A Survey on Challenges and Methods in News Recommendation

Ozlem

Ozg

obek

1,2

, Jon Atle Gulla

and R. Cenk Erdur

Department of Computer and Information Science, NTNU, Trondheim, Norway

Department of Computer Engineering, Ege University, Izmir, Turkey

Keywords:

Recommender Systems, News Recommendation, Challenges.

Abstract:

Recommender systems are built to provide the most proper item or information within the huge amount of data

on the internet without the manual effort of the users. As a speciﬁc application domain, news recommender

systems aim to give the most relevant news article recommendations to users according to their personal inter-

ests and preferences. News recommendation have speciﬁc challenges when compared to the other domains.

From the technical point of view there are many different methods to build a recommender system. Thus, while

general methods are used in news recommendation, researchers also need some new methods to make proper

news recommendations. In this paper we present the different approaches to news recommender systems and

the challenges of news recommendation.

1 INTRODUCTION

The increasing amount of data on the internet makes

harder to ﬁnd what we are really looking for. Even

though the technologies like search engines and RSS

readers help us, it is still hard to ﬁnd the information

we really want to get. On the other hand, we are not

always sure about what we want to get. We can only

search for what we know and we try to ﬁnd some con-

nections to the new information. But this approach of

ﬁnding an item that the user will like mostly depends

on the coincidences, the attention of the user to in-

spect the search results and it requires a lot of effort.

Still there is a high possibility that the user could not

ﬁnding the most suitable item for herself at the end.

Recommender systems are built to help us to eas-

ily ﬁnd the most proper information on the inter-

net. Unlike the search engines recommender systems

bring the information to the user without any man-

ual search effort. This is achieved by using the simi-

larities between users and/or items. There are many

methods to build a recommender system and these

methods can be applied to many speciﬁc domains like

shopping (e.g. Amazon), movies (e.g. Netﬂix) and

music (e.g. Pandora Radio). Since each application

domain has its own speciﬁc needs, the method used

for recommendations differs.

As people are beginning to read news online more

and more, it became a challenge to ﬁnd the interest-

ing news articles. Most of the users spend a lot of

time to ﬁnd an interesting article on a single website

or they just read the front page news which is not ad-

equate. When we consider the different news sources

on the internet, one can spend plenty of time just read-

ing the news. News recommender systems aim to give

the most relevant article recommendations to users ac-

cording to their personal interests and preferences.

Recommending news articles is one of the most

challenging recommendation tasks. The news domain

differs from other domains in many ways. For ex-

ample; the popularity and recency of news articles

changes so fast over time. So focusing on the re-

cency issue becomes more challenging than it is in

other domains. Also some news articles may be con-

nected with each other that the user may want to read

the previous news items related to the one she already

reads or she may want to keep informed about. Only

learning user preferences can be an unsatisfactory so-

lution to news recommendation. This is because the

user may want to read a news article when she is not

really interested in the subject but she thinks it is im-

portant. For example; wanting to read the news about

elections even if she is not generally interested in pol-

itics. Also considering the high number of new arti-

cles published every hour increases the complexity of

other challenges.

In this paper, we summarize the advances in this

very special application domain of recommender sys-

tems which is news recommender systems. The paper

is structured as follows: Section 2 gives an overview

278

Özgöbek Ö., Gulla J. and Erdur R..

A Survey on Challenges and Methods in News Recommendation.

DOI: 10.5220/0004844202780285

In Proceedings of the 10th International Conference on Web Information Systems and Technologies (WEBIST-2014), pages 278-285

ISBN: 978-989-758-024-6

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

of recommender systems. Section 3 summarizes all

the challenges of recommender systems in news do-

main which includes some common challenges with

general recommender systems. Section 4 discusses

the different approaches of news recommender sys-

tems for particular challenges. Section 5, gives the

discussions. And in section 6 the conclusions are pro-

vided.

