Information Technology Reference
In-Depth Information
items to anonymous users. Their method builds context trees based on observed
user behaviors. The authors pay particular attention toward recommending novel
and diverse items. Cantador et al. [ 12 , 13 ] leverage two kinds of data to select more
relevant news items. On the one hand, they derive semantic concepts from an exist-
ing ontology. This represents a content-based approach. On the other hand, they use
contextual features to better account for recent trends. Das et al. [ 17 ] present insights
from a large-scale news recommendation system operated by Google. Their work
emphasizes the requirements which operating recommender systems face. They dis-
cuss how algorithms including MinHash and probabilistic latent semantic indexing
enable news recommender systems to apply the collaborative filtering paradigm in
large-scale settings. Montes-Garcia et al. [ 46 ] propose a news recommender system
tailored specifically towards the needs of journalists. Their approach pays particular
attention toward personal preferences as well as contextual factors. Gao et al. [ 24 ]
analyze how well micro-blogs support news recommendation by indicating trends
in an early stage. They investigate the trade-off between popular news and personal
tastes. Phelan et al. [ 47 ] present a socially-driven news recommendation service
which extracts data from micro-blogging services as well as RSS feeds. The authors
compare whether RSS contents, micro-blog contents, or a combination of both lets
news recommendation services select the most relevant news items. Kompan and
Bielikova [ 32 ] present a news recommender system based on content similarities.
The authors discuss the importance of low computational complexity induced by
short response times. Lv et al. [ 44 ] propose a method utilizing a variety of factors to
estimate articles' relatedness. These factors include relevance, novelty, connectivity,
and transition smoothness. For a detailed survey on personalized news recommen-
dation algorithms, we refer the reader to Li et al. [ 36 ].
Evaluating recommendation algorithms depends on a variety of factors. First, we
have to define the recommendation algorithm's objective. This entails specifying
the notion of a good recommendation. At first, this may appear trivial. Researchers
have come up with several different specifications. Recommender systems attained
increased attention with the “Netflix Prize” challenge [ 7 ]. This competitions seeked
to reduce the error rate when predicting users' preferences for movies. The organiz-
ers decided to use the root mean squared error to compensate for larger deviations.
Subsequently, researchers continued to optimize rating prediction scenarios [ 18 , 29 ,
33 , 50 , 53 , 58 ]. In addition, researchers started to define recommender systems as
ranking mechanisms. They argued that recommender systems ought to rank items
according the user preferences. Accurately estimated preferences yield such rank-
ings. Still, they do not constitute an essential input as long as algorithms keep the
pairwise order of preferences. Optimizing metrics including normalized discounted
cumulative gain (nDCG) and mean reciprocal rank (MRR) provide such rankings
[ 41 , 51 , 56 , 60 ]. Some researchers argue that users refute to consider all available
items. Instead, users limit their attention toward few most relevant items. We find
evaluation criteria accounting for these desires in the field of information retrieval.
Hereby, systems cut rankings at a pre-defined position. We measure recommendation
quality in terms of precision, recall, or a combination thereof [ 4 , 16 , 19 , 30 , 49 , 61 ].
In addition, evaluations may consider further factors determining systems' qualities.
Search WWH ::




Custom Search