Implicit User Profiling in News Recommender Systems
Jon Atle Gulla
1
, Arne Dag Fidjestøl
1
, Xiaomeng Su
2
and Humberto Castejon
2
1
Dep. of Computer and Information Science, Norwegian University of Science and Technology, Trondheim, Norway
2
Telenor Group, Trondheim, Norway
Keywords: Recommender Systems, Personalization, User Profiling, Mobile News, Big Data, Information Retrieval.
Abstract: User profiling is an important part of content-based and hybrid recommender systems. These profiles model
users’ interests and preferences and are used to assess an item’s relevance to a particular user. In the news
domain it is difficult to extract explicit signals from the users about their interests, and user profiling
depends on in-depth analyses of users’ reading habits. This is a challenging task, as news articles have short
life spans, are unstructured, and make use of unclear and rapidly changing terminologies. This paper
discusses an approach for constructing detailed user profiles on the basis of detailed observations of users’
interaction with a mobile news app. The profiles address both news categories and news entities, distinguish
between long-term interests and running context, and are currently used in a real iOS mobile news
recommender system that recommends news from 89 Norwegian newspapers.
1 INTRODUCTION
With the growing popularity of smartphones and
tablets an abundance of news apps have been
introduced over the last few years as alternatives to
online web news sites. News aggregator apps like
News360, Flipboard, Pulse and Feedly are not linked
to one particular media house, but make use of
publically available news stories as soon as they are
published on the Internet. These mobile applications
allow users to read neatly presented news stories
from numerous media houses using interfaces that
require only very limited interaction with the
system. The features of these apps vary, though
information filtering and user friendliness are key
features of many of these applications (Haugen,
2013).
Advanced recommendation technologies are still
rare in commercial news apps, but some of them
now include simplified recommendation features,
and several research prototypes experiment with
new and promising recommendation approaches that
analyze both the content of news articles and the
users’ social networks. This is not very surprising,
as mobile news may benefit from these
recommendation technologies for a number of
reasons: (i) the constant flow of news easily leads to
information overload problems for the reader, (ii)
the small screen prevents the apps from showing
several stories or proper news overviews, and (iii)
the lack of an efficient keyboard hampers the
reader’s interaction with the news apps.
However, an efficient recommender system
needs some understanding of the users’ preferences
and interests. Some news apps do not keep any
particular information about the user and there is no
user-tailoring of news or recommendations offered.
Other apps allow the user to select a category or post
a query. The category or query associated with the
user is then the simplest form of user profile found
in mobile news apps. For full-fledged
recommender systems, though, a more
comprehensive representation of user interests and
preferences is needed to provide personalized news
services. If news stories read in the past is a good
indication of the user’s current preferences, we need
to analyze these stories and find ways of capturing
their content in her user profile. If stories read by
friends or similar people are assumed to be relevant,
we need ways of relating the user’s profile to other
users’ profiles and find exactly those other users that
have read news stories that are most likely to be of
interest to her. There are many aspects of user
profiles, and the choice of recommendation
technique also influences the structure, content and
maintenance of these profiles.
The NTNU SmartMedia project is experimenting
with advanced recommendation techniques in a
185
Gulla J., Fidjestøl A., Su X. and Castejon H..
Implicit User Profiling in News Recommender Systems.
DOI: 10.5220/0004860801850192
In Proceedings of the 10th International Conference on Web Information Systems and Technologies (WEBIST-2014), pages 185-192
ISBN: 978-989-758-023-9
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
news app for Norwegian newspapers. Users are in
this app associated with devices, and the mobile
news recommender engine ranks news using a
combination of freshness, geographical proximity
and match with users’ preferences. Central in the
work is the definition of user profiles that reflect and
interpret user behavior and support both content-
based and collaborative filtering.
The purpose of this study is to investigate to
what extent a minimalistic swipe-based mobile news
interface can be used to infer user preferences
without any explicit signals from the users. The
paper itself is structured as follows: After the
introduction in Section 1, we discuss the particular
challenges and needs in news recommender systems
in Section 2. In Section 3 we briefly present earlier
work on user profiling in recommender systems.
Section 4 introduces the SmartMedia news
recommendation project and its overall method for
extracting user profiles from end user’s interaction
with the mobile app. After discussing the way this
profiling technique can be configured to serve
different purposes in Section 5, we conclude the
paper in Section 6.
2 NEWS RECOMMENDATION
A news recommender system filters incoming news
and presents to each individual user a ranked list of
news that it deems most relevant to the particular
user at hand. Since these are articles that the users
have not already read and evaluated, we need to find
a way of guessing users’ interests based on their
previously observed behavior.
