International Conference on Information Technology Systems and Innovation (ICITSI) 2014

Bandung-Bali, 24-27 November 2014

ISBN: 978-1-4799-6526-7

Film Recommendation Systems using Matrix

Factorization and Collaborative Filtering

Mirza Ilhami Suharjito

Informatics Technology Department Magister in Information Technology
Faculty of Informatics Technology Binus Graduate Program
STMIK Mikroskil Bina Nusantara University
Medan, Indonesia Jakarta, Indonesia
mr.ilhami@gmail.com harjito@yahoo.com

Abstract — Collaborative filtering method was widely used in Based on [6], matrix factorization can be used to increase
the recommendation system. This method was able to provide the prediction accuracy in scalable dataset. Other than, matrix
recommendations to the user through the similarity values factorization also can be used to remove the dimension of the
between users. However, the central issues in this method were item space and retrieve latent relations between items of the
new user issue and sparsity. This paper discusses about how to
dataset.
use matrix factorization and nearest-neighbour in film
recommendation systems. Both of methods will be used in order For experiments purpose, Movielens Hetrec 2011 rating
to make more accurate recommendations. Based on the data will be used in this paper. The rating data will be used for
experiments results, the combination of matrix factorization and computation in recommendation and prediction algorithm.The
classical collaborative filtering (nearest neighbor) could improve combination of matrix factorization and classic collaborative
the prediction accuracy. It can be concluded that the combination filtering (nearest-neighbour) technique will be used to improve
of matrix factorization and nearest-neighbor produced a better the prediction accuracy.
prediction accuracy Preliminary section 3 discusses how the dataset will be
used. In section 4, we describe the mathematical approach of
Keywords— Recommendation Systems; Matrix Factorization;
Collaborative Filtering.
our recommender. Prediction accuracy was evaluated in
section 5. Finally, the conclusions in section 6.
I. INTRODUCTION (HEADING 1)
II. RELATED WORK
Recommendation systems has been rapidly growing due to
huge data informations available. We need a personalized In this section we briefly review the main works in the
systems that can be match to us, based on what we read context. The list of references is not exhaustive due to the
recently, our last activities, and how its relevance to us. page limit. Our main focus was in collaborative filtering
Recommendation systems become more popular today and has which has been successfully applied to several real world
been used in various fields. problems, such as Netflix’s movie recommendation.
Collaborative filtering has been the mostly used in A. Content-based Filtering
recommendation systems nowaday. This method was able to Content-based method is one of the oldest methods and
give recommendation to user based on their similarity to most popular in the recommendation. The principle of this
others. On the other hand, matrix factorization has also been method is recommend object that has similarities to some
familiar since Netflix Grand Prize [1]. other object, the user preferred in the past [2]. The similarity
Generally, recommendation system can be divide into two between objects was determined from the values of the
parts, the Content-based Filtering and Collaborative Filtering. characteristics of the object.
Content-based filtering based on the similarity of the items to Fig. 1 shows an example of a content-based filtering. Well-
the object that the user liked in the past [2]. Meanwhile, known relationship marked by full arrows, calculations or
Collaborative filtering is the process of evaluation using the similarity of objects marked with an arrow point and
information of user behavior or the behavior of other users predictive relationship marked by dotted arrows.
[3].Collaborative filtering grouped by memory-based and
model-based.
Some of the central issues in the collaborative filtering is
sparsity, content analysis, overspecialization and new user
issue (cold start problem) [4]. Hybrid method was used in
order to solve this problems. Hybrid is a combination between
several methods. This method was also a key that brought
Bellkor team as the winner in Netflix Prize in 2009 [5].

