IJRECE VOL.

7 ISSUE 2 (APRIL- JUNE 2019) ISSN: 2393-9028 (PRINT) | ISSN: 2348-2281 (ONLINE)

OPTIMIZATION OF RECOMMENDATION SYSTEM

USING MACHINE LEARNING APPROACH
Anmol Alawadhi1, Vikas Thada2 & Khushboo Tripathi3
1
M.Tech(AI&R), Deptt. of Computer Science & Engineering, Amity University Haryana, India
2
Associate Professor, Deptt. of Computer Science & Engineering, Amity University Haryana, India
3
Assistant Professor, Deptt. of Computer Science & Engineering, Amity University Haryana, India

Abstract
Recommendation system plays important role in Internet Keywords - Recommendation, Collaborative filtering, Model
1 1 1 1 1 1

world and used in many applications. It has created the based, Memory based, Content based, Hybrid.
1 1 1 1 1

collection of many application, created global village and

I. INTRODUCTION 1

growth for numerous information. This paper represents the

Recommendation System is part of Daily life where people
overview of Approaches and techniques generated in
rely on knowledge for making decision of their personal
recommendation system. Recommendation system is
interest. Recommendation system is subclass of information
categorized in three classes: Collaborative Filtering, Content
filtering to predict preferences to the items used by or for
based and hybrid based Approach. This paper classifies
users. Although there are many approached developed in past
collaborative filtering in two types: Memory based and Model
but search still goes on due to it’s often usage in many
based Recommendation.The paper elaborates these
applications, which personalize recommendation and deals 1

approaches and their techniques with their limitations. This

with information overload. These demands throws some
1 1 1 1 1 1 1

survey shows the road map for research in this area.

challenges so different approaches like memory based, model
1 1 1 1 1 1 1 1

In the very recent years, development of recommendation

1 1 1 1 1 1 1 1

based are used. Recommender system still requires

1 1 1 1 1 1 1

system has been a more heated problem due to a higher level of

1 1 1 1 1 1 1 1 1 1 1 1 1

improvement to become better system. 1 1 1 1 1

data consumption and the advancement of machine learning

1 1 1 1 1 1 1 1

Recommendation system is a sharp system that provides idea 1 1 1 1 1 1 1 1 1

techniques. Some traditional approaches such as collaborator

1 1 1 1 1 1 1

about item to users that might interest them some examples are
1 1 1 1 1 1 1 1 1 1 1

filtering, has been widely used in recommendation systems,

1 1 1 1 1 1 1 1

amazon.com, movies in movielens, music by last.fm. In this

1 1 1 1 1 1 1 1 1

have helped recommendation system to give users a quick

1 1 1 1 1 1 1 1 1

paper different approached with their techniques are mentioned

1 1 1 1 1 1 1

access to the data. However, collaborative filtering or content

1 1 1 1 1 1 1 1 1

1to compare the limitation of each technique in proper manner

1 1 1 1 1 1 1 1 1 1

based filtering have limitations in giving a better result with the

1 1 1 1 1 1 1 1 1 1 1

to provide proper future recommendations.

1 1 1 1 1

ignorance of combination factor of lyrics and genre.

1 1 1 1 1 1 1 1

In my paper, I will propose an improved algorithm based on

1 1 1 1 1 1 1 1 1 1 1
1.1 PROBLEM STATEMENT
1 1 1

machine learning on hybrid approach using collaborative

1 1 1 1 1 1 1

This paper is based on recommendation system that

filtering, content based filtering and popularity based filtering.
1 1 1 1 1 1 1 1

recommends different things to users. This system will

The proposed method will make it possible that it could make
1 1 1 1 1 1 1 1 1 1 1

recommend movies to users. This system will provide more

recommendations in a large system to make comparisons by 1 1 1 1 1 1 1 1 1

precise results as compared to the existing systems. The

“understand” the content of data. In this paper, I propose an 1 1 1 1 1 1 1 1 1 1 1

existing system works on individual users’ rating. This may be

end-end model, which is based on machine learning approach
1 1

1 1 1 1 1 1 1 1 1

sometime useless for the users who have different taste from
to predict user’s next most possible data by
1 1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1