2 RECOMMENDATION

TECHNIQUES

There are different methods for recommending an

item. Most commonly they are grouped into three

categories: Content-based ﬁltering, collaborative ﬁl-

tering and hybrid. It is possible to categorize the

recommendation techniques differently. For exam-

ple in (Burke, 2002) ﬁve categories are proposed for

recommendation techniques. These are: Collabora-

tive, content-based, demographic, utility-based and

knowledge-based techniques. This categorization of

recommendation methods is based on the background

data included in the system, input data gathered from

online user interaction and the algorithm used for

the recommendation. Since the ﬁrst categorization is

more widely used we will discuss these three cate-

gories. There is also an alternative semantic approach

(Cantador et al., 2008), in which semantic representa-

tions are added on top of other methods, we will not

go into the details of this work in the rest of the paper.

2.1 Content-based Filtering

In content-based recommendations, the properties of

items are used to make recommendations. Items

which have similar properties with the user’s previ-

ous preferences are recommended to the user. Thus,

for this technique it is important to ﬁnd the similarities

between items. For example; to recommend a movie

to the user, the content-based system should know

about the user’s past movie preferences and the simi-

larity between movies. As discussed in (Adomavicius

and Tuzhilin, 2005) this approach has its roots in in-

formation ﬁltering and information retrieval. Some-

times information ﬁltering is used in the same mean-

ing as content-based ﬁltering (Lee and Park, 2007).

2.2 Collaborative Filtering

In collaborative ﬁltering, recommendations are done

by using the other people’s preferences which are

similar to the user’s past preferences. Collaborative

ﬁltering method can be divided into three as memory-

based, model-based and hybrid methods. In memory-

based (also called the neighborhood-based, user-

based, heuristic-based) collaborative ﬁltering method,

user ratings are used to compute similarities between

users/items. By using the statistical techniques it is

aimed to ﬁnd a similar user to the targeted user. Af-

ter the similarities found the recommendations can be

done by using different algorithms. In the model-

based (also called item-based) method, a model is

created for each user by using data mining and ma-

chine learning algorithms. Probabilistic methods can

be used for recommendation predictions. Methods

like Bayesian network and clustering are included in

this method (Sarwar et al., 2001). The Hybrid col-

laborative ﬁltering method uses both model-based and

memory-based methods.

2.3 Hybrid Approach

These approaches use both content-based and collab-

orative ﬁltering. Generally the aim of these kinds

of approaches is to come up with solutions to the

problems which occur with the use of a single ap-

proach. There may be different combinations of us-

ing the two methods together. (Burke, 2002) groups

the hybridization methods into seven and deﬁnes how

different methods can be joined.

3 PARTICULAR CHALLENGES

IN NEWS

RECOMMENDATIONS

For many people building a recommender system can

be perceived as an easy task at ﬁrst sight. But ﬁnd-

ing the proper item to recommend can be a tedious

task that requires access to information about the user,

the items and the general context. Personal prefer-

ences and interests tend to vary on the basis of age,

culture, gender and personality, and they also change

over time. A successful recommender system needs

to address a number of intrinsic challenges that each

constitute a research ﬁeld. In the news domain, be-

cause of the dynamic properties of news items some

challenges have more importance than the others. The

challenges we explain in this section are all related

but not completely speciﬁc to news domain. Most

of these challenges are the general challenges of rec-

ommender systems where some of the speciﬁc chal-

lenges (e.g. recency) may not be an issue in other

domains.

• Cold-start (First Rater, Ramp Up, Early

ASurveyonChallengesandMethodsinNewsRecommendation

279

Rater) Problem: The ﬁrst-rater problem is one

of the most common problems in recommender

system collaborative ﬁltering applications. Basi-

cally, it is the problem that the system cannot rec-

ommend new items if they do not have any clicks

from other users. Or when there is no data about

the completely new user then it is not possible to

make recommendations for her.

• Data Sparsity: The matrices used for collabora-

tive ﬁltering can be very sparse when there are

not enough ratings from users. The possibility of

data sparsity increases if the number of columns

or rows is much higher than the other. For exam-

ple; if the number of items is much more than the

number of users then it requires too many ratings

to ﬁll the item-user matrix. Data sparsity causes a

decrease in the performance of the system.

• Recency: Recency is one the most impor-

tant challenges in news recommendation domain.

Most of the users want to read fresh news instead

of old dated articles. So the importance of news

items decreases in time. On the other hand, some

news articles may be connected with each other

that the user may want to read the previous news

items related to the one she already reads or he/she

may want to keep informed about that subject (Li

et al., 2011).

