News Recommendation in Real-Time - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 151

to avoid explicit querying. This feature becomes particularly helpful in situations
where users lack a defined information need. Instead, users require systems to pro-
vide information that will likely be of interest to them.
Researchers have proposed a variety of ideas to carry out the selection process.
The ideas range from rather simplistic approaches to highly sophisticated meth-
ods carrying a plethora of parameters with them. Trivial methods include randomly
recommending items as well as suggesting items based on their popularity. Two par-
adigms cover a large fraction of the more advanced methods: collaborative filtering
and content-based filtering. The former builds on the idea of leveraging other users'
preferences to provide recommendations. The latter strives to discover items whose
contents share similarities with items users have liked in the past. A comprehensive
discussion of both exceeds our scope. Still, we present a selection of ideas tailored
for the news domain. We refer readers interested in recommender systems in general
to [ 1 , 43 , 52 ].
Proposed news recommendation approaches either utilize other users' interactions
with news portals, (possibly enriched) news contents, or both. Thus, we recover both
paradigms of regular recommender systems.
Liu et al. [ 40 ] introduce a Bayesian framework to allow hybrid recommendations
of news articles to users in a personalized fashion. They showed that considering
content features increased news consumption compared to a collaborative filtering
baseline. Li et al. [ 38 ] model news recommendation as a contextual-bandit problem.
They show that replaying recorded interactions enables researchers to consistently
evaluate their recommendation methods. They provide the theoretical foundations
for the unbiasedness of such a methodology. De Francisci et al. [ 21 ] make use of
three kind of inputs to their news recommendation system. First, they consider inter-
actions in terms of clicks. Second, they extract contents from micro-blogs. Finally,
they consider the social relation between the micro-blogging service's users. They
represent the problem as learning to rank task. The proposed method considers all
three factors to adjust the ranking of news articles for target users. Son et al. [ 57 ]pro-
pose to consider users' current locations to improve the news item selection process.
Additionally, the authors utilize semantic data to enrich the representations of users'
interests and locations' relevant concepts. Capelle et al. [ 14 ] investigate whether
semantic similarities between named entities in news articles can be leveraged to
improve recommendation quality. The method requires name entity recognition as
a preprocessing step. Bogers and van den Bosch [ 9 ] propose a probabilistic frame-
work to provide better news suggestions. Their work looks at the problem from
an information retrieval perspective. They analyze the impact of the selected rele-
vance model on the recommendation quality. Li et al. [ 37 ] propose a personalized
news recommendation framework. Their work emphasizes the issues arising due
to the dynamics inherent in item collections. Consequently, they propose to rep-
resent the recommendation task as a contextual bandit problem. Li and Li [ 35 ]
propose to leverage co-occurring interactions to improve news recommendations.
Their method models relations between concepts in news texts as hypergraphs. The
approach considers both user behaviors and contents. Garcin et al. [ 25 ] investigate
whether context trees enable news recommender systems to provide relevant news