In formal terms, the task in news
recommendation is to estimate and rank the
evaluations of news articles unknown to a user
(Borges & Lorena, 2010; Jannach et. al, 2010).
Assume a set of users U and a set of news articles A,
the recommender system needs a utility function s
that defines the evaluation v of an article a for a user
u:
:
→,
in which V is a completely ordered set formed by
non-negative values within an internal, e.g. 0 to 1 or
0 to 100. The system will recommend an article a’
that maximizes the utility function for the user:
argmax
∈
,
For those articles that have not been evaluated by the
user, the system needs to estimate their evaluations
from relatedness with other articles that have in fact
been read. Techniques like decision trees, Bayesian
classifiers, support vector machines, singular value
decomposition, clustering and various similarity
scores have been used as part of this estimation
process.
In content-based filtering, which is the primary
concern of this paper, the estimation is all based on
the content of previously read news articles. The
assumption is that users read articles on topics they
find interesting, and the users’ interests do not
change substantially from one day to another. If a
particular user preferred to read about politics one
day, chances are good that she will also be interested
in political news the next day. Moreover, the degree
of interest in a topic may be reflected in the
frequency with which the user visits news stories of
this topic or category.
2.1 News Challenges
Recommender systems for products and services
normally deal with items that are rather stable both
in terms of number and descriptions. The
recommendation task is to compare a fixed number
of items that all have structured and well understood
properties. There is no relevant temporal dimension,
and there is no problem separating one product or
service from another.
The news domain is intrinsically more dynamic
and unpredictable to deal with. There are new
stories coming up all the time. Some of these
uncover new events, while others just report on the
progress of an already published event. There may
be several conflicting stories of the same event, but
there may also be several events discussed in the
same story.
For news recommendation there are three
particular challenges that need to be addressed:
First, the news domain is characterized by
fluctuating and unclear vocabularies and ever
changing news topics. Rather than ranking a fixed
number of products, the system need to rank a
dynamically growing number of events that may
have very little in common with what the user read
yesterday. The vocabularies change as new stories
emerge, making it difficult to detect how stories are
related and thereby estimate their relevance to the
user.
Second, the short life span of news stories
renders most news stories irrelevant, even though
they seem consistent with the user’s interests. For a
developing story, it makes sense to recommend the
latest article, unless there are substantial differences
in quality and depth that make an older article more
WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies
186
informative. On the other hand, if collaborative
filtering is used, it may be difficult to pick up the
latest article before anyone else in your community
has read and rated it.
Third, most media houses consider serendipity to
be important to hold on to its readers. Serendipity
refers to the introduction of news that may not have
been selected or preferred by the user, but may still
be interesting because they are surprising, alarming,
important, have a particular journalistic quality, or
simply add variation into the news stream.
Serendipitous news help media houses expand users
interests in their products, since they persuade users
to read stories that are alien to them and that would
never be read if the user only received stories
consistent with their reading history.
While other recommender systems concentrate
on the utility function itself, news recommender
systems need also to consider content analysis
techniques to extract and structure the topics
pertaining to the news. They need to deal with
ambiguities, uncertainties and inconsistencies that
affect the way articles should be formally
represented and compared to users’ interest profiles.
3 RELATED WORK
A user profile is a representation of user’s interests
and preferences that is used to verify to what extent
news stories are relevant to a particular user. The
profiles are built for each individual user, are
regularly updated, and describe topics, news
categories and relevant features of the users.
In principle, there are two types of user profiles,
profiles based on implicit feedback and profiles
based on explicit feedback (Gauch et al, 2007; Lops
et. al, 2011). Implicit user profiles are automatically
extracted by the recommender systems themselves
and may or may not be a correct representation of
users’ interests. In general, implicit feedback
methods assign relevance scores to user actions on
news articles like saving, sharing, bookmarking, etc.
Explicit profiles are entered or approved by the users
in question, but tend to be slightly less detailed than
the implicit ones. The user-selected categories in
Zite and Flipboard form very simple explicit user
profiles, while the thumbs up/down approach in Zite
and News360 provide explicit signals for what
topics should be included in the profiles. However,
adopting numeric or symbolic scales increases the
users’ cognitive load and may not be adequate for
capturing emotions or attitudes towards the news.
For mobile news recommendation it seems difficult
to require that the user enter and regularly update
extensive representations of her interests, though,
and more advanced techniques for profile
construction on the basis of implicit feedback are
needed.