• Implicit User Feedback: User feedbacks are

quite important to make more precise recommen-

dations. Without explicit feedbacks it may not be

possible to understand if the user liked the arti-

cle she read or not (Fortuna et al., 2010). But it

is not practical for the system to interact with the

user continuously. So the system should be able

to collect implicit feedbacks effectively while pro-

tecting the user privacy.

• Changing Interests of Users: Another key chal-

lenge is predicting the future interests of users for

better recommendations because people may have

changing interests (Liu et al., 2010). For some do-

mains like movie or book recommendations, the

change of user interest happens more slowly. But

for the news domain it is really hard to predict the

changes. Also some people may read the news

not because he/she interested in the topic in gen-

eral but because she found it important.

• Scalability: Recommender systems are aimed to

serve many users, sometimes millions of users

(Das et al., 2007) at a time. Also the number

of items to be recommended can be very high.

To build a really useful recommender system it

is needed the system to be fast. In different news

sources on the internet it is possible to ﬁnd tens of

new headlines within an hour. So in this dynamic

environment of news, the news recommender sys-

tem should have a fast and real time processing

capabilities (Li et al., 2011). Independent from

which approach is used, scalability is one the most

important problems of recommender systems.

• Unstructured Content: For the systems which

require content information, it is hard to analyze

the content, especially for the news domain. For

better news recommendations, news items should

be structured and machine readable (Saranya.K.G

and Sadhasivam, 2012).

• User Modeling/proﬁling (Knowledge of User

Preferences): User proﬁling is an important com-

ponent of recommender systems. To make more

individual speciﬁc recommendations it is needed

to construct a user proﬁle. As it is stated in (Liu

et al., 2010), (Das et al., 2007), (Saranya.K.G and

Sadhasivam, 2012) there are many different ap-

proaches for user proﬁling.

• Gray Sheep Problem: Since collaborative ﬁl-

tering recommends items according to the user’s

common interests with other users, it is not pos-

sible to recommend proper items to people whose

preferences do not consistently agree or disagree

with any group of people (Su and Khoshgoftaar,

2009). When the total number of users increases,

the possibility of this problem occuring decreases

(Borges and Lorena, 2010).

• Serendipity (Over-specialization, Portfolio)

Problem: This is the problem when the system

recommends similar or the same items with

the already recommended ones. For the news

domain, a news item written differently in dif-

ferent news sources may be recommended by

the recommender system as different articles. It

is obvious that the users would not be happy to

get the same or similar recommendations. The

system should always be able to discover new

items to recommend by avoiding the same items.

In (Iaquinta et al., 2008) the problem is discussed

in detail for content-based systems but it is also

a problem for collaborative ﬁltering systems

(Borges and Lorena, 2010).

• Privacy Problems: To make proper recommen-

dations to a user, the system should know about

the users’ past preferences, interests and even the

relations with other people. This requires the stor-

age of detailed data about the user and the analysis

of this data that can cause privacy issues (Garcin

et al., 2013).

• Neighbor Transitivity: Neighbor transitivity oc-

curs when the database is very sparse. Even if

WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies

280

there are two users who have similar interests, the

system cannot detect them because of the lack of

ratings they have on similar items (Su and Khosh-

goftaar, 2009).

• Synonymy: Same items can be named differently

by separate resources and it is not possible for ma-

chines to understand that they refer the same item.

For example; even if the “children’s movie” and

“children’s ﬁlm” have the same meaning, they can

be treated as different items by the recommender

system (Su and Khoshgoftaar, 2009).

4 APPROACHES OF NEWS

RECOMMENDER SYSTEMS

In this section we explain the different approaches to

the most addressed challenges of news recommender

systems. Some of these challenges are not completely

speciﬁc to the news domain. But in most of the pre-

vious works about news recommender systems these

challenges are prioritized and mostly addressed. The

summary of which approach addresses to solve which

particular challenge can be seen in Table 1. The term

N/A is used for deﬁning that particular challenge is

not addressed or it is unknown if the challenge is ad-

dressed or not in that work.

