Concept 1: “Cold start”
The first concept
got the biggest interest from scientists in the following literature review – 9
articles out of 18 contained the discussions about “cold start”. Shah Khusro et
al. (2016) describes this challenge as a situation that appears when there
is entering for either new user or new item in the system. In first case,
system can’t identify user’s preferences and interests and respectively can’t generates
the right and accurate recommendations to the user. In second case, it is
difficult to predict characteristics and give accurate rate for the new item (Khusro, Ali , & Ullah , 2016).
Mohammad-Hossein Nadimi-Shahraki et al.
(2014) explained that there
are two types of dealing with this challenge: adaptive and non-adaptive. First
group of methods can adapt the new user’s preferences easily, that is why the
process of recommendation will be conducted more effective, than in
non-adaptive methods, that recommend the same items to all new users. Authors
explain several adaptive methods on the example of the movie recommendation.
One of them is item-item personalized method, where items are proposed to users
until the time the last ones will finally start to rate them (Nadimi-Shahraki & Bahado, 2014). Item-item
personalized method is not the best one in e-commerce, where there is no factor
of obligation, but it can be easy implemented for the e-learning, since user
has to deal with the following system because of the studying process
adds, that “cold start” problem in e-learning is much bigger than in other
spheres, because system should have all information about all what user knows
and all what he does not know yet and in the same time, what can be interesting
for the learner (Figure 4.1).
Figure 4.1. Good
recommendation system for e-learning (Sacksick , 2017)
Researcher proposes to use spaces and/or knowledge
graphs in order to understand the new user’s level of knowledge (Sacksick , 2017).
Concept 2: Scalability
Shuai Zhang et
al. (2017) mentioned the challenge of scalability as one of the most common
for the 21st century since an enormous amount of data is being added
to the bases every second (Zhang, Yao, & Sun, 2017). Shah Khusro et al. (2016) stated, that such new huge
information creates lots of difficulties for recommendation systems. Authors
give an example of Amazon.com, which has found a good solution for this
challenge: company grouped items in clusters that are recommended with the help
of collaborative filtering (Khusro, Ali , & Ullah , 2016). The same approach
can be done in e-learning: to group all courses in clusters by topics or level
Jianbo Yuan et
al. (2016) proposed to solve both challenges of cold start and scalability
deep learning matcher, that is actually based on the already mentioned by (Nadimi-Shahraki & Bahado, 2014) technique “item-item personalized method”. Jianbo Yuan et
al. (2016) applied it for Careerbuilder’s collaborative filtering-based
recommendation engine and called the model doc2vec
(Yuan, et al., 2016). As it was mentioned before, the same technique can be
implemented on e-learning.
Concept 3: Data Set Sharing Challenges
According to Katrien Verbert et al. (2012), this
challenge is more common for context-based filtering. Authors are sure, that in
e-learning, for a high-quality recommendation data about only item, learner and
teacher is not enough. There should be more information about time, place of
recommendation and so on. But this concerns some privacy issues (Verbert, et al., 2012). That is why this
concept will be discussed in details in the next part.
Olga C. Santos et al. (2011) states that
the reason of this challenge is that mostly educational recommendation systems
take information only from the following educational repositories. If advisory
systems were able to get data from general learning management system (LMS),
recommendation systems would be able to operate with much bigger data sets (Santos & Boticario, 2011).
Concept 4: Shilling Attacks
Shilling Attacks is more common for e-commerce but since education is also a
sphere, where it is possible to get some profit, this problem exist here also. Shah Khusro et
al. (2016) explain this challenge as the one that concerns fake ratings of
some item with the purpose to increase its recommendations for other users.
Mostly, such problem takes place in collaborative filtering and can result in
decreasing users’ trust to the website (Khusro, Ali , & Ullah , 2016).
Shyong (Tony) K. Lam et al. (2004) proposed several ways out of this problem. The most
simple one is to use item-item algorithms that reduce the risks of shilling
attacks (according to the scientists’ experiments). Another advice is to give
extra protection for new items since they are in the high-risk group. This
protection is about asking for rating only from professional critics, trusted users
or filterbots (Lam & Riedl, 2004). The same advices
can be implemented for e-learning. For example, if the educational web-site
proposes courses with money, some people can start shilling attacks in order to
make some course more popular in recommendations and give it fake high ratings.
In order to avoid this, it is better to use item-item algorithm – to recommend
to users courses based on their previous experience or to rate new courses only
by teachers or trusted users