the recommendations shown by the system as every user may

similarity.Experiments made and evaluations based on Dataset
1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1

have different tastes. This system calculates the similarities

and demonstrate how it outperformed the traditional methods.
1 1 1 1 1 1 1 1

1 1 1 1 1 1 1

INTERNATIONAL JOURNAL OF RESEARCH IN ELECTRONICS AND COMPUTER ENGINEERING

A UNIT OF I2OR 2025 | P a g e
 IJRECE VOL. 7 ISSUE 2 (APRIL- JUNE 2019) ISSN: 2393-9028 (PRINT) | ISSN: 2348-2281 (ONLINE)
between different users and then recommend movie to them as
1 1 1 1 1 1 1 1 1 1 order to overcome the limitations of both approach hybrid
1 1 1 1 1 1 1 1 1

per the ratings given by the different users of similar tastes.

1 1 1 1 1 1 1 1 1 1 1 systems are proposed that combines both approaches in some
1 1 1 1 1 1 1 1 1

This will provide a precise recommendation to the user. This is

1 1 1 1 1 1 1 1 1 1 1 manner [15]. 1

a web based as well as android system where there is a movie

1 1 1 1 1 1 1 1 1 1 1 1 1

web service which provides services to user to rate movies, see

1 1 1 1 1 1 1 1 1 1 1
2.2 Collaborative filtering system
1 1 1 1

recommendations put comments and see similar movies. 1 1 1 1 1 1

Collaborative filtering systems work by collecting user remark
in the form of ratings for items in a given field and exploiting
1.2 AIM 1
similarities in rating actions amongst several users in
determining how to recommend an item. Collaborative
To develop recommendation system using machine learning
filtering systems recommend an item to a user based on
approach
opinions of other users. Like, in a movie recommendation
To
application, Collaborative filtering system tries to find other
1.3 Objective 1
like-minded users and then recommends the movies that are
most liked by them. Although there are many collaborative
• To study background related to recommendation and filtering techniques, they can be divided into two major
machine learning approach categories [2]:
: 1

• To design and implement methodology for  Memory Based approaches 1 1 1

recommendation system using machine learning

 Model Based approaches 1 1 1

approach

• To analyse the results for the parameters: Popularity 2.2.1 Memory based Approach
1 1 1 1

based system, Content based system, Collaborative 1

Memory-based techniques continuously analyze all user or 1 1 1 1 1 1 1

Filtering based system using hybrid approach item data to calculate recommendations and can be classified in
1 1 1 1 1 1 1 1 1

1the following main groups: CF techniques, Content-Based

1 1 1 1 1 1 1

II. BACKGROUND 1
(CB) techniques and hybrid techniques. CF techniques
1 1 1 1 1 1 1

A variety of approaches has been used to provide

1 1 1 1 1 1 1 1 1
recommend items that were used by similar users in the past; 1 1 1 1 1 1 1 1 1 1 1

recommendation like collaborative filtering, content based and 1 1 1 1 1 1 1

they base their recommendations on social, community-driven
1 1 1 1 1 1 1

hybrid approach. Different Algorithms and approaches are

1 1 1 1 1 1 1
information (e.g., user behavior like ratings or implicit 1 1 1 1 1 1 1 1

there to provide recommendation that may use rating or content

1 1 1 1 1 1 1 1 1
histories). CB techniques recommend items similar to the ones
1 1 1 1 1 1 1 1 1

1information; however collaborative filtering and content based 1 1 1 1 1 1 1

the learners preferred in the past; they base their
1 1 1 1 1 1 1 1 1

method suffer from same limitations. Several researchers have

1 1 1 1 1 1 1 1
recommendations on individual information and ignore the 1 1 1 1 1 1 1

tried to overcome these limitations by combining both

1 1 1 1 1 1 1 1
offerings from other users. Hybrid techniques combine both
1 1 1 1 1 1 1 1

collaborative filtering and content based method as a hybrid 1 1 1 1 1 1 1 1 1

techniques to provide more accurate recommendations. A 1 1 1 1 1 1 1

approach that combined ratings as well as content information.