4.1 Cold-Start Problem

Since collaborative ﬁltering ﬁnds similarities between

different users’ and makes recommendations by using

the different preferences of similar users, it is impos-

sible to recommend a new item which is not evalu-

ated yet. Another aspect of this problem is that it is

impossible to recommend any items for completely

new users. Cold-start problem is the most common

problem for the applications that use collaborative ﬁl-

tering. As it is mentioned in (Liu et al., 2010) for

some researchers it is one the most important disad-

vantages of collaborative ﬁltering approach. To solve

this problem (Liu et al., 2010), (Fortuna et al., 2010)

and (Lin et al., 2012) are proposed a hybrid method

using the collaborative ﬁltering and content-based ﬁl-

tering together. In (Liu et al., 2010), it is proposed to

use the personalization for recommending new arti-

cles which is building a proﬁle for the user’s genuine

interests. In (Lin et al., 2012), to handle the cold-start

problem, the system includes the opinions of chosen

experts (who uses the social networks have signiﬁcant

inﬂuence on new users). So that the system can make

recommendations to a new user. By using the TF-IDF

method new items can also be recommended. In (For-

tuna et al., 2010), it is proposed another approach that

is grouping the users as an old or a new user according

to the number of articles they read. For each group of

users, a separate model is trained for predicting the

most interesting news category. And the top new arti-

cle from each ﬁltered category is recommended to the

user. (Tavakolifard et al., 2013) proposes an archi-

tecture which considers the users’ long term prefer-

ences, short term preferences and the current context.

So that cold start problem can be eliminated by us-

ing the current contextual information for ﬁrst recom-

mendations. In (Lee and Park, 2007) it is checked that

if the user is new or not. If it is a new user than she is

temporarily placed in a similar segment on the basis

of demographics and the ﬁrst recommendations done

according to the preferences of that demographic seg-

ment. (Garcin et al., 2013) proposes a system which

works for anonymous users. When a user starts to

read a news item the system generates recommenda-

tions and during the session of the user the system up-

dates the model and makes better recommendations.

4.2 Recency

We see that nearly in all the works done on news rec-

ommendation the importance of recency is addressed.

In (Yeung and Yang, 2010) recency mentioned as an

important property of a recommender system and it is

proposed a proactive news recommender system for

mobile devices. Since the environment for a mobile

user is highly dynamic it is a challenge to deliver the

most proper and recent information to the user. In

the proposed system a Hybrid P2P system is used to

deliver the just-in-time information to users’ mobile

device. A pure P2P system is not suitable for mobile

devices since it requires lots of communication with

other devices. In the proposed architecture, mobile

device connects one of the servers in the network and

sends the context information. Recommendation is

done by the server (which gets the recent news articles

constantly from RSS) by using this context informa-

tion of the user which includes user proﬁle, location,

usage patterns, peer ratings etc. As a different ap-

proach, (Wen et al., 2012) includes the time factor in

recommendation process. In this approach, to recom-

mend a news item, time factor is taken into account as

a coefﬁcient in addition to the user interest and pref-

erence models. Similarly, in (Lee and Park, 2007)

weights of articles is calculated by the degree of im-

portance and recency of that article. In (Fortuna et al.,

2010), news categories are determined for each user

according to the user’s preferences and the newest ar-

ticle of each category is selected for recommendation.

On the other hand, in (Li et al., 2011) it is constructed

a news proﬁle for news items which includes dynamic

ASurveyonChallengesandMethodsinNewsRecommendation

281

Table 1: Different works on news recommendation with the challenges they addresses to solve.

Cold-start Recency Implicit Changing User Scalability Data

Feedback Interest Sparsity

(Yeung and Yang, 2010) N/A X X X N/A X

(Wen et al., 2012) N/A X X X N/A N/A

(Liu et al., 2010) X X X X N/A N/A

(Resnick et al., 1994) N/A X N/A N/A X N/A

(Lee and Park, 2007) X X X X N/A N/A

(Li et al., 2011) X X X N/A X N/A

(Das et al., 2007) N/A X X X X N/A

(Fortuna et al., 2010) X X X N/A N/A N/A

(Tavakolifard et al., 2013) X X X X N/A N/A

(Saranya.K.G and Sadhasi-

vam, 2012)

N/A X X X X X

(Garcin et al., 2013) X X X X X N/A

(Lin et al., 2012) X N/A N/A N/A N/A X

characteristics like recency and popularity. (Liu et al.,

2010) considers the news trends to make proper rec-

ommendations. Since news trends mostly composed

of recent news it can also be taken as a challenge of

recency.

