News Recommendation in Real-Time - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 156

6.3.1.5 Recommendation Providers
Recommendation providers capitalize on their algorithms. Typically, portal operators
pay them by click. Hence, recommendation providers seek to maximize the probabil-
ity of visitors clicking on their recommendations. Hereby, they face a dilemma which
we refer to as “exploration exploitation trade-off”. Recommendation providers prefer
to use the methods most likely to maximize click rates. However, even if individ-
ual methods have performed successfully in some scenarios, it stays unclear, which
method suits the current context best. Consequently, they have to evaluate different
methods which in turn may perform worse.
6.3.2 Technical Requirements
Plista have an eco-system at their disposal tailored precisely to news recommen-
dation. In contrast, researchers using ORP may have rather limited resources. A
selection of technical challenges impedes applying highly sophisticated recom-
mendation methods. Real-time response times represent such a challenge. ORP
sets the maximum response time to 100 ms. This affects both computational com-
plexity and model updates. News portal operators require ORP to provide rec-
ommendations within a predefined time slot. Exceeding this time slot, they can-
not include the recommendation into the displayed web page. Simultaneously,
real-time responses require recommendation models to be available at all times.
On the other hand, recommendation models ought to include recent news since
visitors are likely interested in what currently happens. Thus, operators have to
find ways to update their models while concurrently continue to provide recom-
mendations. Thereby, update frequency constitutes a significant parameter. Plista's
observations indicate that decreasing update frequencies negatively affects the click-
through rate. Evaluating recommendation algorithms on recorded data (cf. the
“Netflix Prize” challenge [ 7 ]) cannot cover this time-related aspects. Plista simul-
taneously runs a variety of recommendation algorithms to account for differ-
ent factors determining recommendation quality. The system continues updating
algorithms as news items appear, new interactions occur, and articles get updated.
The frequency with which the system updates algorithms depends on the method.
We report findings which plista observed for certain types of algorithms. Recom-
menders based on content perform well even when updated in low-frequency. In
contrast, collaborative filtering methods require high update frequencies as users'
interests shift. Additionally, collaborative filtering struggles to recommend items
which have not obtained interactions. Further, recommendation algorithms sug-
gesting popular news articles performed best when updated with high frequen-
cies. ORP's users will also have to deal with the technical requirements listed
above.