WSEAS TRANSACTIONS on COMPUTERS Chhavi Rana, Sanjay Kumar Jain

Building a Book Recommender system using time based content

filtering
CHHAVI RANA
Department of Computer Science Engineering, University Institute of Engineering and Technology,
MD University, Rohtak, Haryana, 124001, INDIA.
[email protected]
SANJAY KUMAR JAIN
Department of Computer Engineering, National Institute of Technology, Kurukshetra,
Haryana,136119, INDIA.
[email protected]

Abstract: - Recommender System are new generation internet tool that help user in navigating through
information on the internet and receive information related to their preferences. Although most of the time
recommender systems are applied in the area of online shopping and entertainment domains like movie and
music, yet their applicability is being researched upon in other area as well. This paper presents an overview of
the Recommender Systems which are currently working in the domain of online book shopping. This paper
also proposes a new book recommender system that combines user choices with not only similar users but other
users as well to give diverse recommendation that change over time. The overall architecture of the proposed
system is presented and its implementation with a prototype design is described. Lastly, the paper presents
empirical evaluation of the system based on a survey reflecting the impact of such diverse recommendations on
the user choices.

Key-Words: - Recommender system; Collaborative filtering; Content filtering; Data mining; Time; Book

1 Introduction user navigation history and his similarity with other

Internet and World Wide Web is changing the way users. Recommender System gives a list of
people live and communicate with each other. recommendations to the user which is an attempt of
Together with this, there is lot of information predicting user’s preferences. A website using a
bombardment on the user who wants to access some recommendation system can more effectively
information on the internet. In the midst of this provide a user with useful and relevant suggestion
complex environment of web, Recommender that could fulfil his current information requirement.
System (RS) serves as an agent that helps user in As such, these websites have an edge over others in
getting the relevant information. With the growth of gaining customer loyalty as well as long term
economy and advent of new technology, lots of partnership.
people are using internet as a source of information The term recommendation system was first
for making comparative analysis of products that introduced by Resnick and Varion [13] to generalize
they would like to buy online. As such lot of the concept of collaborative filtering [7]. They
websites are also offering books as one of the implemented the first recommender system using
product and some sites are dedicated websites for collaborative filtering technique. Since then the term
online book shopping People are increasingly using is being used by different researchers and is
internet to choose the best deal to spend their money explained in various ways. The most common
with a widely available list of authors and genres. technique used for building recommendation system
In this competitive market, lots of vendors is collaborative filtering, so much so that both these
are employing different strategies to attract terms are used interchangeably by many
customers. The customers are bombarded with researchers. There are few other techniques also
information on the internet, finding relevant among which are used for making recommendations such
which is a dotting task. Recommendation systems is as content based filtering, demographic and
one such technology that help user by providing knowledge based technique but they are not so
them with relevant information based on their widely applied. Collaborative filtering is more
profile. A user profile is generated on the basis of successful applied technique, prominent example of