4.3 Implicit Feedback

To predict the future interests of a user and to make

proper recommendations, the system needs to know

the past interests of the user. There are two ways of

learning about the past interests of a user: Explicit and

implicit feedbacks. To collect explicit feedbacks it is

needed to interact with the user continuously and ask

if the user liked the item or not, how much she liked

it and maybe other questions about the system in gen-

eral. Both for the users and for the system it is not

practical to continuously interact with the user. Espe-

cially in mobile devices it is hard to manually collect

personal information (especially textual information)

from the user. (Yeung and Yang, 2010) So it is desired

to make user proﬁling and ﬁltering automatically. The

system should be able to collect implicit feedbacks ef-

fectively while protecting the user privacy. In (Garcin

et al., 2013), it is stated that to overcome the lack of

data about the users most systems require the users

to have logged in the system which can cause privacy

issues. Implicit feedbacks are mostly taken from the

log analysis of users’ history. In (Liu et al., 2010), to

predict the future user interests, a large-scale log anal-

ysis is done over the registered users’ history data and

the change of user interests are observed. Similarly in

(Wen et al., 2012) user’s interest and preference mod-

els are constructed by using the user’s navigational

data. Also in (Tavakolifard et al., 2013) the user’s

behaviors are used for detecting some preferences of

the user. In (Lee and Park, 2007) regular analysis is

done over the history of the user and the system learns

about various preferences.

4.4 Changing User Interest

It is known that as time passes the interests of people

change. The preferences of people about movies, mu-

sic or books generally show a slight difference within

short periods of time. But in the news domain it is

again very different from other domains. The news

reading preferences of people can be affected by on

going circumstances in the world as well as their age,

cultural level and even their mood. So, predicting

future interests of users can be a real challenge for

news recommendation. (Liu et al., 2010) addresses

this challenge and proposes an architecture to predict

the future user interests. To do this a hybrid news

recommender system is proposed where a large-scale

log analysis is done over the registered users’ past ac-

tivities. Click distribution over different news cate-

gories is computed both for individuals and groups in

a monthly basis and the change of user interests are

observed. For prediction of user interests, Bayesian

framework is used. Then the predictions are used in

information ﬁltering method. To make proper rec-

ommendations it is combined with the existing col-

laborative ﬁltering method. In (Lee and Park, 2007),

the change of user interests are tracked by observing

and comparing the long-term and short-term prefer-

ences. The comparison of coefﬁcients of long-term

and short-term preferences changes the weight of cat-

egory preferences. So if there is a change in the cat-

egory preference of a user over time, it can be used

for making proper recommendations. (Wen et al.,

2012) creates a model for the user’s degree of inter-

est to a speciﬁc topic by analyzing the navigational

data (frequently visiting a web page related with a

speciﬁc topic shows the user is interested in that sub-

ject) and then it is updated as the user keeps browsing

WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies

282

web pages. To keep track of the changing user in-

terest (Saranya.K.G and Sadhasivam, 2012) proposes

two kinds of user proﬁles: Static and dynamic user

proﬁles. Static user proﬁle includes the user’s sign up

information like user name and favorite topic where

the dynamic user proﬁle is constructed by using the

implicit user data in every session.

4.5 Scalability

Since scalability problem applies to every computer

related system it is also one of the most important

challenges in recommender systems. If we want to

build a useful recommender system it is obvious that

it must be scalable. In news domain, scalability prob-

lem combines with other challenges which makes the

news recommenders more challenging to build. In (Li

et al., 2011) it is aimed a scalable news recommenda-

tion system by clustering the news articles and elim-

inating the unnecessary similarity computations. So

that the system spends less time for computation and

can response faster. In (Das et al., 2007) to be able to

serve millions of users they proposed a new MinHash

(a probabilistic clustering method) based user clus-

tering algorithm, redesigned the PLSI (Probabilistic

Latent Semantic Indexing, a model for performing

collaborative ﬁltering) as a MapReduce (a model for

computing large scale data on clusters) computation

and used item covisitation technique (a method for

determining the user-item relations) for a more scal-

able system. (Saranya.K.G and Sadhasivam, 2012)

discusses the need of scalability in efﬁcient recom-

mender systems. And the Hadoop framework handles

the issues like reliability and scalability in their ap-

plications. Another different approach for news rec-

ommendation includes context trees. In (Garcin et al.,

2013), it is discussed the scalability problem does not

occurs within this approach since it requires only one

tree and the tree structure is very limited because of

the applied context constraints.

