Effective Cross-Domain Collaborative Filtering using Temporal Domain – A Brief Survey

88 International Journal for Modern Trends in Science and Technology
International Journal for Modern Trends in Science and Technology
Volume: 02, Issue No: 10, October 2016
http://www.ijmtst.com
ISSN: 2455-3778
Effective Cross-Domain Collaborative Filtering
using Temporal Domain – A Brief Survey
Swapna Joshi1
| Prof. Manisha Patil2
1,2 Department of Computer Engineering, Smt. Kashibai Navale College of Engineering, Pune, Maharashtra, India
To Cite this Article
Swapna Joshi and Prof Manisha Patil, Effective Cross-Domain Collaborative Filtering using Temporal Domain – A Brief
Survey , International Journal for Modern Trends in Science and Technology, Vol. 02, Issue 10, 2016, pp. 88-92.
Cross-domain collaborative filtering (CDCF) is an evolving research topic in recommender systems. It aims
to alleviate the data sparsity problem in individual domains by transferring knowledge among related
domains. But it has an issue of user interest drift over time. Along with data sparsity, we should also
consider the temporal domains to overcome user interest drift over time problem to predict more accurately as
per the current user’s interest. This paper surveys few of the pilot studies in this research line and the
methods of how to add the temporal domains in the recommender systems. The paper also proposes possible
extensions of using temporal domains with different contexts in current timestamp.
KEYWORDS: Survey, collaborative filtering, temporal domain, cross-domain, recommender syste
Copyright © 2016 International Journal for Modern Trends in Science and Technology
All rights reserved.
I. INTRODUCTION
The essential problem of Collaborative Filtering
(CF) methods is how to find similar users/items
and how to measure similarities between them.
Thus far, most existing CF methods are
single-domain based, which make predictions
based on one rating matrix. In other words, these
methods can only find similar users/items in a
single domain. However, in many recommendation
scenarios, multiple related CF domains may be
presented at the same time and finding similar
users/items across domains becomes possible,
such that common rating knowledge can be shared
among related domains.
Cross-domain collaborative filtering (CDCF) aims
to share common rating knowledge across multiple
related CF domains to boost the CF performance.
CDCF methods exploit knowledge from auxiliary
domains (e.g., movies) containing additional user
preference data to improve recommendation on a
target domain (e.g. books). While relying on a broad
scope of existing data in many cases is a key to
relieving the problems of sparse user-item data in
the target domain, CDCF can also simultaneously
benefits different data owners by improving quality
of service in different domains. Fig. 1 shows the
general architecture of the cross domain
collaborative filtering system.
Fig. 1: Cross Domain CF architecture
Major Issues in current CDCF system –
A.User Interest Drift Over Time -
User’s interests keep changing over time since it
continuously getting affected by multiple factors
such as moods, contexts, cultures, festivals and
different seasons. An example is a person who does
ABSTRACT