978-1-4799-6527-4/14/$31.00 ©2014 IEEE

1
 Fig. 1: Content-based filtering [4]

This method was widely used in text-based applications, to

recommend a document or web site, according to [4]. One
implementation of a content-based is the Music Gnome
Project and WebMate.
B. Collaborative Filtering
Collaborative filtering was one of the most widely used
techniques for recommendation systems. Companies like
Google [7], Amazon [8] using the recommendation, at least in
part based on collaborative filtering. CFwas a popular
recommendation algorithm that bases its predictions and Fig. 3:Hybrid Method [4]
recommendations on ratings or opinions of other users in the
system [3]. Typically, predictions about user interests were Hybrid method can help overcome some of the drawbacks
obtained by collecting taste information from many other of the method recommendations, eg. new item that has not
similar users. The fundamental assumption of this method was been rated by any user. Several methods can be used
that the tastes of other users can be selected and assembled in additional recommendations [9].
such a way to give a reasonable prediction of the active user's D. Matrix Factorization
preference. CF requires a large amount of information to be
Collaborative filtering recommendation method has been
processed, including large-scale data sets such as the e-
widely used in the nearest neighbour (NNH) algorithm
commerce and web applications. Over the past few decades
combined with the pearson correlation or cosine similarity to
CF has been continuously improved and has become one of
calculate the predicted values. In terms of sparse rating, NNH
the most prominent personalization techniques in the field of
approach usually experience difficulties in finding the right
recommendation systems.
match and consequently produce a weak recommendation with
Fig. 2. show an example of a collaborative filtering. Full-
accuracy. Furthermore NNH calculation algorithm complexity
line arrows represent the relationships were already known.
tends to increase with number of users and the number of
Both subjects had ties with the goods and therefore they were
items, which means the recommendation system will have
considered to have similarities. Recommendations marked
serious problems in scalability [6].
with dotted arrows. An object can be recommended for
To overcome this problem, a matrix factorization approach
selected of objects related to the same subject, so the subject
has been proven to be an efficient method for rating-based
was not related to the object.
recommendation [10].

III. METHODOLOGY
There were several things that need to be done before the
recommendation can be made. Firstly, it needsto extract the
dataset, in order to process recommendations. This section
will explain how the datawill be used, a hybrid
Fig. 2:Collaborative filtering [4] recommendation techniques and evaluation methods.

Based on the approach of unknown relationships, A. Dataset Analysis

collaborative filtering can be divided into two groups: a The Movielens dataset was used in this research, contains
memory-based and model-based. 855,598 ratings for 10197 movies and 2113 users. This dataset
was an extension of Movielens10M dataset, published by
Groupleans research group. It links the movies of Movielens
C. Hybrid Recommendation Method dataset with their corresponding web pages at Internet Movie
Hybrid method combining two or more methods of Database (IMDb)and Rotten Tomatoesmovie review systems.
recommendation with the aim to improve recommendation The dataset was released in HetRec 2011. The ratings data
results. There were a number of approaches to incorporate represent as a list of triples userID, itemID and rating. Split
practical advice among collaborative filtering, content-based the dataset into two parts, 80% for training and 20% for
filtering, knowledge-based demographics and testing the algorithm. The dataset contains actors, countries,
recommendations. The most popular was a combination of directors and genres. For experiment purpose, it only use the
content-based filtering and collaborative filtering [4]. data rating from users to make recommendations. Fig. 4. show
The combination of memory-based collaborative filtering how the data format look like.
and collaborative filtering-based models were often used in
contemporary commercial recommendations.Fig. 3. show an
example of hybrid method.

2
 between matrix factorization and the the combination of
matrix factorization and nearest-neighbour (NNH). Combining
two or more technique was called hybrid [11]. The objective
of hybrid was to increase the recommendation accuracy and
reduce the error. Fig. 2. shows the architecture of hybrid.