4.6 Data Sparsity

Even though it is one of the most important chal-

lenges of collaborative ﬁltering method, data sparsity

is not addressed in most of the news recommender

system approaches. Data sparsity occurs when the

number of users or items are much higher than the

other one. In this case when the user-item matrix

is constructed, the matrix would be very sparse. In

(Yeung and Yang, 2010), it is discussed that using

Bayesian Network makes it easy to eliminate the data

sparsity. On the other hand in (Saranya.K.G and Sad-

hasivam, 2012) HBASE (a non-relational distributed

database) is used to store the data where it can pro-

vide fault tolerant storage for sparse data. Another

solution for data sparsity is to use the hybrid approach

(Lin et al., 2012).

5 DISCUSSION

News recommendation is a speciﬁc domain in rec-

ommender systems which has special challenges and

characteristics. Even though some of the challenges

are shared with recommender systems in general, oth-

ers may require different approaches to solve. As it

is seen in Table 1, different works addresses to solve

particular challenges. We see that all the approaches

try to come up with solutions to as much challenges

as possible. Some of them highlights one or two chal-

lenges and addresses the others as secondary chal-

lenges.

Since they are addressed and solved nearly in all

of the works, recency and implicit feedback chal-

lenges seem the easiest ones to solve. For recency,

most approaches prefer to recommend the latest head-

lines. Calculating the recency by using a coefﬁcient

which decreases in time is also another commonly

used solution (Wen et al., 2012), (Lee and Park,

2007). Implicit feedback is also one of the most ad-

dressed challenges. It highly depends on the data ex-

traction from users’ navigational history or log anal-

ysis. The analysis and storage of this huge amount

of data about users can cause privacy issues when it

is required the users to log in to the system (Garcin

et al., 2013). Also it reduces the scalability which is

another important challenge. As we can see in Table

1, nearly all approaches have solutions for these two

challenges. On the other hand, we see that scalability

and data sparsity are the challenges which less solu-

tions are offered. For some approaches like (Tavako-

lifard et al., 2013) data sparsity challenge is not avail-

able since it is a problem only for collaborative ﬁlter-

ing method. It is also possible to eliminate the prob-

lems which belong to only one of methods by using

hybrid systems.

As we can see in Table 2, some of the methods

have minor differences but even though there are sim-

ilarities between methods some approaches like con-

text tree approach (Garcin et al., 2013) really differ

from others. Some approaches include only one of

the ﬁltering methods where others include both of

them (hybrid methods). In approaches that use hybrid

method, it is possible to see that they use or propose

different algorithms. (Yeung and Yang, 2010) pro-

poses a new method for news ranking which is called

AHP (Analytic Hierarchy Process) model. AHP of-

ASurveyonChallengesandMethodsinNewsRecommendation

283

Table 2: Methods used in different news recommender systems.

Type Algorithm Log

Analysis

(Yeung and Yang, 2010) Hybrid Bayesian Network, AHP X

(Wen et al., 2012) Hybrid TF-IDF, Naive Bayes Model X

(Liu et al., 2010) Hybrid Bayesian framework X

(Resnick et al., 1994) Collaborative

ﬁltering

Matrix Correlation -

(Lee and Park, 2007) Collaborative

ﬁltering

Speciﬁc equations used for calculations X

(Li et al., 2011) Hybrid LSH (Locality Sensitive Hashing) MinHash, NLP,

LDA (Latent Dirichlet Allocation)

(Das et al., 2007) Collaborative

ﬁltering

For model-based - Item covisitation, for memory-

based - PLSI and MinHash

(Fortuna et al., 2010) Hybrid For model-based - SVM (Support Vector Machine) X

(Tavakolifard et al., 2013) Content-

based ﬁltering

TF-IDF, NLP, NER (Named Entity Recognition) X

(Saranya.K.G and Sadhasi-

vam, 2012)

Hybrid Adaptive user proﬁling, dynamic neighborhood cal-

culation, document ranking calculation

(Garcin et al., 2013) Hybrid Context tree, BVMM, LDA (Latent Dirichlet Allo-

cation)

(Lin et al., 2012) Hybrid TF-IDF, probabilistic matrix factorization models -

fers a solution to assign weights for different ranking

factors. (Fortuna et al., 2010) proposes an SVM (Sup-

port Vector Machine) (a machine learning model)

based news recommender system. (Saranya.K.G and

Sadhasivam, 2012) proposes calculation methods for

adaptive user proﬁling, dynamic neighborhood calcu-

lation and document ranking calculation.

