Anda di halaman 1dari 10

1.

Article 1
The goal of this article is to describe and highlight the differences between
traditional and modern approaches and algorithms used for developing recommender
systems. It addresses both traditional and modern approaches, exemplifying them in
movies recommendations.

1.1 Problem
One of the big problems in the, so called, Information Era is the overload of
information. This issue was only partially resolved by the creation of search engines, as it
lacked the personalization of information in an infinite search space. Therefore, the long
term solution is the Recommender System, a tool for searching and filtering data of
interest for users.

1.2 Methods
There are three traditional approaches that are used in the creation of
recommender systems: content- based filtering, collaborative filtering and the hybrid
filtering.
On the other hand, the article presents also some modern approaches used for
modeling recommender systems: the context- aware, the cross- domain based and the
cross- lingual.

A. Content- based Filtering


The content- based filtering approach recommends items that are similar to items
that a particular user has liked in the past. These systems build up firstly the user’s profile
based on the user’s taste and afterwards recommend those items close to the user by
similarity of profiles. In order to calculate the similarity between items, the cosine
similarity function is used.

B. Collaborative Filtering
The collaborative filtering approach makes predictions about a user’s interests by
compiling his preferences with those of a user in the same neighborhood. Users rate items
using a ranking option. This approach applies best to e- commerce solutions.

C. Hybrid Filtering
The hybrid approach combines both the content- based and the collaborative
filtering approaches and is thought to be more performant.
The hybrid approach is used to develop the Cinema Screen Application, where the
system plays the agent role for making suggestions about the best movie running at the
cinema. Each user creates an account and ranks the movies watched at the cinema.
Afterwards, the collaborative filtering technique is applied and then, on the output, the
content- based filtering is used [2].

D. Context- aware Filtering


The context- aware approach takes into consideration the circumstances under
which rankings and ratings are given by the user when making recommendations. For
example, an entertainment recommendation would recommend rather going to the
swimming pool then to the beach when outside is raining.
The high quality of the recommendations is a strength point of this approach. A
drawback of this approach is obtaining the context data: via surveys, via location,
extracting it from reviews using text mining algorithms.
The COMPASS mobile application offers tourist recommendations of activities
that might be of interest in the visited area. This application was implemented using
context- aware filtering technique [3].

E. Cross- domain Filtering


This approach addresses the process of building an accurate neighborhood in the
collaborative filtering context. The problem is that, although two users’ profiles share
similar rankings in one domain, this doesn’t always apply to other domains. Therefore,
the cross- domain approach considers domain dependent neighborhoods.

F. Cross- lingual Filtering


When using cross- lingual filtering the user can receive recommendations even if
their descriptions do not match the user’s spoken languages. Once the language barriers
are broken, the recommendation space increases considerably. This is accomplished by
building a language- independent representation of a text using semantic analysis.

1.3 Challenges and Issues


The article describes the main challenges and problems that might arise in the
recommendation process: cold- start, trust, scalability, privacy and sparsity (lack of
recommendations).

1.4 Remarks
 Advantages
A strength point of this article is that most of the theoretical approaches presented
are also exemplified, making them easier to understand.
Another important section is the part regarding challenges and issues, as it arises
concerns about performance, security and scalability.
The organized structure of the article is an advantage because each section
presents the theoretical part, the used approach and a relevant example.
 Drawbacks and missing elements
A negative aspect is the lack of specification or mathematical proofs for
algorithms. For example, mathematical solution is missing in the process of obtaining
different neighborhoods related to the collaborative filtering approach.

1.5 Conclusions
In conclusion, the article describes very well the theory around recommender
systems using masterpiece examples.
2. Article 2
The article presents the developed Cinema Screen Recommender Agent, a system
designed to make movie recommendations using the hybrid filtering technique. The users
can select a cinema in their area and receive recommendations of movies that are
currently running on the screens.

2.1 Problem
The most common shortcoming of a movie recommender system using the
collaborative filtering technique is the impossibility to recommend new movies or movies
that have not been rated by any user before.
The solution to this problem is to use instead the hybrid filtering technique, a
combination between collaborative and content- based filtering techniques. The
experimental results prove an improved accuracy of recommendations based on content.

