News Recommendation in Real-Time - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 162

frequently induces a power-law distribution of interactions. A power-law distribution
manifests as few items comprise a relatively large fraction of interactions. Conversely,
a large fraction of items comprises only relatively few interactions. Popularity has
been found to affect establishing users' trust into the recommender system [ 48 ].
Recommender systems which suggested popular items had a better chance to engage
users to interact with them. Figure 6.3 shows the popularity distribution of a news
portal's articles along with the Movielens movie rating data set. We observe that
both exhibit similar shapes. Few individual items comprise a majority of interaction.
Conversely, the majority of items comprises only few interactions.

6.4.3 Item Collection Dynamics

Continuously adding new items to existing collections represents a major reason for
the information overload. Additions incur as film studios create new movies, music
labels release new albums, or editors publish new books. Some of the novel items
may become popular ones attracting plenty interactions. Others may remain barely
known. The frequencywithwhich items enter collections depends on the type of item.
According to [ 22 ], European publishers released about 535,000 books in 2013. In
contrast, news articles represent a muchmore high-frequency type of item. Individual
news portals account for hundreds of thousands articles published per year.
News consumption differs from other domains. On the one hand, movies, songs,
and topics attract users throughout longer periods. For instance, we consider the
rating data from the Movielens data set. Each interaction conveys a timestamp. Thus,
we compute the duration in between the last and first interaction for each movie.

Oct Jan

entry exit

2000

1500

1000

500

Oct Jan

Fig. 6.4 Number of items entering and exiting the news portals over time. On the left-hand side ,
the figure shows how many items we observed whom users interact with for the first time. On the
right-hand side , the figure illustrates how many items we observe no future interactions with