Swapna Joshi and Prof Manisha Patil : Effective Cross-Domain Collaborative Filtering using Temporal Domain – A
Brief Survey
not like animated movies, may like to watch them
in future due to evolving superior 3D animation
technologies with very good visual effects or else if
the person having baby may like to watch it now
which he has never watched before. Another
example is a person buying outfit items of a specific
type may like to buy the outfits of different style if
he/she is relocating to an altogether different
geographical area.
Reviewers express their opinions about a particular
product or a service (or restaurants, hotels, movies,
etc.) through some review sites, social networking
sites like facebook, twitter, etc. or even over blogs.
All these act as sources for opinions.
B. Data sparsity –
In practice, many commercial recommender
systems are used to evaluate large item sets (e.g.,
Amazon.com recommends books and CDnow.com
recommends music albums). In these systems,
even active users may have purchased well under
1% of the items (1% of 2 million books is 20,000
books). Accordingly, a recommender system based
on nearest neighbor algorithms may be unable to
make any item recommendations for a particular
user. As a result the accuracy of recommendations
may be poor.
C. Scalability –
Nearest neighbor algorithms require computation
that grows with both the number of users and the
number of items. With millions of users and items,
a typical web-based recommender system running
existing algorithms will suffer serious scalability
problems.
D. Synonyms –
Synonyms refers to the tendency of a number of the
same or very similar items to have different names
or entries. Most recommender systems are unable
to discover this latent association and thus treat
these products differently. Topic Modelling (like the
Latent Dirichlet Allocation technique) could solve
this by grouping different words belonging to the
same topic.
E. Diversity and the Long Tail –
Collaborative filters are expected to increase
diversity because they help us discover new
products. Some algorithms, however, may
unintentionally do the opposite. Because
collaborative filters recommend products based on
past sales or ratings, they cannot usually
recommend products with limited historical data.
This can create a rich-get-richer effect for popular
products, akin to positive feedback. This bias
toward popularity can prevent what are otherwise
better consumer-product matches.
II. MOTIVATION
Many In most CDCF approaches more stress is
given on referring the ratings from multiple sites of
the related domains. But there is one more problem
in recommender systems that the user interests
keep drifting over a period of time. Hence the
temporal domains should also be considered in the
cross domain recommender systems. As users’
interests keep changing over time and a user is
more likely to be interested in different item
categories at different time, we cannot simply view
the multiple historical buying patterns of the same
user in different historical temporal domains.
III. RELATED WORK
The Cross domain collaborative filtering is an
evolving research topic in recommender systems.
The survey was cried to study two major issues in
current CDCF systems, data sparsity and user
interest drift over time.
[1] was studied to understand the basic
Cross-domain recommender systems. In
cross-domain, we refer multiple auxiliary domains
for user ratings/inputs from related domains and
then use these ratings to transfer the collective
rating matrix to the target domain for
recommendations. For example, to recommend a
movie of 3D animation genre to the end user, the
target domain is 3D animation genre movie. So we
consider multiple input auxiliary domains as
books, music, and other movie genre similar to the
movie subject etc. It used a derived method based
on a Bayesian latent factor model which can be
inferred using Gibbs sampling. This addresses the
challenge of modeling user-interest drift over time
by considering the historical time slice patterns of
the user.
[2] and [3] were studied for dealing with data
sparsity reduction problem. In [3], principled
matrix-based transfer learning framework is used
that takes into account the data heterogeneity.
The principle coordinates of both users and items
in the auxiliary data matrices are extracted and
transferred to the target domain in order to reduce
the effect of data sparsity.
The survey was carried out on [4], [5] and [6] to
study different methods of cross domain