2.2 Method
The Cinema Screen Recommender Agent uses the hybrid filtering technique. The
key point of this approach is that it applies content based filtering on the output set
obtained using collaborative filtering.

A. The Collaborative Filtering


The first step is to determine for the current user a set of users that share similar
movie taste in terms of ratings. These are called potential peers. Further, a list of films
that are rated by at least one such peer is generated. Then, the weighted average for each
film in the list is computed. There are two mathematical methods used at this step: the
Pearson’s product- moment correlation coefficient and the weighted mean, which is the
predicted rating for a film.
The output of this step is the list of movies, the corresponding predicted ratings
and the number of peers that rated the movie.

B. The Content- based Filtering applied on the output of the Collaborative Filtering
A scoring mechanism is used for each film in the list. Initially, the weighted
average is summed up with the score of each actor, director and category. Further, an
average score is calculated for each actor, director and category. Then, separately, for
each actor, director and category sum up the average score with the current score for each
film they played in, directed or belong to.
In the end, the recommender agent displays movies recommendations in
ascending order by their score.
2.3 Performance
In terms of performance, the process of computing recommendations uses a lot of
server resources. In order to avoid frequent recalculations, each set of recommendations
is cached until the end of the session.
In the testing phase, a data set provided by the University of Minnesota’s
GroupLens Research Group and containing 943 users and 10 000 movie ratings has been
considered. The tests are meant to provide recommendations for a random set of 200
users.

2.4 Comparisons
The proposed approach was compared with other four approaches: collaborative
filtering used in isolation, content- based filtering used in isolation, collaboration- via-
content filtering and running the results of content based filtering through collaborative
filtering.
The evaluation metrics used are precision and recall. Precision defines the
probability that the movies selected by the agent will be relevant to the user. Recall
defines the probability that the whole set of relevant films will be selected by the
recommender agent.
Moreover, coverage metrics are used to compute the number of movies the
system is able to make recommendations for. In the testing phase, three types of coverage
are applied: standard coverage, catalog coverage and prediction coverage.
All in all, at the end of the testing phase it is demonstrated that the proposed
approach is able to generate more recommendations than the collaborative filtering
technique and that the quality of the recommendations increases compared to the content-
based filtering technique.

2.5 Remarks
 Strong points
A strength point is that, due to the content- based filtering technique applied, the
system is able to recommend also new movies, not just the already rated ones.
3. Article 3
This article presents the mobile application COMPASS, which was designed to display
any touristic information about a city at any time.

3.1 Problem
In our society access to information is a crucial need. A good example is the travellers
who need guidance about what tourist attractions the city offers. With the rise of mobile
phones, personal digital assistants and position systems (GPS), this need is easy fulfilled.

3.2 Method
Using a server side component, the interaction manager, the application finds the best
way to communicate the user’s request. For example, if the user clicks on a POI that
represents a restaurant, the application and show him all sorts of information like menu,
timetable, distance. Another component of the application is the POI retriever that receives a
request from the interaction manager when the user context changes. After that, the POI
retriever searches for the best match of POIs in the new user context and the application
displays the new information on the user’s map. The application also uses a recommendation
engine that applies multiple prediction strategies to predict how interesting each POI is for
the user. Used prediction methods include social filtering, item-item filtering, case-based
reasoning and category learning.

3.3 WASP Platform


The COMPASS application is built upon the WASP platform that supports services such
as context manages and service registry. The WASP platform components are request
dispatcher, notification manager, context manager, service registry and matchmaker. The
request dispatcher is responsible for forwarding user requests to the 3G-network. The
notification manager allows applications to send and receive notification when the context
changes. The context manager retrieves information about the user’s context. The service
registry contains information about the third-party applications. The matchmaker uses the
service registry to discover services that match with the requests received from an
application.