A system combining explicit feedback and
automatic learning is described in Singh et. al
(2006). After building an initial interest category
hierarchy on the basis of explicit feedback on a
number of articles, the system analyzes user
feedback from ongoing news sessions and
automatically adds new leaf categories or update
existing ones in the interest hierarchy. A similar
approach is taken by Kim & Chan (2003), though
they depend less on explicit user feedback.
Billsus and Pazzani (2000) have developed an
approach in DailyLearner for interpreting implicit
user feedback on news articles. A user click on the
headline of an article is taken as a signal of interest,
and an initial score of 0.8 is set. If the user is
requesting more pages of the story, the score will be
gradually increased until it reaches a maximum of
1.0 when all pages have been consulted. Similarly, a
skipped article is assumed to be of no interest and is
given a negative score that is subtracted from the
system’s prediction score for the article. All these
scores are combined into a user profile that lists
weighted informative words typical for the user’s
interests and preferences.
Liu, Dolan and Pedersen (2010) build user
profile vectors that express users’ interests in
specific news categories over time. For each user
they record the distribution of clicks and associate
these click rates with categories on a monthly basis.
This allows them to assess for every user the
proportion of time spent on reading news from each
category as well as to reflect on the development of
her interests from one month to another.
In Cantador et. al (2008) they use semantic
expansion in combination with a standard user
preference algorithm for content-based
recommendation. They observe that automatically
learned profiles tend to be dominated by the main
characteristics of user’s preferences, preventing the
recommendation engine from recommending news
that are related albeit not directly addressed by the
profile. Their solution is to use an ontology to
include additional weighted concepts in the profile
that are related to the original learned concepts or
terms. Then they combine this expanded profile
with a running context, which is a weighted set of
concepts from the user’s latest interactions with the
system – to produce a contextualized version of the
user’s preferences that filters out topics that are out
ImplicitUserProfilinginNewsRecommenderSystems
187
of focus.
Implicit signals of user interests require that the
user interface is designed in such a way that
different levels of interests in a news article are
reflected in the user’s interaction with the system.
This is a challenge in itself, as most news apps try to
minimize the required interaction between system
and reader. More interaction is good for analysing
user interests, though it may lead to a less attractive
user experience.
4 THE SMARTMEDIA PROJECT
The SmartMedia project at NTNU was initiated in
2011 as part of the university’s collaboration with
the Norwegian media industry and major Telecom
companies. The project targets recommendation
technologies for the mobile news domain and
addresses content-based and collaborative
techniques in combination with semantic and
linguistic theories (in Gulla et. al (2014).
Another important issue is the use of platforms,
frameworks and methods for designing efficient
cross-platform mobile user interfaces. Figure 1
shows the current mobile news recommender
systems for the iOS platform. Details of this
implementation can be found in Tavakolifard et. al
(2013). A new version of the system in
HTML5/JavaScript is under development and will
introduce new features for hybrid news
recommendation.
Figure 1: SmartMedia’s news recommender app.
The current iOS news app indexes about 1,500 news
articles per day. The articles are sourced from 89
Norwegian newspapers and are mostly in one of
Norway’s two official languages. During the
indexing process, we extract and index the article’s
category (e.g. sports or lifestyle), its location (e.g.
Oslo), and relevant keywords characterizing the
article’s content. The keywords are partly named
entities (e.g. Lionel Messi) from a NER (named
entity recognition) component, partly important
nominals that refer to concepts discussed in the
article. The underlying news index is in SolR, an
open source search platform from the Apache
Lucene project, and is built up as a traditional large-
scale search engine index (see for example Gulla et.
al (2002) and Solskinnsbakk & Gulla (2010)).
5 USER PROFILING
In SmartMedia’s user profiling component a set of
10 user acts have been identified as implicit signals
of user interest. These are Opened article view,
Article view time, Preview time, Clicked category,
Shared twitter, Shared facebook, Shared mail,
Starred article, Viewd map, and Viewed similar
article. We monitor the time spent reading a news
story or looking at its preview, assuming that large
view dwell times indicate a stronger interest in the
topic.
The construction of a user profile for a particular
user follows a three-step process:
1. Extract interests from user acts
2. Build running context from all user acts
3. Combine running context and long-term interests
into new user profile
In the following, we go into detail of each of
these three steps.
5.1 User Acts
Whenever the user is interacting with the news app,
her actions are logged and stored with all relevant
data associated with the act. This includes not only
the user ID and the act itself, but also the time of the
act, the location of the news reported in the article,
the news category, as well as entities and important
key phrases describing the content of the article.