Brief Survey
collaborative filtering with temporal domain.
In [4], the temporal domains are considered. The
ratings provided by the same user at different time
may reflect different interests as those ratings are
provided by different users. The proposed
algorithms are a variant of standard
neighborhood-based (either user based or item
based) CF. The main idea behind the proposed
approaches is to enhance inter-domains edges by
both discovering new edges and strengthening
existing ones. In [5], Factorization model and
item-item neighborhood model are used. In both
factorization and neighborhood models, the
inclusion of temporal dynamics proved very useful
in improving quality of predictions, more than
various algorithmic enhancements
While working in CDCF recommender system,
one critical aspect is how to collect and transfer the
knowledge from different input auxiliary domain to
the target domain for giving the recommendations.
The survey was done on [7], [8] and [9] to
understand the methods of transferring the
knowledge across the domain. These Transfer
learning methods are used to first individually
collect the ratings matrix for each auxiliary input
domain and then transferring the collective ratings
to the target domain.
[10] was studied rigorously which deals with the
cross domains over different sites (domains),
transferring the rating knowledge from these sites
to recommend in target domain by using
knowledge transfer method. Along with using
multiple sites/domain to collect the user ratings, it
also extends the model of [4] for using temporal
domain to deal with the issue of user interest drift
over time. It uses the principle that user has
multiple counterparts across temporal domains
and the counterparts in successive temporal
domains are different but closely related. A series
of time-slices are viewed as related CF domains and
a user at current time-slice depends on herself in
the previous time-slice. Model is built on a
cross-domain CF framework by viewing the
counterparts of the same user in successive
temporal domains are different but related users. If
we can find the unchanged rating patterns (static
components) shared across temporal domains, the
drifting factors of users (changing components) in
each temporal domain can be easily captured.
Bayesian generative model to generate and predict
ratings for multiple related CF domains on the
site-time coordinate system, is used as the basic
model for the cross-domain CF framework which is
extended for modeling user-interest drifting over
time, where a series of time-slices are viewed as
related CF domains and a user at current
time-slice depends on herself in the previous
time-slice.
Table 1 provides an overall evaluation of related
work.
Table 1: Overall Evaluation of Related work
Reference
Number
Paper Name Data
Sparsity
User
Interest
drift over
time
1
Cross-domain
recommender systems
No Yes
2
Can movies and books
collaborate? Cross
domain collaborative
filtering for sparsity
reduction
Yes No
3
Transfer learning in
collaborative filtering for
sparsity reduction
Yes No
4
Cross-domain
collaborative filtering over
time
No Yes
5
Collaborative filtering
with temporal dynamics
No Yes
6
A spatio-temporal
approach to collaborative
filtering
Yes Yes
8
Transfer learning for
collaborative filtering via
a rating-matrix generative
model
Yes No
10
Rating Knowledge
Sharing in Cross-Domain
Collaborative Filtering
Yes Yes
Table 2: Specific Evaluation of Related work (for user
interest drift over time issue)
Paper Name Using
historical time
slice data
Using
Current time
context
Cross-domain recommender
systems [1] Yes No
Cross-domain collaborative
filtering over time [4] Yes No
Collaborative filtering with
temporal dynamics [5] Yes No
A spatio-temporal approach to
collaborative filtering [6] Yes No
Rating Knowledge Sharing in
Cross-Domain Collaborative
Filtering [10] Yes No
Proposed System Yes Yes
IV. PROPOSED WORK
After the rigorous survey, the issue specific to
user interest drift over time needs to be handled
effectively using current time context. As per the
survey of existing work [10], the user interest drift
over time problem in Cross domain collaborative
filtering techniques is handled by introducing the
concept of temporal domain. The historical rating
data is time stamped. With the input historic data