Using context trees for news recommendation is

proposed in (Garcin et al., 2013) where it is used to-

gether with LDA (Latent Dirichlet Allocation) and

BVMM (Bayesian Variable-order Markov Model) al-

gorithms and deﬁned as a scalable and effective so-

lution to most of the challenges. LDA is also used

in (Li et al., 2011) for the representation of topic

distributions in a text. Bayesian Network is the

most widely used technique to model user interests

(Yeung and Yang, 2010), (Wen et al., 2012), (Liu

et al., 2010). Another technique used for user proﬁl-

ing is NLP (Natural Language Processing) (Li et al.,

2011), (Tavakolifard et al., 2013). There are different

tools used for NLP technique like GATE and Apache

OpenNLP library. For news items clustering LSH

(Locality Sensitive Hashing) and MinHash (a prob-

abilistic clustering method, a scheme for LSH (Das

et al., 2007)) techniques are used in (Li et al., 2011).

MinHash is used together with PLSI (Probabilistic

Latent Semantic Indexing) in (Das et al., 2007). In

content based ﬁltering TF-IDF (Term FrequencyIn-

verse Document Frequency) is one of the mostly

used techniques. In addition to TF-IDF, (Tavakolifard

et al., 2013) uses NER (Named Entity Recognition)

model to identify the names and location.

In Table 2 we can see that nearly in all of the works

it is done log (or click) analysis over the usage data.

Most systems require to users log in to the system.

They gather data both from the sign up process and

from the actions of the user while she is using the

system. The approach in (Garcin et al., 2013) does

not require any log in to the system, thus it makes

recommendations only for the active session without

determining who the user is. The other systems learn

about users and they make recommendations based

on the long term preferences of the user. By using log

analysis it is also possible to track the change of user

interests.

As the number of researches grow in recom-

mender systems and speciﬁcally in news recommen-

dation domain, we see that the number of hybrid sys-

tems increases. Recent evaluations show that hybrid

systems tend to outperform other systems.

Evaluation and quantitative comparison of differ-

ent recommender systems are other challenging as-

pects of recommender system research. Even though

there are several evaluation methods for measuring

the performance of the system, it is hard to measure

the qualitative aspects like user satisfaction. There are

some reasons which makes the quantitative compari-

son of the references we addressed in this paper not

possible for us. First, they use different evaluation

metrics which are not comparable. Second, because

of the challenges they solve are different, the systems

are not identical to each other to make quantitative

comparisons.

WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies

284

6 CONCLUSIONS

Beginning from the ﬁrst half of 90’s, recommender

system research continues to grow in different appli-

cation domains. Nowadays it is not hard to see a rec-

ommender system working in the background of the

web site you have visited and recommends you music

or shopping items. Even though these useful appli-

cations of recommender systems exist, there are still

many challenges for a true personalized recommender

system. As the number of online news sources in-

creases it becomes harder for an end user to ﬁnd what

she is looking for. Thus the need for a news recom-

mender system increases.

In this paper, the challenges and different meth-

ods of news recommendation domain are presented. It

is pointed out which approaches solve different chal-

lenges and how they do this. Our current frame-

work for comparing recommender systems in news

domain deals with content-based, collaborative and

hybrid recommendation approaches, though we in-

tend to expand it with recent results from semantically

based recommender systems. Including semantic rep-

resentations in the recommendation process helps us

to understand user needs and news content and can

be valuable when several of the challenges above are

addressed.

REFERENCES

Adomavicius, G. and Tuzhilin, A. (2005). Toward the

next generation of recommender systems: A survey of

the state-of-the-art and possible extensions. Knowl-

edge and Data Engineering, IEEE Transactions on,

17(6):734–749.