Figure 2 shows how a particular user act is
represented in the system. As seen from the
timestamp, this is the act of reading a particular
news story at 16:41 on 2nd June 2013. The user
spent 1.4 seconds reading it, and the article itself
was a news article about an Indian man that resisted
an arrest in Kristiansand in Norway. The content is
WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies
188
revealed from the keywords (named entities and
important nominal phrases) that are listed under the
‘tags’ label.
As different types of user acts may signal
different levels of user interest, each type of act t is
associated with a constant
t
. For timed user acts,
like reading a preview or the full version of a news
story, we only want to consider this reading if it is a
significant act, i.e, she spends more time than usual
on the story.
Figure 2: User act logging.
At the content level, we now assume that the
interests of the user are reflected by the important
keywords found in the article she has read, modified
by the importance of the user act. We model this as
a user act interest vector
= (e
1
,..., e
k
) for act i,
where e
j
is defined as
0
∗
t
denotes the importance of user acts of type t,
while
is the frequency of keyword j. In the current
implementation
t
is set to 1 for all types t, though
experiments show that some of these acts are more
significant than others.
A similar analysis is used for categories, giving
us also a user act category vector.
5.2 Running Context
A running context for a user is built from all user
acts of that user that have happened after the last
time a full user profile was generated. It describes
the overall topics of what the user has been clicking
on or reading lately without reflecting what she
might have been interested in at earlier occasions.
The running context gives us the user’s current news
focus.
Formally, the running content context
is
defined as follows:
∈
where i T are the user acts in the time span T after
the last time her profile was generated.
A similar running category context is also
defined.
Implementationally, the mobile news app records
every gesture from the user and maintains an
updated running context at all times. Whenever the
user decides to reset his context or consult her stored
profile, the running context is compared with her old
user profile on the server side. When the user ends
her session, the running context is incorporated into
her long-term user profile, so that she will start with
an updated user profile and an empty running
context next time.
5.3 User Profile
Our user profiles combine long-term interests with
users’ focus on current news stream. The idea is that
the user may be interested in following news events
that are unfolding, even though they are not
necessary perfectly in line with her general interests
and preferences.
For each user the system stores an old user
profile
. The new user profile
is calculated
from the old profile and the running context as
follows:
1
If c is set to 1, the user profile does not take into
account the running context and will never change.
A value of 0 makes the old profile irrelevant, and
only the running context is used to recommend news
stories to the user. I practice the constant c should be
carefully selected to both address the users’ need for
relevant news and the content providers’ need to
promote the latest news.
As before there is a similar calculation for
constructing the user profile at category level.
It is important to note that we also assume that
long-term interests fade unless they are renewed
when similar topics come up in the news stream.
Since c will always be below 1, the features of old
ImplicitUserProfilinginNewsRecommenderSystems
189
Figure 3: Generated user profile.
user profiles gradually disappear as they are updated
with running contexts. The lower the c value is, the
faster the system will forget the user’s old profile.
Figure 3 shows a generated user profile for a user
that has just started using the system. Both the
category part and the content part of the profile are
shown. The category part tells us that this user is
mostly interested in news, though she has also
enjoyed some stories about traveling and sports. At
the content level, we notice that the user has
preferred local news, as Norway, Trondheim and
Sør-Trøndelag are all location names relevant to his
home place. The rest of the profile exhibits some
sports personalities as well as culture persons in the
Norwegian media. At the top of the profile, there
are two variables, mean preview time and mean read
time, that are used to determine if a timed user act is
significant or not. If the user spends less than 4.2
seconds reading a full news story, the system will
not consider this reading act relevant to the
calculation of her user profile.
6 DISCUSSION
The SmartMedia news recommender app sources
news from all major newspapers in Norway. Some
of these news come with meta data like category and
tags, though most news stories are just unstructured
of texts that need to be analyzed to locate the news,
categorize the stories and extract informative
keywords that describe their content. Since these
keywords are critical for both the construction of
profiles and subsequent recommendation process,
their quality and availability are important to the
recommendation process.
With the exception of finance articles, news
stories in Norway are on average 220 words (tokens)
long. Finance news tend to be more than 1,000
words long and written in a more factual and
objective manner.
On the average our news articles contain the
following named entities:
Named entity Occurrences per story
Locations 1.6
Persons 2.3
Organizations 2.3
Roles 0.8
Since there are so few named entities in news
stories, our user profiles also include important
nominal phrases that constitute more concepts than
references to particular entities.