Fig. 4: Rating data format

Data Contributing
Our main focus was on the userID, movieID and rating. Training Recommender
Those field will convert into matrix two-dimensional space,
first dimension was the number of users and second dimension
is the number of films (M2113 x 10197). The valueswere
represented with numeric. However, the result of matrix M Data Actual
Testing Recommender Prediction
will be sparse, mostly the user-item value was empty field.
Thus, only the available rating data will used in training
algorithm.
Fig. 5: Hybrid Recommendation System Architecture
B. Matrix Notation [11]
Our experiment willfocus on the film and the user entity.
Movie data, and user ratings were available in different data Fig. 5. shows the big picture of hybrid recommendation
formats. Each file consists of fields that were interrelated and system works. The underlying from its architecture was data
can be extracted to form a matrix notation that will be used to training and data testing. The ratio of data training and data
make recommendation system. Table1. shows the matrix testing were usually 80:20, however, some papers also use
notation that will be used to recommendation system. 70:30. Both of the data will be used to produce the prediction
value.
TABLE 1: Matrix notation D. Evaluation Method
user/item film1 film2 film3 n-films The recommendation systems accuracy depend on several
user1 1 - 2 - factors. Such as, number of data, k-values, normalization, and
user2 2 3 - - many others. There were a lot of performance metrics that can
user3 - 2 1 3 be used in research. Generally, it will be used to evaluate and
user4 5 4 - 1 analyze the system performance. Some of metric tools to
user5 3.5 1.5 5 2 evaluate recommender system according to microsoft research
were Root Mean Square Error (RMSE) and Mean Square
user6 2.5 2 3 1
Error (MAE) [20]. These metric will be used in our research,
n-user 1 3 - 1 in order to computes how close the estimates were to the
values actually observed. This experiment want to minimize
From the table above, rows represented the number of the prediction Root Mean Square Error (RMSE) on test sets:
users and columns represented the number of movies.
Whereas, the cell represented the rating value that user given.
ே
In table above show the construction of the matrix. Then, that ͳ
matrix will be used into recommendation system. The matrix ܴ‫ ܧܵܯ‬ൌ  ඩ ෍ሺ‫ݎ‬ҧ௜ െ ‫ݎ‬௜ ሻଶ
ܰ
dimension is number of users multiple number of movies. ௜ୀଵ
Using matrix factorization can find the missing values from
matrix notation. In equation above ‫ݎ‬ҧ௜ denotes the predicted rating and riwere the
C. Hybridization actual rating of the movie. Whereas, the mathematical notation
for MAE is:
Various techniques have been proposed as a basis for
recommendation systems: collaborative filtering, content- ͳ
based filtering, knowledge-based screening and demographic ‫ ܧܣܯ‬ൌ  ෍ ȁ‫ݎ‬Ƹ௨௜ െ ‫ݎ‬௨௜ ȁ
ȁܶȁ
techniques. Each of these techniques has its drawbacks, such ሺ௨ǡ௜ሻ‫்א‬
as the cold-start problems in collaborative filtering and
content-based filtering [11]. From the equation above ‫ݎ‬Ƹ௨௜ is prediction value from user u to
There a lot of technique that can be used in item i dan ‫ݎ‬௨௜ is actual value.
recommendation system, the mostly used was nearest-
neighbour, item-based filtering [8] and matrix factorization
[12]. Matrix factorization will be used in this paper. Then,
several experiments to compare the prediction accuracy

