News Recommendation in Real-Time - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 153

These factors include diversity [ 34 ], novelty [ 60 ], stability [ 3 ], and scalability
[ 5 , 54 , 58 ]. Having decidedwhich criteria to optimize, we face another design choice:
Do we rely on recorded data or do we aim to interactively conduct experiments with
users [ 27 , 55 ]? Both alternatives have advantages. Offline experiments entail little
costs. Additionally, other researchers can reproduce results as the data used for eval-
uation is fixed. Conversely, conducting experiments with actual users may better
reflect the actual use-case. User studies as well as deploying novel algorithms into
existing recommender systems represent two alternatives for online experimentation.
Related work covers a wide spectrum of news recommendation's aspects. Most
recent works focus on two of these aspects. First, researchers seek to improve recom-
mendation quality by using additional data sources. These sources provide textual
descriptions, interaction with users, and social relations. We still cannot satisfy-
ingly tell how to determine additional data's value in advance. Second, research
investigates potentials to algorithmically improve recommendations. Due to inher-
ent requirements, we struggle to transport established, sophisticated methods to the
news domain. Besides these two major aspects, researchers seek to discover better
evaluation protocols along with means to deal with the real-time character of news
recommendation.
6.3 The Plista Case
We introduced recommending news articles as a challenge for science and industry
in Sect. 6.1 . Subsequently, we outlined methods enabling news portals to suggest
news articles in Sect. 6.2 . Both occurred on a rather abstract level. In this section,
we present an actual news recommendation scenario. The scenario focuses on the
plista GmbH. Plista runs a content and advertisement recommendation service on
thousands of premium websites. These websites include portals dedicated to news
and entertainment among other topics. Having a large customer base, plista processes
millions of user visits on a daily basis. Each visit has to be handled in real-time as web
portals attempt to instantly deliver their contents. Portals include recommendations
by means of a widget.
The quality of their recommendations represents a major asset to plista. Users
accepting recommendations do not only provide revenues. Evidence for increased
visitor satisfaction facilitates acquiring new portals to serve with recommenda-
tions. Consequently, plista continuously seeks to improve their recommendation
algorithms. Similarly, Netflix seeked to improve their movie recommendations thus
releasing a large rating data set in 2006. The Netflix Prize competition has shown that
combinations of recommendation algorithms provide better recommendations [ 6 ].
Combinations of algorithms have shown to better reflect contextual factors [ 2 ].
Hence, plista seeks to acquire new algorithms thus improving their system's rec-
ommendation quality.