The evaluation of user profiles has so far focused
on the contribution of each type of user acts on the
quality of the profile (Nilsen, 2013). In one
experiment we used sensitivity analyses to assess the
influence of each type of user acts, and to what
extent they are consistent with each other. The
results show that the Opened article view act in
many cases provided a good average of the other
acts taken together, though the other acts were
needed to deal with particularly interesting or
uninteresting news stories. Another experiment
analyzed the impact on user profile size. Not
surprisingly, the Preview time act accounted for 82%
of the profile vectors’ contents after some time of
regular use. This may be an indication that our
threshold for taking this act into account is set to low
or has to reflect a more dynamic rule than just
accepting any reading act longer than the average
one. Quite likely the lengths of the news stories
WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies
190
need to be a factor when considering the reading
time.
Experiments with the news recommender system
shows that there are three issues that need to be
carefully considered when user profiles are
constructed:
Granularity. In the current implementation we
consider category interests and content interests
equally important. This is a choice that seems
reasonable to most users, though there are many
users that would prefer a stronger focus on either
the category as a whole (“i like all sports”) or on
its particular topics (“I don’t care about football,
but like Messi and Barcelona”).
Persistence. The balance between stable long-term
user interests and short-term news context is
delicate. If long-term interests are preferred, the
user risks that there is no relevant news available
and she will leave the app. Emphasizing the
running context too much is also unfortunate, as
this will make the app appear similar from one
user to another.
Serendipity. There is also a balance between
profile-relevant news stories and serendipitous
news. Most content providers like to include also
news that are not directly within a user’s profile to
trigger new interests and widen her perspective
(Das et. al, 2007). This is in our case achieved
using content-based filtering for profile-driven
news and collaborative filtering for serendipitous
news, and the weighting of the techniques decide
the balance between these two kinds of news.
In sum, it is difficult to identify an ideal user
profile in news recommender systems. Users have
different priorities than content providers, and the
various weighting schemes that may be used to
address the granularity, persistence and serendipity
of news typically depend on whether they reflect the
users’ or the content providers’ perspective.
It is also worth noting that freshness and location
seem very important to users, even if the stories
themselves are not necessarily the best match with
their user profiles. In this respect news
recommendation differs fundamentally from product
recommendation on online shopping sites, where
users have a clear and stable interest that calls for all
relevant products independently of age or origin.
Our profile construction approach is in many
ways similar to what is used in SCENE (Li et al.,
2011). However, whereas we combine content
phrases and entities in one vector, they split it up and
also consider more complex phrases as part of their
content description. A more fundamental difference,
though, is the granularity of the user click analysis
itself. As opposed to SCENE and most other news
apps that look at news stories as either read or not
read, we consider in more detail what the user is
doing with the stories and use these smaller acts as
cues for refining the users’ real interests and
preferences. This technique is to some extent also
exploited in DailyLearner (Billsus and Pazzani,
2000), though the SmartMedia app takes every user
gesture as an indication of user satisfaction or
dissatisfaction.
7 CONCLUSIONS
The Smartmedia mobile news recommender system
demonstrates both the value and serious challenges
of implicit user profiling. Our mobile news
recommender system uses extensive user profiles to
recommend news stories that deal with the same
topics as the profiles. However, the system is highly
configurable, with a number of parameters that
seriously affect the news stories recommended to the
users. As there is no obvious best configuration of
the system, since users and content providers have
different agendas, only time will tell what is an
acceptable weighting scheme for both readers and
media houses.
The news app has been available online since the
summer of 2013, and we are now in the process of
building a new mobile user interface based on
feedback from the early months. We are also –
together with the media industry – collecting a large
training set with real news and real users over
several months, which we intend to use to refine our
weighting schemes for balancing different
recommendation strategies.
Research-wise we are gradually expanding the
recommender system with semantic features for
modeling news events and the entities involved in
these events. This involves the definition of news
taxonomies not very different from what is used in
semantic search solutions (see e.g. Brasethvik &
Gulla, 2002). With a new log-in feature we also
intend to use social media sites like Twitter to
expand and deepen the understanding of users’
interests and preferences (e.g. O’Banion, 2012).
ACKNOWLEDGEMENTS
This research was supported by Telenor Group as
part of their collaboration with the Department of
ImplicitUserProfilinginNewsRecommenderSystems
191
Computer and Information Science at the Norwegian
University of Science and Technology in
Trondheim.