3
 IV. RECOMMENDER SYSTEM
In this section, the step-by-step of proposed method will be A rating prediction approximation will very close to the
more details. It will describe more about the recommendation original rating values, and it also deliver some predictions of
technique, data normalization, matrix factorization and its the unknown values or missing values. Neighbourhood
combination with classic collaborative filtering. algorithm uses the ratings of the similar users (or items) to
predict the values of the input matrix. The only difference with
A. Normalization plain SVD is the way how it computes the predictions. To
The data rating from MovieLens, it was a range between compute the prediction, it uses this equation:
rating. The range was from 1 until 5 (0, if no rating). The
normalizationwas used to remove the bias. One common σ௝‫א‬ௌ ೖ ሺ௜Ǣ௨ሻ ܵ௜௝ ‫ݎ‬௨௝
method of normalization involves having values of each ‫݃݊݅ݐܽݎ‬ሺ‫ݑ‬ǡ ݅ሻ ൌ 
σ௝‫א‬ௌ ೖ ሺ௜Ǣ௨ሻ ܵ௜௝
feature range from 0 to 1.
In order to increase the prediction accuracy, the “global Where Sk (i; u) is the set of k that rate by u, which were most
effect” has to be removed. Generally the combination of user, similar to i. Sij is the similarity between i and j.
item and all average rating value (‫ݎ‬ҧ , ‫ݎ‬ҧ‫ݎ& ݑ‬ҧ݅ ) were substracted
from the original data ruito remove the popularity effect [13].
D. Nearest Neighbour
‫ݎ‬෤‫ ݅ݑ‬ൌ ‫ ݅ݑݎ‬െ ߙ‫ݎ‬ҧ െ ߚ‫ݎ‬ҧ ‫ ݑ‬െ ߛ‫ݎ‬ҧ ݅ Another method that will be used in this paper was the
nearest neigbour approach. Nearest neighbor was a classical
collaborative filtering technique that still used in
After normalize the data, then perform matrix factorization
recommendation systems. This technique will predict the
to get predictions and recommendations.
value and compare it with matrix factorization. Performed by
B. Matrix Factorization this technique was to find the similarity to the films or users.
In many cases, matrix factorization was used to remove the From the experiments carried out in section 5.2 proves that the
dimension of the item space and retrieve latent relations combination of matrix factorization and nearest neighbor can
between items of the dataset [12], [14]. Based on the improve the prediction accuracy for better recommendation
demonstration on Neflix Grand Prize, matrix factorization system.
model was the best nearest-neighbour technique in V. EXPERIMENTS
recommendation system.
Recommendation systems have a large enough data and In this section we carried out tests on datasets that have
very sparse, this causes a slow process and computational been mentioned above. Tests performed several times on the
predictions inaccurate. So therefore, it need special technique value of k to determine the accuracy of the results of the
in order to minimize the spread of data and speed up the methods that have been mentioned in section IV.
computation process, it can be done by matrix factorization A. MF Experiment
[12]. There were many methods used to process factorize the
The most difficult thing to get better accuracy was to
matrix, such as Singular Value Decomposition (SVD), Non-
determine the optimal k values as paramater in computation.
negative Matrix Factorization (NMF), and others. SVD
Therefore, multiple test by iterating the k valueswere
approach will be used in this paper, it factorize matrix rating
performed.
to be three low dimensional space, left-singular vector (V),
As ilustrated in Fig. 4, the prediction accuracy was strongly
singular value (S) and right-singular vector (W).
depend on the k values.
்
‫ܪ‬௞ ൌ ܸ௠௫௞ Ǥ ܵ௞௫௞ Ǥ ܹ௞௫௡
0.85
Multiple k examples were used to make the experiment and 0.8
investigate the result for next experiment. 0.75
Results

0.7
C. Predictions 0.65
0.6
The comparisonof prediction value and original rating 0.55
were performed in order to get the evaluation about how 0.5
accurate the algorithm works. To predict rating rui, SVD class 5 10 15 20 25 30 35 40 45 50
reconstructs the original matrix
K
‫ܯ‬ǯ ൌ ܷσ௞ ܸ ்
MAE RMSE
And the rating prediction equals to:

‫݃݊݅ݐܽݎ‬ሺ‫ݑ‬ǡ ݅ሻ  ൌ ‫ܯ‬ǯ݆݅ Fig. 6: MF in k-dimensional space

4
 From Fig. 6.it shows that matrix factorization can result in factorization method, the smallest RMSE value was 0.789649,
an average prediction accuracy well.From several tested k, which k = 35. While for MF + CF method, the smallest RMSE
obtained the best of RMSE and MAE, respectively were 0,78 value was 0.788649 k = 30.
and 0,59 with k values were 20 and 35. Results metrics that
close to 0 were the most good. VI. CONCLUSIONS
The combination of MF and CF can indeed be used to
B. MF + CF Experiment make predictions and recommendation systems. But it doest
not have a huge if we compare it with the MF only. In
After getting the results of experiments with matrix experiment we have done, the difference were only 0.001.
factorization method, the next experiments will be performed Of course, there were many other parameters that can
by using the combination of matrix factorization and
determine the accuracy of predictions, such as the number of
collaborative filtering.
datasets, content features, normalization, and others. With the
It needs neighbourhood algorithm, which use the similarity
existance of this paper we hope it can enrich the knowledge in
values of user ratings to predict the value of the input matrix.
the field of recommendation systems and can be used in the
The prediction accuracy resulting from experiment using MF future for further experiments to improve the prediction
+ CF based on the value of k ilustrated in the Fig. 5. accuracy.

0.85 ACKNOWLEDGMENT
0.8 For the finishing of this experiment, thank to Mr. Bens
0.75
Results

Pardamean, M.Sc, P.hD., for any criticsm and suggestions.