E-ISSN: 2224-2872 27 Issue 2, Volume 11, February 2012

 WSEAS TRANSACTIONS on COMPUTERS Chhavi Rana, Sanjay Kumar Jain

which is amazon.com. In this technique, the user is Section 4 describes the evaluation of the proposed
given recommendation on the basis of similar user system with respect to diversity and temporality.
profiles which are calculated through various Finally, section 5 provides concluding remarks.
measures. Mostly Pearson correlation coefficient,
Cosine similarity measure and Euclidian distance
are taken as widely applied similarity measures. 2 Current Recommender System
A number of recommender systems are Now a days Recommendation systems are applied
applied in various domains on the internet and each to a wide range of domains, though initially their
one of them tries to accurately predict user focus primarily was on the entertainment domain
preference. Also, various researchers studied and online shopping sites. Currently there use is
different facts of recommender system like being extended to social network sites, e-leaning
semantics, scalability etc. Another dimension which and travel domain. The focus area of this paper is
is being highlighted lately is the temporal dimension related to book and as such we present here an
that incorporates changes in the recommendation overview of the Book Recommender Systems. They
process over a period of time [8]. It is being help users in managing their reading list by learning
observed that with time user preference changes as their preference. There are two category of such
well as the system content also gets updated [3]. system, one which gives a list of recommendation
Thus, the recommendation models that predict user based on user profile in a library automation system
preference needs to be updated by refreshing the and others which tells user specifically what should
user profile at regular intervals to present relevant he read next according to the current requirements.
recommendation for users. The effect of temporal There are sites like whichbook.net, what should I
dynamics in collaborative filtering [8, 9] is being read next, lazy library, library think etc., each of
studied lately, but its effect on content based which uses a particular strategy to fulfil user current
filtering [2] is yet to be taken up. requirement. A number of such recommender
In this paper, we have developed a book systems are described hereby.
recommendation system that is based on content Whichbook.net allow users to select a
based recommendation technique and takes into particular category, plot and settings for giving
account the choices of not only similar user but all consequent recommendations. Thus, the focus of the
users to predict new recommendations for the user. system is on users current information need. Also
This approach combines user choices with the most this system allows user to select a particular mood
widely visited links over a period of time and thus like happy, funny, unusual and unpredictable and at
gives a set of diverse as well as most widely used the same time allows them to change their search
preferences. The paper also analyzes the effect of criteria between moods and plot setting. The
temporal factor on content based recommendation. diversity quotient of such recommendation is
Though a number of books Recommender System narrowed within the unpredictable option and
already exist, but none have so far implemented the therefore user could find the system repeating
time factor on content based recommendation. We similar books with passage of time.
implemented a system which will use a counter for WhatshouldIreadnext.com uses
each item that gets updated with time in relation to collaborative filtering technique and give
other items and combined it with content based recommendations to the user by using an author or
recommendations. As such, the paper presents a title of a book and matching it with others reading
completely new outlook on the effect of time factor list. Such technique generally makes an assumption
on the recommendation process in a content based that similar users have similar preferences [4]. Thus,
environment. Though collaborative filtering is the users having some books common in their
widely used methodology in most of the reading list can be given recommendations of others
recommendation systems, content based technique books which are not in their reading list. Such
can have leverage where there are some preferences could get stale after a period of time as
characteristic values that represent item content and similar books will be repeatedly given as
this is the case with books. recommendations.
Our proposed system ensures that Lazylibrary.com search books with topic
recommended items remains relevant to the wise listing and have a limit on size of book. Thus
changing user preferences. In Section 2 literature the system uses content based filtering approach that
survey of existing book recommendation system is chooses books with similar content [1]. The novelty
presented. Section 3 introduces the overall design about this system is the restriction on the number of
and implementation of the content based RS.

E-ISSN: 2224-2872 28 Issue 2, Volume 11, February 2012

 WSEAS TRANSACTIONS on COMPUTERS Chhavi Rana, Sanjay Kumar Jain

page of the book which can also be a stumbling implements content based filtering within the
block in giving relevant recommendations. similar category of books to present
Librarything.com is a completely different recommendation. The drawback of the system lies
system which is best known for cataloguing book in its ability to handle diverse recommendation.
into category for each user. It uses a hybrid
technique that combines recommendation that are
based on library of congress categories and what 3 Proposed System
other similar uses have in their reading lists. It also The book recommender system implemented till
includes links to a number of books stores as well now utilizes either similar user interest or item
[11,14, 5]. The major drawback is that it assumes characteristic for matching it with user profile for
that users have already looked at all the books by giving future recommendation. Our recommender
same author and thus select something completely system adds another important dimension to this
different which could be irrelevant for the user. which is the temporal dimension. This dimension
Abebook Corporation which own the world takes into account the number of time an item is
largest online book store have also developed and liked by the user over a period of time. A counter is
launched a new online book recommender system stored for each item which is updated whenever a
called Bookhints.com based upon title owned, read user checks that items in its favourite links. The
and enjoyed by members of Librarything.com, the greater the counter for any particular book, the more
online cataloguing site. It uses the database of are the chances of it being recommended. Moreover
library thing and generates recommendations based instead of using counter rating of similar user, our
upon titles found on the catalogues of similar users algorithm uses counter ratings of all users that
who also own the book originally sought. “It is the improves diversity of the recommendation process.
first time that unique information added by an Over a period of time, the counter for different items
online community has been directly applied to will change in relation to other items which could
retailing”, said Boris Watz, COO of Abebook. either become more favourite or less favourite in
Another book recommended system comparison to each other. Thus, the overall system
Booklamp.com uses a novel technique to predict reflects an updated favourite list with respect to a
user preference. It matches readers to a book by particular category at any given point of time. User
analyzing the writing style, perspective and preferences are incorporated through content based
description of the whole book together and thus filtering where only a particular category of book
allows users to find books with similar tones, action will be recommended. The user will be given a
level and dialog level. It removes outside influences search option as well a category listing to choose a
such as advertising, author marketing or social book and once it chooses a category or book title,
networking. As the system has grown, it has also recommendation will be enlisted. The timely
incorporated human feedback and mixed it with updation of recommendation is a very important
computer analysis. It allows user to match books aspect as user need as well as the system content
with far greater details than any other such system. keeps changing with time [3]. Thus, there should be
Goodsreads.com, a social network for book a mechanism to precisely incorporate these changes
lovers was started in 2006. Initially it does not have into the recommendation process. This will give
a recommendation system embedded in it, but lately user more relevant, diverse as well as updated
it acquired discoverread.com which has a strong information which establishes the effectiveness of
recommendation algorithm. Discoverread.com Recommender System.
analyzes book trends and give recommend on the
bases of user rating by matching it with friends and
similar users [13]. The major drawback with the 3.1 System Overview
system is that as the system becomes popular, more The proposed approach is very simple and it
and more irrelevant ratings are added which could effectively improve the recommendation process by
harm the accuracy of recommendations. incorporating diverse as well as updated
Bookexplorer.com is another recommendations by maintaining a simple counter
recommendation engine in which the books are for each item. The recommendation process is
arranged category wise for user and there is also a divided into four categories:
search option for users. Once user selects a book (a) Gathering of useful web data
from a category or from the search option, lists of (b) Pre-processing of the collected data
recommended books are also presented to the user. (c) Interpretation of the web data
The recommendation system is fairly simple as it (d) Decision making or Final Recommendation.