1 1 1 1 1 1 1 1 1
hybrid RS could combine CF (or social-based) techniques with
1 1 1 1 1 1 1 1 1

Recommendation system will always remain active search area 1 1 1 1 1 1 1 1

CB (or information-based) techniques. If no efficient
1 1 1 1 1 1 1

for researchers [1].

1 1 1
information is available to carry out CF techniques, it would 1 1 1 1 1 1 1 1 1 1

2.1 Approaches of Recommendation System

1 1 1 1 1
switch to a CB technique [4].
1 1 1 1 1

Recommendation system is usually classified on rating 1 1 1 1 1 1 1

estimation 1

 Collaborative Filtering system 1 1 1

 Content based system 1 1 1

 Hybrid system 1 1

In content-based approach, similar items to the ones the user

1 1 1 1 1 1 1 1 1 1

preferred in past will be recommended to the user while in 1 1 1 1 1 1 1 1 1 1 1

collaborative filtering, items that similar group people with 1 1 1 1 1 1 1 1

similar tastes and preferences like will be recommended. In

1 1 1 1 1 1 1 1 1

INTERNATIONAL JOURNAL OF RESEARCH IN ELECTRONICS AND COMPUTER ENGINEERING

A UNIT OF I2OR 2026 | P a g e
 IJRECE VOL. 7 ISSUE 2 (APRIL- JUNE 2019) ISSN: 2393-9028 (PRINT) | ISSN: 2348-2281 (ONLINE)
Fig 1 Block Diagram of Memory Based RS [4]
1 1 1 1 1 1 1 1 1 the range of preference scores for CF is normalized, the
1 1 1 1 1 1 1 1 1 1

The prediction process in memory-based CF contains three

1 1 1 1 1 1 1 1 Spearman Rank Correlation in the CF field shows comparable 1 1 1 1 1 1 1 1 1

steps. They are similarity evaluation, generation of nearest

1 1 1 1 1 1 1 1 performance to the Pearson Correlation[8] 1 1 1 1 1

neighborhoods and score prediction. For evaluation of the 1 1 1 1 1 1 1 1 B) Formation of Nearest Neighbor
1 1 1 1 1

performance, the CF system considers the mean absolute error 1 1 1 1 1 1 1 1 1 The second step after the similarity evaluation is generation of
1 1 1 1 1 1 1 1 1 1

(MAE), precision and recall. The CF performance varies 1 1 1 1 1 1 1 1 nearest neighborhoods. To improve performance, many 1 1 1 1 1 1

according to the processing method of each step[5]. 1 1 1 1 1 1 1 1 methods have been proposed by CF researchers. The methods 1 1 1 1 1 1 1 1 1

A) Existing Similarity Measures

1 1 1 1 for selecting nearest neighborhoods include classification using
1 1 1 1 1 1

The most important first step in memory-based CF is similarity

1 1 1 1 1 1 1 1 1 1 1K-means, a threshold for the number of common rating items 1 1 1 1 1 1 1 1 1 1

evaluation. The CF system in this step evaluates the similarity 1 1 1 1 1 1 1 1 1 1 and a graph algorithm. In general, it selects similar users
1 1 1 1 1 1 1 1 1 1

between the target user and other users for common rating 1 1 1 1 1 1 1 1 1 1 greater than a given threshold or high rank users[8][10].
1 1 1 1 1 1 1 1 1

items. The similarity is used as a weight for predicting the

1 1 1 1 1 1 1 1 1 1 1 C) Prediction of Preference Score
1 1 1 1 1

preference score. Various similarity metrics have been
proposed in previous studies. These are as follows [8][10][17]: 1
Tanimoto coefficient. It is similarity between two sets. It is a
ratio of intersections. Assume that set X is {B,C, D} and set Y
1 1
is {C, D, E}. The Tanimoto coefficient T of two set A and B is
1 1 1 1
0.5. This metric doesn‟t consider the user rating but the case of
1 1
a very sparse data set is efficient[8].
1 1
1 1 1 1 1 1 1 The last step in memory-based CF is to predict the preference
1 1 1 1 1 1 1 1 score of the target user for non-rating items. It predicts the
1 1 1 1 1 1 1 1 1 1 1 preference score of non-rating items for the target user, based
1 1 1 1 1 1 1 1 1 1 1 on the rating of nearest neighborhoods. Various methods have
1 1 1 1 1 1 1 1 1 1 1 been proposed, and Weighted Mean is used as most general
1 1 1 1 1 1 1 1 1 1 algorithm. PSU1,Ii is the predicted score of item i for U1 , and
1 1 1 1 1 NNUi is the nearest neighbor i[8].

D) Performance Evaluation
1 1 1

Fig 1 Block Diagram of Memory Based RS [17]

1 1 1 1 1 1 1 1 1 In the CF system, there are two types of measure for the
1 1 1 1 1 1 1 1 1 1 1 1

The prediction process in memory-based CF contains three

1 1 1 1 1 1 1 1 performance evaluation. The first type is prediction accuracy, 1 1 1 1 1 1 1 1

steps. They are similarity evaluation, generation of nearest

1 1 1 1 1 1 1 1 which is evaluated by MAE. Pi is the real preference score of
1 1 1 1 1 1 1 1 1 1 1 1

neighborhoods and score prediction. For evaluation of the 1 1 1 1 1 1 1 1 item i and qi is the predicted score of item i[8].
1 1 1 1 1 1 1 1 1 1

performance, the CF system considers the mean absolute error 1 1 1 1 1 1 1 1 1

(MAE), precision and recall. The CF performance varies 1 1 1 1 1 1 1 1

Merits and Demerits of Memory Based Approach 1 1 1 1 1 1 1

according to the processing method of each step[17]. 1 1 1 1 1 1 1 1

User-based techniques correlate users by mining their (similar) 1 1 1 1 1 1 1 1

Cosine similarity. The Cosine similarity is known as the

1 1 1 1 1 1 1 1 1
ratings and then recommend new items that were preferred by
1 1 1 1 1 1 1 1 1 1

Vector similarity or Cosine coefficient. This metric assumes

1 1 1 1 1 1 1 1
similar users. Item-based techniques correlate the items by 1 1 1 1 1 1 1 1

that common rating items of two users are two points in a

1 1 1 1 1 1 1 1 1 1 1 1
mining (similar) ratings and then recommend new, similar 1 1 1 1 1 1 1 1

vector space model, and then calculates cosӨ between the two
1 1 1 1 1 1 1 1 1 1
items. The main advantages of both techniques are that they use
1 1 1 1 1 1 1 1 1 1

points[10][8][18]. 1
1information that is provided bottom-up by user ratings, that 1 1 1 1 1 1 1 1 1

Person’s Correlation. In Equation, SU1 is the standard 1 1 1 1 1 1 1 1

they are domain-independent and require no content analysis
1 1 1 1 1 1 1 1

deviation of user U1. The Pearson Correlation measures the 1 1 1 1 1 1 1 1 1

and that the quality of the recommendation increases over time.
1 1 1 1 1 1 1 1 1

strength of the linear relationship between two variables. It is

1 1 1 1 1 1 1 1 1 1
1CF techniques are limited by a number of disadvantages. First
1 1 1 1 1 1 1 1 1 1

usually signified by r, and has values in the range [-1.0,1.0].

1 1 1 1 1 1 1 1 1 1 1
of all, the so-called „cold start‟ problem is due to the fact that
1 1 1 1 1 1 1 1 1 1 1 1 1

Where -1.0 is a perfect negative correlation, 0.0 is no

1 1 1 1 1 1 1 1 1 1
CF techniques depend on sufficient user performance from the
1 1 1 1 1 1 1 1 1

correlation, and 1.0 is a perfect positive 1 1 1 1 1 1 1

past. Even when such systems have been running for a while,
1 1 1 1 1 1 1 1 1 1 1

correlation[4][10][8][18]. 1
this problem emerges when new users or items are added. New
1 1 1 1 1 1 1 1 1 1 1

Spearman’s Rank Correlation. The Spearman Rank 1 1 1 1 1 1

users first have to give a sufficient number of ratings for items
1 1 1 1 1 1 1 1 1 1 1 1

Correlation also measures the strength of the linear relationship 1 1 1 1 1 1 1 1

in order to get accurate recommendations based on user-based
1 1 1 1 1 1 1 1 1

1between two variables. Unlike the Pearson Correlation, this 1 1 1 1 1 1 1 1

CF (new user problem)[9]. New items have to be rated by a
1 1 1 1 1 1 1 1 1 1 1 1

metric considers rank of scores. So this similarity measure has

1 1 1 1 1 1 1 1 1 1
sufficient number of users if they are to be recommended. 1 1 1 1 1 1 1 1 1 1

more general applicability than the Pearson Correlation, which

1 1 1 1 1 1 1 1
Another disadvantage for CF techniques is the sparsity of the 1 1 1 1 1 1 1 1 1 1

isn’t suitable outside a normalized preference range. Because

1 1 1 1 1 1 1 1
past user actions in a network. Since these techniques deal with
1 1 1 1 1 1 1 1 1 1 1

community-driven information, they support well-liked tastes 1 1 1 1 1 1

INTERNATIONAL JOURNAL OF RESEARCH IN ELECTRONICS AND COMPUTER ENGINEERING

A UNIT OF I2OR 2027 | P a g e
 IJRECE VOL. 7 ISSUE 2 (APRIL- JUNE 2019) ISSN: 2393-9028 (PRINT) | ISSN: 2348-2281 (ONLINE)
more strongly than unpopular tastes. The learners with an
1 1 1 1 1 1 1 1 1 collaborative filtering3 because they compare the user to a 1 1 1 1 1 1 1 1 1

unusual taste may get less qualitative recommendations, and

1 1 1 1 1 1 1 1 controlled number of segments rather than the entire customer
1 1 1 1 1 1 1 1 1

learners with common taste are unlikely to get unpopular items

1 1 1 1 1 1 1 1 1 1 base. The complex and expensive clustering computation is run
1 1 1 1 1 1 1 1

of high quality recommended. Another common problem is

1 1 1 1 1 1 1 1 1offline. However, recommendation quality is low.1 Cluster
1 1 1 1 1 1 1

scalability. RSs which deal with large amounts of data, like 1 1 1 1 1 1 1 1 1 1 models group numerous customers together in a segment,
1 1 1 1 1 1 1 1

amazon.com, have to be able to provide recommendations in 1 1 1 1 1 1 1 1 1 match a user to a segment, and then consider all customers in
1 1 1 1 1 1 1 1 1 1 1 1

real time, with the number of both the users and items
1 1 1 1 1 1 1 1 1 1 1 the segment similar customers for the purpose of making
1 1 1 1 1 1 1 1 1

exceeding millions[10]. 1 1 recommendations. Because the similar customers that the 1 1 1 1 1 1 1

2.2.2 Model Based Approach1 1 1 1 cluster models find are not the most similar customers, the
1 1 1 1 1 1 1 1 1 1

K-MEANS CF: k-means clustering is applied to identify the 1 1 1 1 1 1 1 1 1 recommendations they produce are less relevant[2]. 1 1 1 1 1

segments. k- means is a clustering method that has found wide 1 1 1 1 1 1 1 1 1 1 1

application in data mining, statistics and machine learning. The 1 1 1 1 1 1 1 1 1