Borges, H. L. and Lorena, A. C. (2010). A survey on rec-

ommender systems for news data. In Smart Infor-

mation and Knowledge Management, pages 129–151.

Springer.

Burke, R. (2002). Hybrid recommender systems: Survey

and experiments. User modeling and user-adapted in-

teraction, 12(4):331–370.

Cantador, I., Bellog

ın, A., and Castells, P. (2008).

News@hand: A semantic web approach to recom-

mending news. In Proceedings of the 5th interna-

tional conference on Adaptive Hypermedia and Adap-

tive Web-Based Systems, AH ’08, pages 279–283,

Berlin, Heidelberg. Springer-Verlag.

Das, A., Datar, M., Garg, A., and Rajaram, S. (2007).

Google news personalization: Scalable online collab-

orative ﬁltering. In WWW ’07 Proceedings of the 16th

international conference on World Wide Web, pages

271–280.

Fortuna, B., Fortuna, C., and Mladenic, D. (2010). Real-

time news recommender system. In ECML PKDD’10

Proceedings of the 2010 European conference on Ma-

chine learning and knowledge discovery in databases:

Part III, pages 583–586.

Garcin, F., Dimitrakakis, C., and Faltings, B. (2013). Per-

sonalized news recommendation with context trees.

CoRR, abs/1303.0665.

Iaquinta, L., Gemmis, M. D., Lops, P., Semeraro, G., Filan-

nino, M., and Molino, P. (2008). Introducing serendip-

ity in a content-based recommender system. In Hy-

brid Intelligent Systems, 2008. HIS’08. Eighth Inter-

national Conference on, pages 168–173. IEEE.

Lee, H. and Park, S. J. (2007). Moners: A news recom-

mender for the mobile web. Expert Systems with Ap-

plications, 32(1):143–150.

Li, L., Wang, D., Li, T., Knox, D., and Padmanabhan, B.

(2011). Scene : A scalable two-stage personalized

news recommendation system. In SIGIR ’11 Pro-

ceedings of the 34th international ACM SIGIR con-

ference on Research and development in Information

Retrieval, pages 125–134.

Lin, C., Xie, R., Li, L., Huang, Z., and Li, T. (2012).

Premise: personalized news recommendation via im-

plicit social experts. In Proceedings of the 21st ACM

international conference on Information and knowl-

edge management, pages 1607–1611. ACM.

Liu, J., Dolan, P., and Pedersen, E. R. (2010). Personalized

news recommendation based on click behavior. In IUI

’10 Proceedings of the 15th international conference

on Intelligent user interfaces, pages 31–40.

Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P., and

Riedl, J. (1994). Grouplens: an open architecture for

collaborative ﬁltering of netnews. In Proceedings of

the 1994 ACM conference on Computer supported co-

operative work, pages 175–186. ACM.

Saranya.K.G and Sadhasivam, G. (2012). A personalized

online news recommendation system. International

Journal of Computer Applications, 57(18):6–14.

Sarwar, B., Karypis, G., Konstan, J., and Riedl, J. (2001).

Item-based collaborative ﬁltering recommendation al-

gorithms. In Proceedings of the 10th international

conference on World Wide Web, pages 285–295.

ACM.

Su, X. and Khoshgoftaar, T. M. (2009). A survey of col-

laborative ﬁltering techniques. Advances in artiﬁcial

intelligence, 2009:4.

Tavakolifard, M., Gulla, J. A., Almeroth, K. C., Ingvaldesn,

J. E., Nygreen, G., and Berg, E. (2013). Tailored news

in the palm of your hand: a multi-perspective transpar-

ent approach to news recommendation. In Proceed-

ings of the 22nd international conference on World

Wide Web companion, pages 305–308. International

World Wide Web Conferences Steering Committee.

Wen, H., Fang, L., and Guan, L. (2012). A hybrid approach

for personalized recommendation of news on the web.

Expert Systems with Applications, 39(5):5806–5814.

Yeung, K. F. and Yang, Y. (2010). A proactive personal-

ized mobile news recommendation system. In De-

velopments in E-systems Engineering, pages 207–212.

IEEE.

ASurveyonChallengesandMethodsinNewsRecommendation

285