E-ISSN: 2224-2872 29 Issue 2, Volume 11, February 2012

 WSEAS TRANSACTIONS on COMPUTERS Chhavi Rana, Sanjay Kumar Jain

certain domain like books, e-learning and Academic

When user sends request for a particular book or research content based filtering can be more useful
author in the form of a query to the Search Engine, than collaborative filtering [10]. This is due to the
the Search Engine return the response by giving fact that such domains offer lots of features that can
URL’s of all the matching links depending upon the be easily used for matching users and items. As such
algorithm used by the Search Engine. The proposed we have focused on content based filtering and
system uses content based filtering algorithm over a added a feature that could combine updated as well
categories of books to search similar books. as diverse information into the system with passage
Moreover, books which are listed as favourites by of time. There is no need to communicate with
maximum number of users are given as other users in such systems [2]. The major strength
recommendations. When user enter query for its of content filtering is that even if the system has
particular search request, the Query is first fewer ratings or none at all which usually is in case
processed to filter out irrelevant data and then of most of the new users, recommendations can be
analyzed to discover interesting navigation patterns made. The only requirement is to have some
and relation between books category and counter information about each item which is easily
value of items in that category. A response available in case of book recommender system.
repository is formed which serves as a knowledge
base. After building Knowledge Base, system can
give recommendations based on the navigation and 3.2 Process flow of the System
most visited links by particular user in the featured The personalized book recommender system
domain. The Personalized data is then given as a consists of the following workflow -:
input to the ranking algorithm for personalizing the 1) Firstly user casually uses the system without
result according to the most visited links or registration and collects raw information i.e.
favourite links of the particular user by evaluating information which is available to the every
web mining results. Thus, the System helps the user visitor’s of the website. After analyzing this
to retrieve his/her interested domain by clicking on information, user login to the system and
the list of his/her favourite links with more generates its profile which contain user
convenience. The current recommendation is based personal details and most visited link
on the recent most last selected item by the user. recommended by the system based on the
The content are filtered according to this category URL ranking algorithm (collectively most
and the items with highest number of favourite visited favourite links of all users).
checks at this point of time are given as 2) After the registration is completed and a set
recommendations. Fig 1 shows the overall of general recommendation with most
architecture of the proposed book recommender visited links provided, the system proceeds
system. on to next level. Based on the user
navigation pattern, system analyzes user
behaviour and predict its interest by
performing content filtering on the basis of
this analysis. As a result, system provide
favourite link to the user based on its profile
via personalized web page and
simultaneously update already exiting links
on the basis of periodical temporal update.
3) In the last step, the system again perform
analysis on the previous result and deliver
top n recommendations of that category of
books in which user is most interested in by
filtering the contents available on the system
by performing content filtering.
The result of the overall process workflow is that
Fig. 1. Architecture of Book Recommender System system deliver web page to user which contain its
personal detail, most visited link of the site, user’s
The proposed system tries to give diverse favourite link and contents recommended by the
recommendations to the uses which are also relevant system of his/her interested domain in single
to his current information needs. In the area of customized format.