Brief Survey
from the same or different user having similar
interests of the past buying patterns, the data set is
divided in multiple time slices. Each time slice is
considered as one time domain and then the
principles of cross domain filtering are applied on
these multiple time domains. The knowledge is
transferred along successive temporal domains to
benefit the user-counterparts in each domain. A
user in different temporal domains can be treated
as a set of user-counterparts with similar but
different interests. Thus, the ratings provided by
the same user at different time may reflect different
interests as those ratings are provided by different
users. In this current work, only the historical time
domain is being considered by gathering the
historical ratings provided by same or different
user having same taste.
The proposed work take into account the current
time domain context along with historical data.
This can be viewed in multiple contexts/aspects.
Contexts can be added based on the season
changes, location changes, cultural changes,
festival periods ect. By using these current
contextual parameters along with the existing work
of historical time domain, the user
recommendations will be more effective and more
current addressing the current context of the user.
The dataset pre-processing will divide the input
dataset of each year into 12 different slices having
one slice per month as opposed to the 4 slices per
year in the base system. By this more granular
level rating matrix can be generated which is more
aligned with the ever changing user interests. This
can provide better ratings recommendations. Also
the contextual parameters will be applied on these
slices on granular levels so that the season,
holidays, locations will be applied more closely to
the current matching context.
The block diagram in fig. 2 shows the proposed
work. First the dataset will be initialized by
acquiring the data from various sites. Dataset
pre-processing will create the user * movies matrix
in 12 different time slices.
Then the average rating will be calculated by
applying the current contextual parameters and
comparing the historical ratings as per the time
slices. Ten the final ratings recommendations will
be presented to the user.
Fig. 2: Proposed work
V. CONCLUSION
We have given a brief survey of cross domain CF
over Temporal domain and summarized the related
works for different models dealing with the issues
mentioned above. Studied papers are using only
the previous history of either same user or user
groups for recommendation of current time slice.
As a part of future work, we can consider the
current time slice parameters as well depending on
different time aware contexts such as festivals,
seasons, regional and cultural changes etc. as per
current user’s geographic locations. This will
enhance the recommendations more in line with
the current time domain.
ACKNOWLEDGEMENT
This survey would not have been possible
without the support and help of many individuals. I
would like to extent my sincere thanks to all of
them. I am indebted to Prof. Manisha Patil,
Department of Computer Engineering of Smt.
Kashibai Navale College of Engineering affiliated to
Savitribai Phule Pune University for her guidance
and constant supervision as well as all the other
staff members for providing important information
regarding the survey.
REFERENCES
[1] P. Cremonesi, A. Tripodi, and R. Turrin,
“Cross-domain recommender systems,” in Proc.
IEEE 11th Int. Conf. Data Mining Workshops
(ICDMW), Vancouver, BC, Canada, 2011, pp.
496–503.
[2] B. Li, Q. Yang, and X. Xue, “Can movies and books
collaborate? Cross domain collaborative filtering for
sparsity reduction,” in Proc. 21st Int. Joint Conf.
Artif. Intell. (IJCAI), 2009, pp. 2052–2057.
[3] W. Pan, E. W. Xiang, N. N. Liu, and Q. Yang,
“Transfer learning in collaborative filtering for
sparsity reduction,” in Proc. 24th Conf. Artif.Intell.

Brief Survey
(AAAI), 2010, pp. 230–235.
[4] B. Li et al., “Cross-domain collaborative filtering over
time,” in Proc. 22nd Int. Joint Conf. Artif. Intell.
(IJCAI), 2011, pp. 2293–2298.
[5] Y. Koren, “Collaborative filtering with temporal
dynamics,” in Proc. Int. Conf. Knowl. Discov. Data
Mining (KDD), Paris, France, 2009, pp. 447–456.
[6] Z. Lu, D. Agarwal, and I. S. Dhillon, “A
spatio-temporal approach to collaborative filtering,”
in Proc. 3rd ACM Conf. Recommender Syst., New
York, NY, USA, Oct. 2009, pp. 13–20.
[7] R. Salakhutdinov and A. Mnih, “Bayesian
probabilistic matrix factorization using Markov
chain Monte Carlo,” in Proc. 25th Int. Conf. Mach.
Learn. (ICML), Helsinki, Finland, 2008, pp. 880–887.
[8] B. Li, Q. Yang, and X. Xue, “Transfer learning for
collaborative filtering via a rating-matrix generative
model,” in Proc. 26th Annu. Int. Conf. Mach. Learn.
(ICML), 2009, pp. 617–624.
[9] S. J. Pan and Q. Yang, “A survey on transfer
learning,” IEEE Trans. Knowl. Data Eng., vol. 22, no.
10, pp. 1345–1359, Oct. 2010.
[10] Bin Li, Xingquan Zhu, Ruijiang Li, and Chengqi
Zhang “Rating Knowledge Sharing in Cross-Domain
Collaborative Filtering” in IEEE TRANSACTIONS ON
CYBERNETICS, VOL. 45, NO. 5, MAY 2015
[11] B. Li, “Cross-domain collaborative filtering: A brief
survey,” in Proc. 23rd IEEE Int. Conf. Tools Artif.
Intell. (ICTAI), Boca Raton, FL, USA, Nov. 2011, pp.
1085–1086.

Effective Cross-Domain Collaborative Filtering using Temporal Domain – A Brief Survey

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Effective Cross-Domain Collaborative Filtering using Temporal Domain – A Brief Survey (20)

Recently uploaded (20)

Effective Cross-Domain Collaborative Filtering using Temporal Domain – A Brief Survey