Information Technology Reference
In-Depth Information
6.4 News Consumption Phenomena
This section introduces a variety of phenomena which we observed as users interact
with online news portals. These phenomena distinguish the case of recommending
news fromother subjects such asmovies, songs, or books.We dedicate a subsection to
the aspects
sparsity
(
6.4.1
),
popularity
(
6.4.2
),
dynamics
(
6.4.3
), and
context
(
6.4.4
).
Recommender systems have established in a variety of use-cases. They support
users' decision making. Typical use cases include deciding which movie to watch,
which song to listen to, and which product to buy. Recommender systems have
proofed to be valuable in those scenarios. In contrast, suggesting news entails a variety
of challenges. We discuss sparsity, popularity biases, dynamic item collections, and
contextual factors. These aspects represent the major challenges for operators of
news portals running recommender systems.
6.4.1 Sparsity
We observe users interactingwith items. Interactions cover a range of actions depend-
ing on the items. For instance, users may buy products, listen to music, watch movies,
or read news articles. We can quantify interactions by the cardinalities of the involved
sets of users and items. Let
u
∈
U
and
i
∈
I
denote users and items. Further, let
(
·
)
=|·|
card
denote the function returning the number of elements contained in
a set. Equation
6.1
defines sparsity. Sparsity reflects the fraction of interactions we
actually observed by the number of possible interactions. Note that
I
(u,
i
)
represents
the indicator function returning 1 if
u
interacted with
i
and 0 otherwise (see Eq.
6.2
).
u
∈
U
i
∈
I
I
(u,
i
)
sparsity
=
1
−
(6.1)
|
U
||
I
|
1
:
if we observe an interaction between
u
and
i
I
(u,
)
=
i
(6.2)
0
:
otherwise
Recommender systems operate on domains with high sparsity. Recommending
items with almost complete profiles represents a rather trivial problem. The lack of
such comprehensive information induces the need for intelligent suggestion mecha-
nisms. Table
6.1
displays sparsity levels of a selection of datasets. We observe that
most datasets include less than 3% of potential interactions. We determine potential
interactions by multiplying the numbers of users and items. Additionally, Table
6.1
shows the relation of observed interactions to potential interactions. For instance, the
Netflix
data set exhibits 1 in 86.4 potential interactions. In contrast, we recorded data
from two news portals where we observe 1 in 66622.8 potential interactions. This
illustrates the difficulty to select appropriate news articles as recommendations.
Search WWH ::
Custom Search