do Users Stay on Web Pages. Web Document. 15 June
2014. http://www.nngroup.com/articles/how-long-do-users-stay-on-web-pages/
15. Pasquier, N., Bastide, T., Taouil, R., Lakhal, L.: Discovering Frequent Closed
Itemsets for Association Rules. Proceedings of the 7th International Conference
on Database Theory, Jerusalem, Israel, pp. 398416 (1999)
16. PDFBox, Java PDF. ”processing Library.” Link: http://www. pdfbox. org,
(2014).
17. Shankar, R., Kiran, G., Vikram, P.: Evolutionary Clustering using Frequent
Itemsets. In Proceedings of StreamKDD’10, July (2010)
18. Show-Jane, Y., Yue-Shi, L., Cheng-Wei, W., Chin-Lin, L.: An Efficient Al-
gorithm for Maintaining Frequent Closed Itemsets over Data Stream. Next-
Generation Applied Intelligence, Lecture Notes in Computer Science Volume
5579, pp 767-776 (2009)
19. Show-Jane, Y., Cheng-Wei, W., Yue-Shi, L., Tseng, V.S., Chaur-Heh H.: A fast
algorithm for mining frequent closed itemsets over stream sliding window. IEEE
International Conference on Fuzzy Systems (FUZZ), pp. 996-1002, 27-30 June
(2011)
20. Yu, H., Searsmith, D., Li, X., Han, J.: Scalable Construction of Topic Directory
with Nonparametric Closed Termset Mining, In Proc. of Fourth IEEE Intl. Conf.
on Data Mining (2004)