3.4 Performance
After a user testing 57 people consisting of 23 females and 34 males, within the age of 10
to 70 years old, the result show that most of the participants think that predictions are useful
and that the COMPASS application would resolve their need when they are travelling.
3.5 Remarks
 Strength Points
The application can give access to information really fast and can adapt to the user’s
needs. It can increase awareness of the surrounding and can save a lot of time.
It offers a detailed overview of the application’s architecture.

 Drawbacks
In order for the application to work, a powerful mobile phone is needed. Also, a
permanent and stable connection to the 3G-network is mandatory. Those factors can lead to a
quick battery drain.
There are no insights on how exactly the recommendation process is done: how the
recommendations are generated or which mathematical formulas are used for the
recommendation algorithm.
4. Article 4
This article proposes a news recommendation system for Google News based on click-
behavior.

4.1 Problem
The information overload problem makes it harder and harder for the user to select
relevant content on the Internet.

4.2 Method
The system builds user profiles based on the click- behavior gathered from the web
history. Then, using the click logs, a Bayesian framework is built up to predict the user’s
current news interests.
In order to generate recommendations, the collaborative filtering technique is combined
with an information filtering technique based on learned user profiles. The main idea of the
methodology implies to predict user’s news interests based on the clicks from the past time
period. Then, the predictions are combined to obtain an accurate prediction of the user’s
genuine news domain of interest and the final step involves the combination of news interests
with the current news trend from the user’s location.
In the log analysis phase, the click distribution is computed. Then, in order to be able to
update the user’s profile over time, a comparison between the click distributions of the
current month and the ones from all the previous months is done. In this scope, a distance
function between click distributions is defined.
In a similar manner, in order to establish the news trend, the click distribution of the
public per country is computed. In order to generate the list of recommended articles, the
proposed system computes and information filtering score (IF) and a collaborative filtering
score (CF) for each article.

4.3 Data
The anonymous click logs of randomly selected 16.848 users from more than ten different
countries and regions were used as input data.

4.4 Performance and Comparisons


For testing purposes, the application was deployed on Google News. The results show
that the presented approach increased the web traffic and the quality of news
recommendation has increased.
In comparisons with other similar approaches existing in the literature, the information
filtering approach presented in this article is original due to its capability to dynamically
determine the user’s current news interest by always comparing the genuine interest of the
user with the current news trend.
4.5 Challenges and Issues
One constraint of the news information filtering algorithm is that the user’s interest in
pieces of news changes very often, therefore, a constant update of the user’s profile should be
done.

4.6 Remarks

The process of combining information filtering and the collaborative


filtering information improves the performance of the collaborative method.

The experiment described the situation of 10.000 users for the live traffic of
Google News. Users were assigned to a control group and to a test group. The
users from the control group received news from the classical collaborative
method and the users from the test group received news from the enhanced
method. The experiment runs for 34 days from 1/10/2009 to 2/17/2009. The
metrics used to measure the performance of the recommender systems were:
click-through rates of the recommended news section and frequency of visiting
Google News website.

The combined method improved the quality of news recommendations and


increased the number of visitors for the Google News website.

References

[1] Asanov D., Algorithms and methods in recommender systems, Institute of Technology
Berlin, Germany, 2011.

[2] J. Salter and N. Antonopoulos, “Cinema Screen recommender agent: Combining


collaborative and content-based filtering,” IEEE Intelligent Systems, vol. 21, pp. 35–41, January
2006.
[3] M. van Setten, S. Pokraev, and J. Koolwaaij, “Context-aware recommendations in the
mobile tourist application compass,” in Adaptive Hypermedia and Adaptive Web-Based
Systems, ser. Lecture Notes in Computer Science, P. De Bra and W. Nejdl, Eds. Springer Berlin
/ Heidelberg, 2004, vol. 3137, pp. 515–548.

[4] Liu, J., Dolan, P., Pedersen, E. R.: Personalized News Recommendation Based on Click
Behavior. IUI’10: Proceedings of the 2010 International Conference on Intelligent
User Interfaces (2010). 31–40.

[5] Das, A. S., Datar, M., Garg, A., Rajaram, S. Google news personalization: scalable online
collaborative filtering, Proceedings of the 16th international conference on World Wide Web,
2007.

Anda mungkin juga menyukai