News Recommendation in Real-Time - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 164

6.4.4 Contextual Factors
News consumption is subject to a variety of contextual factors. Our news consumption
differs with respect to the time of day, day of week, location, device, mood, and
more. Determining the current context represents a difficult problem. In particular,
confounding contextual factors impede recognizing situations correctly. Contexts
manifest as combinations of contextual factors. For instance, users reading news
on a weekday at noon on their desktop in good mood represent a specific context.
Altering an individual factor may provide a context requiring a different kind of
suggestions. For instance, users reading news on a weekday at noon in a good mood
but on their tablet devices may dislike reading comprehensive articles due to their
limited screen sizes. Figure 6.5 shows the relative frequencies of interactions grouped
by daytime, weekday, and device. The majority of interactions recorded for desktop
computers concentrates on the working times. Contrarily, phones as well as tablets
account for larger proportions of interactions during evenings as well as weekends.
Generally, we observe neglectable proportions of interactions during the night times
for all device types. Suppose we ought to select a recommendation algorithm for
a particular request. Context represents an important aspect we need to consider.
Requests are more likely to arise from mobile devices on the weekend. Mobile
devices provide less space to display recommendations on. Thus, we should consult
the recommendation method which performs best under these circumstances.
We have seen that sparsity, popularity, dynamics, and context represent major
impeding factors for news recommendation. Sparsity hampers establishing valuable
user and itemprofiles. Sparsity represents a particular challenge for newly added users
and items. This is due to the system having almost no knowledge about preference
relation with the entity. The system struggles to determine what items a new user
will like. Conversely, it cannot reliably select potential consumers. Popularity skews
consumption distributions as few items concentrate large amounts of interactions.
Contrarily, unpopular items see hardly any interactions. Dynamics refer to the system
characteristic of fluctuating item collections. In established domains songs, movies,
and books remain recommendable items. Conversely, news' relevance fades with
time. Finally, systems have to consider users' current context to select enjoyable
articles. Users may dislike reading comprehensive articles on mobile devices. In
Sect. 6.5 , we discuss a selection of algorithms and their abilities to deal with these
specificities.
6.5 Recommendation Algorithms for News
Recommendation algorithms are subject to a vigorous research community.
Researchers continuously propose and evaluate novel methods or extend existing
ones. Methods differ with respect to complexity, applicability, and the underlying
ideas. In the following, we introduce and discuss four kinds of such underlying ideas
and their implementations: