International Journal of Computer Applications Technology and Research

Volume 6Issue 2, 97-100, 2017, ISSN:-23198656

Machine Learning Algorithms for Recommender

System - a comparative analysis

Satya Prakash Sahu Anand Nautiyal Mahendra Prasad

University of Hyderabad University of Hyderabad University of Hyderabad
Hyderabad, India Hyderabad, India Hyderabad, India
[email protected] [email protected] [email protected]

Abstract: Recommendation system is one of the most popular applications of Artificial Intelligence which attracts many researchers all
over the globe. The advent of the Internet era has brought wide implementation of recommendation system in our everyday lives. There
are many machine learning techniques which can be used to realize the recommendation system. Among all these techniques we are
dealing with Content Based Filtering, Collaborative Based Filtering, Hybrid Content-Collaborative Based Filtering, k-mean clustering
and Naive Bayes classifier. We have exploited these algorithms to their extreme in order to achieve the best possible precision and have
presented a comprehensive comparative analysis. The strength of all these algorithms can be clearly realized by the significant
enhancement in the accuracy, depicted by the experimental analysis taking cold start problem into consideration.

Keywords: Recommender System, Classifier, Content Based, Collaborative Based, Cluster, Correlation.

1. INTRODUCTION In this paper, we have used the MovieLens dataset[8]. All the
Recommendation system[10] is an application which is used above algorithms deal with this dataset in order to recommend
for prediction in various domains throughout the internet. A the movies and calculate the precision along with tackling the
large amount of data flows through the internet and it gives cold-start problem[3]. Cold-start problem is one of the most
away a lot of information regarding the user searching activity. commonly encountered challenges of the recommendation
The information extracted from the pattern of previously system. It is also known as the new user problem as it creates
searched data can be molded into the prediction of relevant data problem of generating recommendations for the new user. We
for the user[1]. The implementation of the system can be have divided this analysis into various sections. Section II
performed by various techniques. In this paper, we have describes the different state-of-the-art techniques for the
discussed Content Based Filtering, Collaborative Filtering[10], recommendation system. Section III gives the experimental
Hybrid Content-Collaborative Based Filtering, k-mean results for all these techniques. Section IV concludes the study.
clustering Based and Naive-Bayes Classifier based techniques. Section V describes the future work that we propose.

The Content Based Filtering approach takes into account a 2. ALGORITHMS

users profile which is constructed based on his previous 2.1 Content Based Filtering
ratings[2]. His ratings determine his inclination and interests,
forming the basis for recommending a new item. A higher The Content Based Filtering considers the items rated by a user
rating denotes a higher likelihood of the user to visit similar to formulate the future recommendations while exploring the
items. So, a new item is recommended according to the internet services. A user tends to rate an item which he likes or
maximum number of ratings given by the user in a genre[3]. dislikes. His ratings reflect his response towards that item. If he
In the Collaborative Based Filtering, recommendation for a likes an item, he rates it higher and lesser ratings denote that he
user is governed by other users profiles. An item is is not much interested. These rated items serve as the 'content'
recommended based on the ratings of other users who have in the Content Based Filtering[2][3]. Based on this content, the
similar interests as the user under consideration[2][4]. In user is recommended future items which he might approve of.
another approach, the content and collaborative based filtering Here, the user is recommended movies which fall in a particular
are combined to form the Hybrid Content-Collaborative Based genre of his liking.
Filtering. It includes the advantages of both the methods and Algorithm 1. Content Based Filtering
outperforms both of them.
Input: users X, movies m, rating r, movie genre mg , Number
In the k-mean clustering, the similarity between the objects is of movies to be recommended().
calculated by the means of various distance measures such as
Euclidean distance[5], Pearson Correlation, etc. The value of k Output: Recommended movies R
determines the number of clusters to be formed[6]. The nearest
1. for all users do
k objects are the most similar to one another. These clusters of
similar objects drive the recommendation of new arriving 2. Select seen movies s, unseen movies s', association
objects. Naive Bayes is another popular and efficient classifier of unseen movies asi' w.r.t X, association of each
based on Bayes theorem. It is a conditional probability based genre agj w.r.t s' , where i is 1 to n and j is 1 to m.
classifier. The prior knowledge of the classifier assists learning.
The naive assumption is that the features are conditionally 3. Calculate scorej .
independent[7]. 4. Select highest three scorej
5. Select m' s' according to highest three scorej

www.ijcat.com 97
 International Journal of Computer Applications Technology and Research
Volume 6Issue 2, 97-100, 2017, ISSN:-23198656

6. Calculate score me' where e m' 3. Calculate scorej .

7. Return top score recommendations. 4. Select highest three scorej
8. end for 5. Select m'' s of the ith user according to highest
In this algorithm, the notations used have the following three scorej
meaning : association of each movie asi represents total 6. Find similarity (simj) w.r.t m''
number of users who rated movie i s', association of each
genre agj represents total number of movie belonging to genre 7. Select highest simj user.
j.
8. Select m' according to its highest three scorej s of
scorej = agj / m user obtained in step 7 and s' of the ith user under
score(me') = ame / total count of m' consideration.
9. Calculate weight W(me') where e m'
2.2 Collaborative Filtering
10. Return top weight recommendations.
There can be many users who must be having the same pattern
of rating an item as the user intended. This similar pattern of 11. end for
their ratings with the user guides the Collaborative
Filtering[2][3][10]. The notion behind the Collaborative 2.4 K-Mean Clustering
Filtering is the recommendation of an item based on the
preferences of like-minded users.
The k-mean is a non parametric classification technique. It
Algorithm 2. Collaborative Filtering
distributes the items into k clusters according to their proximity
Input: users X, movies m, rating r, Number of movies to be to one another. In this paper, this proximity is being measured
recommended(). by using the Euclidean distance[11]. For calculating the
Output: Recommended movies R. Euclidean distance we have taken rated and unrated movies as
binary. Each cluster possesses a centroid which is the mean of
1. for all users do
all the items in the cluster. All the objects in a cluster move
2. Select seen movies s, unseen movies s' towards the centroid and the centroid is updated in each
3. Find similarity (simi) w.r.t s, where i = 1 to n. iteration. The iteration continues until a saturation point arrives,
when the centroid stops altering. By following this approach
4. Select highest simi user
we are decreasing the search space which results in reduced
5. Select m' s of user obtained in step 4 and s' of ith computational complexity[6]. These computations are
user. performed off-line which helps the classification to be efficient
in terms of time complexity.
6. Calculate weight W(me') where e m'
7. Return top weight recommendations. Algorithm 4. k-mean clustering
8. end for
Input: users X, movies m, rating r, Number of movies to be
In this algorithm, the notations used have the following recommended , value of k.
meaning : simi represents common movies between user i and
other users. Output: Recommended movie R.
weight(me')= rating of particular moviee / max rating.
1. begin
2.3 Hybrid Filtering
To cater better precision, a hybrid filtering method is used 2. Randomly select k centroids.
which can provide the advantages of both the content and the
collaborative approaches[4] and can overcome their 3. Calculate euclidean distance (eucd) for X from k
shortcomings. Suppose, the user appreciates mostly movies in centroids.
g G genres, and the collaborating users also give high ratings
to the g G genres, then g will be taken as the metric to 4. Allocate X to kth cluster according to eucd.
recommend movies to the user.
5. Update centroid for each cluster with (summation(ki )
Algorithm 3 .Hybrid Filtering from 1 to p)/p, where p is the number of members in
Input: users X, movies m, rating r, movie genre mg , Number ki cluster
of movies to be recommended().
6. Repeat step 3 to step 5 until centroid(t) centroid
Output: Recommended movies R.
(t+1).
1. for all users do
2. Select seen movies s, unseen movies s', association 7. for all users do
of each genre agj w.r.t s', where i is 1 to n and j is 1
to m. 8. Select seen movies s, unseen movies s'.

www.ijcat.com 98
 International Journal of Computer Applications Technology and Research
Volume 6Issue 2, 97-100, 2017, ISSN:-23198656

9. Find similarity (simi) w.r.t s, where i = 1 to p. recommendations (T), the ones which are also present in the
converted movies are the correct recommendations(tc).
10. Select highest simi user.
Precision = (tc / T ) * 100
11. select m' s of highest simi and s' of ith user.
For all the experiments, we are taking value of = 5 and value
12. Calculate weight W (me') where e m' of k = 10.
13. Return top weight recommendations.

14. end for Algorithm/Size 10K 50K 100K

15. end
Content Based 18.45 18.66 19.10

2.5 Naive Bayes

Collaborative 17.97 18.69 19.95
The Naive Bayes is based on the Bayes theorem. The
probabilistic approach followed by Naive Bayes Classifier Hybrid 20.31 21.03 22.20
determines the probability of the classification and helps in
finding the uncertainty about the model[9]. It is an efficient K-Mean 21.05 21.93 22.67
learning algorithm which uses the prior knowledge of the
observed data. The Naive assumption is that the features are Naive Bayes 24.62 25.19 25.73
conditionally independent[1].
Table 1. Precision of Different Algorithms.
Algorithm 5. Naive Bayes

Input: users X, movies m, rating r, number of movies to be

recommended()

Output: Recommended movies R.