0.7
0.65 And also thank to Grouplens for the dataset and for the rec-sys
0.6 python framework, thanks to Ocelma.
0.55
0.5
5 10 15 20 25 30 35 40 45 50 REFERENCES
K [1] Yehuda Koren. (2009) The BellKor Solution to the
Netflix Grand Prize. [Online].
MAE RMSE http://www.netflixprize.com/assets/GrandPrize2009_B
PC_BellKor.pdf
[2] Michael J Pazzani and Daniel Billsus, "Content-Based
Fig. 7: MF+ CF in k-dimensional space Recommendation Systems," The Adaptive Web Lecture
Notes in Computer Science, vol. 4321, pp. 325-341,
From the fig. 7. shows that the value of k = 30 was the lowest, 2007.
meaning that k = 30 was the most optimum done in this case.
[3] Michael D Ekstrand, John T Riedl, and Joseph A
When we compared with the matrix factorization (MF) in
Konstan, "Collaborative Filtering Recommender
previous experiment, the combination of MF & CF has a
Systems," Foundations and Trends in Human–
better accuracy rate, were proven by the smaller RMSE value
Computer Interaction, vol. 4, no. 2, pp. 81 - 173, 2010.
or close to 0. All the experimental results can be seen in the
Table 4. [4] Gediminas Adomavicius and Alexander Tuzhilin,
"Towards the Next Generation of Recommender
TABLE 2: Comparison of RMSE metrics for both Systems: A Survey of the State-of-the-Art and Possible
methods Extensions," IEEE Transaction On Knowledge and
K MF MF + CF Data Engineering, vol. 17, no. 6, pp. 734-749, 2005.
5 0,800966 0.810545 [5] Yehuda Koren, "Factorization meets the neighborhood:
10 0,798626 0.79244 a multifaceted collaborative filtering model," in
15 0,793362 0.803481 Proceeding of the 14th ACM SIGKDD international
20 0,789719 0.794638 conference on Knowledge discovery and data mining,
25 0,791834 0.79704 New York, NY, USA, 2008, pp. 426-434.
30 0,789986 0.788649 [6] Yehuda Koren, "Factor in the Neighbors: Scalable and
35 0,789649 0.80757 Accurate Collaborative Filtering," ACM Transactions
40 0,79193 0.802361 on Knowledge Discovery from Data, pp. 1-24, 2010.
45 0,792893 0.804966 [7] Abhinandan Das, Mayur Datar, and Ashutosh Garg,
50 0,791348 0.803322 "Google News Personalization: Scalable Online
Collaborative Filtering.," in International World Wide
Table 2. above show the experimental results of the Web Conference Committee (IW3C2), 2007, pp. 271-
comparison between two proposed methods. For the matrix 280.

5
[8] Greg Linden, Brent Smith, and Jeremy York, "Matrix Factorization Techniques for Recommender
"Amazon.com Recommendation: Item-To-Item Systems.," in IEEE Computer Society., 2009, pp. 42-49.
Collaborative Filtering," 2003. [13] Upendra Shardanand and Pattie Maes, "Social
[9] Paolo Cremonesi, Roberto Turrin, and Fabio Airoldi, information filtering: algorithms for automating “word
"Hybrid algorithms for recommending new items," in of mouth”," in Proceedings of the SIGCHI Conference
Proceedings of the 2nd International Workshop on on Human Factors in Computing Systems, New York,
Information Heterogeneity and Fusion in 1995, pp. 210-217.
Recommender Systems, New York, 2011, pp. 33-40. [14] Xiaoyuan Su and Taghi M Khoshgoftaar, "A Survey of
[10] G Takács, I Pilászy, B Németh, and Domonkos Tikk, Collaborative Filtering Techniques," Journal Advances
"Investigation of Various Matrix Factorization Methods in Artificial Intelligence, pp. 1-19, 2009.
for Large Recommender Systems," in IEEE
International Conference on Data Mining Workshops, .
Pisa, 2008., pp. 553-562.
[11] Robin Burke, "Hybrid Web Recommender Systems," in
Vol. 4321 of Lecture Notes in Computer Science. Berlin
Heidelberg: Springer-Verlag, 2007, p. 377.
[12] Yehuda Koren, Robert Bell, and Chris Volinsky,