III. LITERATURE REVIEW 1 1

input to k-means is the pair-wise distance between the items to

1 1 1 1 1 1 1 1 1 1 1

Bonnin and Jannach in [7] review and compare the mostly

1 1 1 1 1 1 1 1 1 1

be clustered, where the distance means the dissimilarity of the

1 1 1 1 1 1 1 1 1 1

used methods for automatic playlist generations and

1 1 1 1 1 1 1

items. The number of clusters, k is also an input parameter. It is

1 1 1 1 1 1 1 1 1 1 1 1 1

recommendations. As important issues in music playlists are 1 1 1 1 1 1 1 1

an iterative algorithm and starts with a random partitioning of

1 1 1 1 1 1 1 1 1 1

the co-occurrence of songs and the smooth transition among

1 1 1 1 1 1 1 1 1

the items into k clusters. Each iteration, the centroids of the

1 1 1 1 1 1 1 1 1 1 1

them, Markov models using songs as states and association

1 1 1 1 1 1 1 1 1

clusters is computed and each item is reassigned to the cluster

1 1 1 1 1 1 1 1 1 1 1

rules or sequential patters mining are among the techniques

1 1 1 1 1 1 1 1 1

whose centroid is closest. 1 1 1

used in this domain. The usual recommendation approaches

1 1 1 1 1 1 1 1

CLUSTER MODEL: To find customers who are similar to 1 1 1 1 1 1 1 1 1

treating playlists as users and comparing them through cosine
1 1 1 1 1 1 1 1 1

the user, cluster models divide the customer base into many
1 1 1 1 1 1 1 1 1 1
similarity measures or combined with rank prediction 1 1 1 1 1 1 1

segments and treat the task as a classification problem. The 1 1 1 1 1 1 1 1 1 1

algorithms or content based approaches to find tracks with 1 1 1 1 1 1 1 1 1

algorithms goal is to assign the user to the segment containing 1 1 1 1 1 1 1 1 1 1 1

similar musical features, can also be used. The major
1 1 1 1 1 1 1 1 1

the most similar customers. To find customers who are similar

1 1 1 1 1 1 1 1 1 1
limitations of these methods are that they are computationally
1 1 1 1 1 1 1 1 1

to the user, cluster models divide the customer base into many
1 1 1 1 1 1 1 1 1 1 1
expensive and sometimes work based on strong assumptions
1 1 1 1 1 1 1 1

segments and treat the task as a classification problem[2]. 1 1 1 1 1 1 1 1 1

while their overall effectiveness depends heavily on the type of
1 1 1 1 1 1 1 1 1 1

The algorithm’s goal is to assign the user to the segment

1 1 1 1 1 1 1 1 1 1 1
the used data. Finally, due to the long tail distribution present in
1 1 1 1 1 1 1 1 1 1 1

containing the most similar customers. It then uses the 1 1 1 1 1 1 1 1 1

1the music domain, the authors propose two popularity-based
1 1 1 1 1 1 1 1

purchases and ratings of the customers in the segment to 1 1 1 1 1 1 1 1 1 1

recommendations approaches that also include some artist 1 1 1 1 1 1 1

generate recommendations. The segments typically are created

1 1 1 1 1 1 1
information. 1

using a clustering or other unsupervised learning algorithm,

1 1 1 1 1 1 1 1

although some applications use manually determined 1 1 1 1 1 1

Hyeong-Joon Kwon, Tae-Hoon Lee, etal.[8] base their 1 1 1

segments. Using a similarity metric, a clustering algorithm 1 1 1 1 1 1 1 1

reasoning on the implicit likeliness that can be found in the
groups the most similar customers together to form clusters or
1 1 1 1 1 1 1 1 1 1
playlists generated by professionals, like music radio stations
segments. Because optimal clustering over large data sets is 1 1 1 1 1 1 1 1 1
and Djs. Usually, the items placed into sets together with
impractical, most applications use various forms of greedy 1 1 1 1 1 1 1 1
higher frequency have some common characteristics, like a
cluster generation. They then repeatedly match customers to
1 1 1 1 1 1 1 1
particular genre, pairwise suitability, relative popularity etc.
the existing segments, usually with some provision for creating
1 1 1 1 1 1 1 1 1
They model the transmissions between musical items based on
new or merging existing segments. Once the algorithm
1 1 1 1 1 1 1 1
a graph representation, where adjacent songs are represented
generates the segments, it computes the user’s similarity to 1 1 1 1 1 1 1 1 1
as nodes and each arc has weight equal to the number of times
vectors that summarize each segment, then chooses the
1 1 1 1 1 1 1 1
that this transition was observed. The resulting graph is
segment with the strongest similarity and classifies the user
1 1 1 1 1 1 1 1 1
transformed into a Markov random field where a playlist can
accordingly. Some algorithms classify users into multiple 1 1 1 1 1 1 1
be generated as a random walk starting from a given song
segments and describe the strength of each relationship. Cluster1 1 1 1 1 1 1 1
(node) and using the Markov transition probabilities.
1models have better online scalability and performance than
1 1 1 1 1 1 1 1
Baccigalupo and Plaza in [9], present an interesting Case- 1 1 1 1 1 1 1 1 1