E-ISSN: 2224-2872 30 Issue 2, Volume 11, February 2012

 WSEAS TRANSACTIONS on COMPUTERS Chhavi Rana, Sanjay Kumar Jain

This figure below depicts overall process of which five focuses on demographic data (gender
workflow for book recommender system. and age) so that results could be categorized. The
other questions were specifically focusing on
recommendation relevancy to the user, accuracy
with temporal changes and diversity.

4.1 Implementation
The system was implemented using ASP.net as front
end and SQL server as backend on an Intel
Core2duo processor with 2GB ram and 300GB hard
disk.

4.2 Participants
Participants were college students who were already
using some book recommendation system and are
requested to use the current system. They
volunteered to take the survey by clicking on the
link on the website. Around 150 respondents
participated in the survey among which 50 girls and
100 boys were there. Each participant belongs to
Fig. 2. Process workflow for Book age group 18-25 years.
Recommender System
4.3 Results
The results were inferred by using descriptive
statistics to generate percentages and further
3.3 Personalized Recommendation analysis of the data gathered from the survey
The process of recommendation can enhance the through qualitative methods to discover major
overall user interface of a particular website. The themes [6].
user interface can further be enhanced if the
recommendations are given in a particular manner 4.3.1 Diversity Impact
that helps user in further navigation. In case of In the first experiment shown in Figure 3, it is the
books, most of recommender system provides a evaluation of the recommendation preferences with
mechanism through which user is presented with respect to diversity. The users are asked to give a
further links to buy recommended books from score to both types of recommendations. The higher
online stores. Other recommender system gives an score indicates the more useful the
option of borrowing such books from library which recommendations are. It can be seen that most of the
could be limited to particular country or city, as is recommendation that comes under diversity were
case with librarything.com. The emphasis suggest more liked by the users. Participants were asked to
that calculating recommendation list is as important explain their preferences regarding diverse
as presenting them in certain manner. Finally some recommendation and content based similar
feedback mechanism could also be incorporated at recommendation, and from the data five participant
the end of the recommendation list that will take preference categories were developed: Diverse,
suggestion from users and thereby improve the Similar, both, neither, and no preference. Of the 74
recommendation process. This will act as an participants that responded to the survey, the
unobtrusive way of getting suggestion from the majority of the members, 54% (40) preferred
users at the end of the whole process. diverse recommendations while only 22% (16)
preferred similar recommendations. The other 24%
(18) of the participants preferred neither, both or
4 Research Design had no preference.
The performance analysis of this system is done
using an online survey to explore the effect of
temporal updation on user satisfaction. The study
also analyses how user reacts to a set of diverse
recommendations. There was no identifying private
data which was sought. There were 10 questions out