1. for all users do

2. Select seen movies s, unseen movies s' .

3. Find similarity (simi) w.r.t s, where i = 1 to n.

4. Select x' X where simi > 10.

5. Calculate association of unseen movies asi ' w.r.t to

x'

6. Calculate score (se') where e s'.

7. Return top score recommendations.

8. end for
Fig. 1. Precision Comparison

4 CONCLUSION
3 EXPERIMENTAL RESULT
All the algorithms described in this paper are compared with
We now illustrate the analysis of the experiments performed
respect to their precision rates. This comprehensive analysis
and provide a comparison of all the state-of-the-art methods
depicts the strength and the weakness of each one of them in
described above. To compare their accuracy we have used the
different versions of the MovieLens dataset. The experiments
MovieLens dataset of 10K, 50K and 100K. The dataset varies
performed are the witness of the sparsity handling by these
in sparsity. For example, the 100K MovieLens dataset has
algorithms. Our experiments have shown promising results and
100K ratings, 943 users and 1682 movies of 19 different
this paper conforms that out of all these approaches Naive
genres. The analysis of these algorithms is demonstrated based
Bayes gives the best precision.
on precision measure. For each test user, we convert 30% of the
users seen movies into unseen movies and apply the
algorithms described above. Out of the total number of

www.ijcat.com 99
 International Journal of Computer Applications Technology and Research
Volume 6Issue 2, 97-100, 2017, ISSN:-23198656

5 FUTURE WORK [4] Shinde U., and Shedge R. Comparative Analysis of

Collaborative Filtering Technique, IOSR Journal of
Computer Engineering, Volume 10, pp 77-82 2013.
With this paper, we have achieved encouraging results from all
[5] CHENG X., WANG J., Danqian LU. Research of
these algorithms. In the real time sophisticated Question Analysis Based on HNC and K Nearest
recommendation systems there is a need of high accuracy. Such Neighbor, Journal of Computational Information Systems,
systems still have space for improvement. There are several 6:10, pp 3449-3455, 2010.
machine learning algorithms which can be applied to these real [6] Campos P.G., Bellogn A., Dez F., and Chavarriaga J.E.
time systems. It is worthwhile to examine those other Simple Time-Biased KNN-based recommendations,
algorithms to improve the precision further. ACM, 978-1-4503-0258-6, 2010.
[7] Puntheeranurak S., and Pitakpaisarnsin P. Time-aware
6. REFERENCES Recommender System Using Nave Bayes Classifier
Weighting Technique, International Symposium on
[1] Katore L.S., and Umale J.S. Comparative Study of
Computer, Communication, Control and Automation,
Recommendation Algorithms and Systems using WEKA,
3CA, pp 266-269 ,2013.
International Journal of Computer Applications, Volume
[8] Bellogn A. Castells P. and Cantador I. Precision-Oriented
110 No. 3, pp14-17, 2015.
Evaluation of Recommender Systems: An Algorithmic
[2] Tewari A.S., Kumar A., and Barman A.G.,. Book
Comparison, ACM, 978-1-4503-0683-6, 2011.
Recommendation System Based on Combine Features of
[9] Puntheeranurak S. and Sanprasert S. Hybrid Naive Bayes
Content Based Filtering, Collaborative Filtering and
Classifier Weighting and Singular Value Decomposition
Association Rule Mining, IEEE, 978-1-4799-2572-8, pp
Technique for Recommender System, IEEE, 978-1-4244-
500-503, 2014.
9698-3, pp 473-476, 2011.
[3] Wanaskar U.H., Vij S.R., and Mukhopadhyay D. A
[10] Laishram A., Sahu S.P., Padmanabhan V., and Udgata S.
Hybrid Web Recommendation System Based on the
K. Collaborative Filtering, Matrix Factorization and
Improved Association Rule Mining Algorithm, Journal of
Population Based Search: The Nexus Unveiled. ICONIP ,
Software Engineering and Applications, 6, pp 396-404
Part III, LNCS 9949, pp. 352361, 2016.
2013.
[11] Prasad M. and Singh A. A Novel Hybrid Ant Colony
Optimization Approach to Terminal Assignment Problem.
ACM, AICTC 16, Article 29, August 12-13, 2016.

www.ijcat.com 100