AD ALTA
JOURNAL OF INTERDISCIPLINARY RESEARCH
in the system, he has not rated items yet. Thus, the system
will not be able to provide accurate recommendations.
The systems should contain rated items in order to
recommend some items to the users. When a new item
enters the systems, the item has not rated by users yet.
Therefore, the systems will not be able to recommend it to
the users.
Sparsity is a major problem for collaborative filtering
approach. The total number of ratings is important in the
recommendation systems. In order to provide accurate
recommendations by the recommendation systems,
sufficient number of ratings should exist in the systems.
For example, in movie recommendation systems, there are
many movies that have been rated by only a few people.
The systems will rarely recommend these movies [13].
In many practical collaborative filtering recommendation
systems, the number of users and items increase rapidly in
the system [8]. Therefore, the system needs to provide
more and complicated computational process, and this
leads the computational resources going beyond the
acceptable levels.
5 Demographic approach
A demographic approach recommends items to the user based on
the user’s demographic information such as gender, age, and
date of birth. It puts the users into groups based on their
demographic characteristics. The system will put the users who
belong to a certain zip code into one group. Also, the users of
ages ranging from 18 to 25 years-old will be in one group. The
recommendation systems based on demographic approaches
assume that the users in the same group or category share the
same interests and preferences [13]. The demographic approach
tracks the buying or rating behavior of the users within the same
group or category. If there is a new user entering the system, the
system first will place the user into a particular group based on
the user’s demographic information. Then, the system will
recommend products or items to the user based on the buying or
rating behavior of the other users in the group.
The purpose of the system is to recommend books to library
visitors based on their personal information that is gathered from
them through an interactive dialogue. Another recent example of
a recommendation systems based on demographic groups is
Lifestyle Finder. The system uses demographic groups from
marketing research to recommend a range of products and
services, and it gathers the data from users through a short
survey. The advantage of the demographic approach is: the
system does not require maintaining a history of user ratings like
in content based and collaborative filtering approaches [8].
5.1 Limitation of demographic approaches
The demographic approach suffers from is how to identify
the group or category that the user belongs to when the
user is new to the system.
The demographic approach how to identify the interests
and preferences of users within the same group.
The demographic approach is the demographic system
works well when the demographic data is available to the
system.
The accuracy of recommendation systems based on
demographic data is less than those recommendation
systems based on content or collaboration filtering.
6 Hybrid recommendation approach
Since all above-mentioned approaches have complementary
strengths and weaknesses, so a hybrid recommender system
combines two or more recommendation techniques to gain better
system performance and mitigate the weaknesses of individual
ones.
However, recommendation approaches previously introduced
may be combined in hybrid approaches. Many of the studied
approaches have some hybrid characteristics. For instance,
content-based approach uses global relevance attributes to rank
the candidates, or graph methods are used to extend or restrict
potential recommendation candidates.
Therefore, hybrid recommendation technique used so-called
“feature augmentation”. It is a weak form of hybrid
recommendation technique, since the primary technique is still
dominant. In true hybrids, the combined concepts are similarly
important, among the approaches reviewed; only TechLens
approaches may be considered true hybrid approaches.
TechLens [31] is one of the most influential research-paper
recommender systems. TechLens was developed by the
GroupLens31 team. Currently the GroupLens team is still active
in the development and research of recommender systems in
other fields. Between 2002 and 2010, Konstan, Riedel, McNee,
Torres, and several others published six articles related to
research-paper recommender systems. Often, McNee et al.’s
article from 2002 is considered to be the original TechLens
article [34]. However, the 2002 article introduced some
algorithms for recommending citations, which was introduced in
2004 by Torres et al. [31]. Two articles about TechLens
followed in 2005 and 2007 with respect to recommendations. In
2006, McNee et al. analyzed potential pitfalls of recommender
systems [35]. In 2010, Ekstrand et al. published another article
on the approaches of TechLens [36].
TechLens’ algorithms were adopted from Robin Burke [8] and
consisted of three content-based approach variations, two
collaborative filtering approach variations, and five hybrid
approaches.
Pure-content-based approach served as a baseline in the form of
standard content-based approach in which a term-based user
model was compared with the recommendation candidates. In
the case of TechLens, terms from a single input paper were used.
In content-based approach -Separated, for each paper being cited
by the input paper, similar papers are determined separately and
at the end the different recommendation lists are merged and
presented to the user. In combined content-based approach,
terms of the input paper and terms of all papers being cited by
the input paper are combined in the user model. Then the papers
most similar to this user model are recommended. [38, 39]
Pure-collaborative filtering approach served as another baseline
and represented the collaborative filtering approach from McNee
et al., in which papers were interpreted as users and citations
were interpreted as votes [34].
Hybrid: With Pure-CF->CBF Separated, recommendations were
first created with Pure- collaborative filtering. These
recommendations were then used as input documents for CBF-
Separated. Similarly, Pure-CF->CBF Combined, CBF
Separated->Pure-CF, and CBF-Combined->Pure-CF were used
to generate recommendations. Fusion created recommendations
with both CBF and CF independently and then merged the
recommendation lists.
The previously discussed filtering techniques can be combined
to produce a hybrid filtering [37]. Although any filtering
techniques are possible to be combined, we will only focus on
combining collaborative filtering approach and content-based
approach. There are several ways to combine two filtering
techniques [13], as shown in Figure 1:
Case A: implementing a collaborative filtering approach
and a content-based approach separately and combining
their recommendations afterwards;
- 261 -