REFERENCES
Billsus, D. and Pazzani,M. J., 2000. User Modeling for
Adaptive News Access. User Modeling and User-
Adapted Interaction, 10, pp. 147-180.
Borges, H. L. and Lorena, A. C., 2010. A Survey of
Recommender Systems for News Data. In Szczerbicki
& Nguyen (eds.), Smart Information and Knowledge
Management, SCI 260, pp. 129-151. Springer.
Brasethvik, T. and Gulla, J. A., 2002. A conceptual
modeling approach to semantic document retrieval. In
Proceedings of the 14
th
international Conference on
Advanced Information Systems Engineering
(CAiSE’02), pp. 167-182. Springer.
Cantador, I. Bellogin, A. and Castells, P., 2008.
Ontology-Based Personalised and Context-Aware
Recommendations of News Items. In Proceedings of
the 7
th
International Conference on Web Intelligence,
pp. 562-565. IEEE.
Das, A. S. Datar, M. Garg, A. and Rajaram, S., 2007.
Google news personalization: scalable online
collaborative filtering. In Proceedings of the 16
th
international conference on World Wide Web, pp. 271-
280. ACM.
Gauch, S., Speretta, M., Chandramouli, A., and Micarelli,
A., 2007. User profiles for personalized information
access. The adaptive web, pp. 54-89. Springer.
Gulla, J. A. Auran, P. G. and Risvik, K. M., 2002.
Linguistic Techniques in Large-Scale Search Engines.
In Proceedings of the 6
th
International Conference on
Applications of Natural Language to Information
Systems (NLDB’02), pp. 218-222.
Gulla, J. A., Ingvaldsen, J. E., Fidjestøl, A. D., Nilsen, J.
E., Haugen, K. R., Su, X., 2014. Learning User
Profiles in Mobile News Recommendation. Accepted
for publication in Journal of Print and Media
Technology Research.
Haugen, K. R., 2013. Mobile News: Design, User
Experience and Recommendation. MSc thesis.
NTNU, Trondheim.
Jannach, D. Zanker, M. Felfernig, A. and Friedrich, G.,
2010. Recommender Systems: An Introduction.
Cambridge University Press.
Kim, H. R. and Chan, P. K., 2003. Learning implicit user
interest hierarchy for context in personalization. In
Proceedings of the 8th international conference on
Intelligent user interfaces, pp. 101–108. ACM.
Li, L., Wang, D., Li, T., Know, D., and Padmanabhan, B.,
2011. SCENE: a scalable two-stage personalized news
recommendation system. In Proceedings of SIGIR’11,
pp. 125-134. ACM.
Liu, J. Dolan, P. and Pedersen, E. R., 2010. Personalized
news recommendation based on click behavior. In
Proceedings of the 15th international conference on
intelligent user interfaces, pp. 31–40. ACM.
Lops, P., de Gemmis, M. and Semeraro, G., 2011.
Content-based Recommender Systems: State of the
Art and Trends. In Ricci, Rokach, Shapira and
Kantor (Eds.), Recommender Systems Handbook,
Chapter 3, pp. 73-106. Springer.
Nilsen, J. E., 2013. Large-Scale User Click Analysis in
News Recommendation. MSc thesis, NTNU,
Trondheim.
O’Banion, S. Birnbaum, L. and Hammond, K., 2012.
Social media-driven news personalization. In
Proceedings of the 4
th
ACM RecSys workshop on
Recommender systems and the social web. pp. 45-52.
ACM.
Rajaraman, A. and Ullman, J. D., 2011. Mining of Massive
Datasets. Cambridge University Press.
Singh, S., Shepherd, M., Duffy, J. and Watters, C., 2006.
An Adaptive User Profile for Filtering News Based on
a User Interest Hierarchy. In Proceedings of the
American Society for Information Science and
Technology, Volume 43, Issue 1, pp. 1-21, 2006.
Solskinnsbakk, G. and Gulla, J. A., 2010. Combining
ontological profiles with context in information
retrieval. Data & Knowledge Engineering, 69(3), pp.
251-260.
Tavakolifard, M. Gulla, J. A. Almeroth, K. C. Ingvaldsen,
J. E. Nygreen, G. and Berg, E., 2013. Tailored News
in the Palm of your HAND: A Multi-Perspective
Transparent Approach to News Recommendation. In
Proceedings of 22
nd
International World Wide Web
Conference (WWW’13), Companion Volutme, pp. 305-
308, May, Rio de Janeiro.
WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies
192