Based Reasoning approach to music playlists’ recommendation

1 1 1 1 1 1

INTERNATIONAL JOURNAL OF RESEARCH IN ELECTRONICS AND COMPUTER ENGINEERING

A UNIT OF I2OR 2028 | P a g e
 IJRECE VOL. 7 ISSUE 2 (APRIL- JUNE 2019) ISSN: 2393-9028 (PRINT) | ISSN: 2348-2281 (ONLINE)
1with aim to generate playlists of a desired length, being both
1 1 1 1 1 1 1 1 1 1 1 Collaborative filtering recommendation technology is used to 1 1 1 1 1 1 1

varied and coherent. Every playlist is treated as a case and their

1 1 1 1 1 1 1 1 1 1 1 1 explore the user s new interest points via user rating data for
1 1 1 1 1 1 1 1 1 1 1 1

relevance is computed based on their songs co-occurrences and 1 1 1 1 1 1 1 1 1 music. In order to fully mine the user s preferences, it often
1 1 1 1 1 1 1 1 1 1 1 1

a recommendation is formed as a combination of the items in

1 1 1 1 1 1 1 1 1 1 1 combined with music tags to study. 1 1 1 1 1

the most similar lists. The authors also analyze the properties
1 1 1 1 1 1 1 1 1 1

that may bias the effectiveness of their approach, namely

1 1 1 1 1 1 1 1 1 IV.METHODOLOGY
songs’ popularity and sub-lists’ length. As the context within
1 1 1 1 1 1 1 1 1

which a music item is consumed or a playlist is generated is of

1 1 1 1 1 1 1 1 1 1 1 1 1
Hybrid approach- 1

high importance, 1 1

Popularity based + Content based + Collaborative based 1 1 1 1 1 1 1

Domingues et al. [10] present an interesting approach of 1 1 1 1 1 1 1 1 1

incorporating the contextual parameters within the 1 1 1 1 1 1

• It has been taken into account minimum rating while 1 1 1 1 1 1 1

recommendation model. More than performing a pre or post 1 1 1 1 1 1 1 1 1

fetching our data i.e. popularity based. 1 1 1 1 1

filtering based on the actual context, they represent it as “virtual

1 1 1 1 1 1 1 1 1 1

1items”. Their results show that these additional dimensions are

1 1 1 1 1 1 1 1 1 • After that, took the loss function called WARP 1 , 1 1 1 1 1 1 1

able of improving the recommendations’ accuracy when

1 1 1 1 1 1 1 (Weighted Approximate Rank Pairwise) 1 1 1

combined with the usual recommendation algorithms. In 1 1 1 1 1 1 1

addition this contextual modeling may also enable the access to

1 1 1 1 1 1 1 1 1 1
• Warp helps us to create recommendations for each 1 1 1 1 1 1 1 1

less popular or novel but however relevant to the active user,

1 1 1 1 1 1 1 1 1 1 1
user by looking at the existing user rating pairs and
1 1 1 1 1 1 1 1 1 1

items, Finally a slightly different approach to music

1 1 1 1 1 1 1 1
predicting rankings for each. 1 1 1 1

recommendations is presented by Rosa et al. [11]. The authors

•
1 1 1 1 1 1 1 1 1 1

It uses the gradient descend algorithm to find the

associate music songs with users’ sentiments as these can
1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1 1

weights that improve our prediction over time. This

extracted from the users’ posts in their social networks by using
1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1 1 1