E-ISSN: 2224-2872 31 Issue 2, Volume 11, February 2012

 WSEAS TRANSACTIONS on COMPUTERS Chhavi Rana, Sanjay Kumar Jain

45
40 4 Conclusion
35 In this paper, we propose a book
30 Diversity recommender system that uses content based
25 Preffered
approach for recommending item to a particular
20 Similiarity
Preffered
uses. In this approach we embedded a new
15
10 dimension called temporal dimension using a
Others
5 counter for each item which gets update with
0 passage of time and thereby improving the whole
1st 2nd 3rd 4th recommendation process.
Week Week Week Week
The major goal of the current system is to
provide diverse recommendation to the user even if
there are fewer ratings about an item for a particular
Fig. 3. Diversity Effect user. The process also get updated with passage of
time which makes different item less or more
4.3.2 Temporal Impact relevant according to their counter value which gets
In the second experiment, preferences of users over changes with time in relation to other items. The
a period of time were analyzed. The users were literature survey listed a number of recommender
asked to rate the recommendation after using the system but none takes into account this criteria for
system for a period of time on a daily basis and then suggesting diverse as well as updated
after a week they were asked to give score about the recommendation. After a certain amount of time, the
nature of recommendation. Participants were asked accuracy of the recommendation process gets stuck
specific question about whether the and users keep on receiving similar items. Further
recommendation were changing with time and were accuracy can only be increased by taking into
those changes helpful or not. account temporal dimension which is what this
Users were also asked to rate the quality of paper addresses. Result shows that proposed
recommendation on a scale of 1 to 10. Figure 4 architecture for the book recommendation engine
shows that only 20 (26%) participants did not notice provide diverse and temporally updated
any changes in their recommendations while 54 recommendation to the users which are more useful
(74%) participants see a noticeable change in their and relevant.
recommendations daily, weekly or periodically. The current system applies this approach in
Most of the participants 37 (70%) among the 54 the domain of book recommender system which
who noticed the changes, find those changes useful could further be extended to other domains as well.
and give a score 8 and above to the new Further the system can be extended in other
recommendations which were changing with time. dimension that includes incorporating contextual
The above result shows that the newly developed information, taking multiple ratings and providing a
RS adapt to the user changing requirement in a very more flexible recommendation that could range into
effective manner. different domains as well. The performance of the
system and its comparison with the other similar
systems is scheduled for future ongoing work.

References:
60
[1] Baudisch, P., and Terveen, L. Interacting with
50
recommender systems. CHI 99 extended
40 Visble
Changes
abstracts on Human factors in computing
30
Non Visible
systems CHI 99, 1999, pp. 164.
20 [2] Chen, T., and Han, W. Content recommendation
10 Useful system based on private dynamic user profile.
0
Machine Learning, Vol. 4, 2007, pp. 2112-
1st 2nd 3rd 4th 2118
Qtr Qtr Qtr Qtr [3] Chu, W. and Park, S. T. Personalized
Recommendation on Dynamic Content Using
Predictive Bilinear Models. 18th International
Fig. 4. Temporal Effect

E-ISSN: 2224-2872 32 Issue 2, Volume 11, February 2012

 WSEAS TRANSACTIONS on COMPUTERS Chhavi Rana, Sanjay Kumar Jain

WWW Conference, Madrid, Spain, 2009, pp.

691-706.
[4] Davies, J. E. What shall I read next? Developing
tools for reader support. Proceedings of IFLA
General Conference and Council, Vol. 68,
2002.
[5] DeZelar-Tiedman, C. Doing the LibraryThing in
an academic library catalog. Proceedings of the
2008 International Conference on Dublin Core
and Metadata Applications, 2008, pp. 211–211.
[6] Glaser, B.G. and Strauss, A.L. The constant
comparative method of qualitative analysis, In
The discovery of grounded theory: Strategies
for qualitative research. 1967, Aldine de
Gruyter: Hawthorne, New York. P. 101-115.
[7] Goldberg, D., Nichols, D., Oki, B. M., and
Terry, D. Using collaborative filtering to weave
an information tapestry. Communication of the.
ACM, Vol. 35, No. 12, 1992, pp. 61-70.
[8] Koren, Y. Collaborative filtering with temporal
dynamics. Proceedings of the 15th ACM
International Conference on Knowledge
Discovery and Data Mining, Vol. 53, No. 4,
2009, pp. 447.
[9] Lathia, N., Hailes, S., Capra, L., and Amatriain,
X. Temporal Diversity in Recommender
Systems. Changes, 2010, pp. 210-217.
[10] Garden, M., and Dudek, G. Mixed
Collaborative and Content-Based Filtering with
User-Contributed Semantic Features. AAAI
Conference On Artificial Intelligence, Vol. 21,
No. 2, 1999, pp.1307.
[11] Naughton, R., and Lin, X. Recommender
Systems: Investigating the Impact of
Recommendations on User Choices and
Behaviors. ACM RecSys UCERST, Barcelona,
Spain, 2010, pp. 9-13.
[12] OReilly, T. Goodreads vs Twitter: The
Benefits of Asymmetric Follow. OReilly Radar,
2009.
[13] Resnick, P. and Varian, H.R. Recommender
Systems. Communications of the ACM, Vol. 40,
No. 3, 1997, pp. 56–58.
[14] Thomas, M., Caudle, D. M., and Schmitz,
C. Trashy tags: problematic tags in
LibraryThing. New Library World, Vol. 111, pp.
5/6, 2010, pp. 223-235

E-ISSN: 2224-2872 33 Issue 2, Volume 11, February 2012