takes into account both the users past rating history i.e.
1 1 1 1 1 1 1 1 1

1lexicon-based sentiment metrics. However, these last papers 1 1 1 1 1 1 1

1 content based and similar users rating i.e. 1 1 1 1 1 1 1

focus on song recommendations and would possibly need to be

1 1 1 1 1 1 1 1 1 1

collaborative based. 1

reformulated or extended in order to be used for playlist 1 1 1 1 1 1 1 1 1 1

recommendations. 1

Approach – 1 1 1

Personalized recommendation is a typical way of personalized 1 1 1 1 1 1 1 1

service, which actively to push targeted resources to users

1 1 1 1 1 1 1 1 1 Popularity based- 1

according to the user s preferences and the user s evaluation or 1 1 1 1 1 1 1 1 1 1 1 1

feedback on the project, so as to achieve the purpose of 1 1 1 1 1 1 1 1 1 1 1

It creates an instance of popularity based recommender class
1 1 1 1 1 1 1 1 1

decision support and information services. For music 1 1 1 1 1 1 1

and feed it with our training data. This achieves the following
1 1 1 1 1 1 1 1 1 1 1

personalized recommendation, the commonly used methods 1 1 1 1 1 1

goal: based on the popularity of each song, create a
1 1 1 1 1 1 1 1 1 1

include content-based recommendation technology, the 1 1 1 1 1

recommender that accept a user_id as input and out a list of 1 1 1 1 1 1 1 1 1 1 1 1

collaborative filtering recommendation technology and hybrid 1 1 1 1 1 1

recommended song of that user 1 1 1 1

recommendation technology. The content-based 1 1 1 1

The logic can be seen more clearly here - Based on the number
recommendation is concerned with some of the characteristics
1 1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1

of users or guests that rated place1 and place 2, we’d say that
1 1 1 1 1 1 1 1 1 1 1 1 1

of the music itself, which mainly use the metadata related with
1 1 1 1 1 1 1 1 1 1 1

place 1 is more popular than place 2, so based on popularity,

1 1 1 1 1 1 1 1 1 1 1 1

the user s favorite music to match the information. First of all,

1 1 1 1 1 1 1 1 1 1 1 1

place 1 would be recommended over place 2

1 1 1 1 1 1 1

obtain the metadata of user s favorite music through the user s

1 1 1 1 1 1 1 1 1 1 1 1

historical records. Then the music with similar content is 1 1 1 1 1 1 1 1 1

Approach – 2 1 1

obtained by calculating the similarity between the metadata and1 1 1 1 1 1 1 1

1recommend to the user. By comparing the acoustic features 1 1 1 1 1 1 1 1 1

Content Based- 1

that are extracted from the music with the user s preferences,
1 1 1 1 1 1 1 1 1 1 1

music that has similar acoustic features is recommendation to

1 1 1 1 1 1 1 1 1
Content based systems predict what you like based on what you
1 1 1 1 1 1 1 1 1 1

the user [12].

1 1 1
have liked in the past.
1 1 1 1 1

INTERNATIONAL JOURNAL OF RESEARCH IN ELECTRONICS AND COMPUTER ENGINEERING

A UNIT OF I2OR 2029 | P a g e
 IJRECE VOL. 7 ISSUE 2 (APRIL- JUNE 2019) ISSN: 2393-9028 (PRINT) | ISSN: 2348-2281 (ONLINE)
Starting from an explicit set of music tracks provided by the
1 1 1 1 1 1 1 1 1 1 1 (forward-selection, backward-elimination, tree-induction, etc.). 1 1 1

user as evidence of his/her music preferences, it is infered high-

1 1 1 1 1 1 1 1 1 1Wrapper methods are independent of the prediction 1 1 1 1 1 1 1

level semantic descriptors, covering different musical facets,

1 1 1 1 1 1 1 algorithm[13]. 1

such as genre, culture, moods, instruments, rhythm, and tempo.

1 1 1 1 1 1 1 1 1 2) Filter methods are typically based on heuristic measures,
1 1 1 1 1 1 1 1 1

On this basis, two of the proposed approaches employ a

1 1 1 1 1 1 1 1 1 1 such as Mutual Information or Pearson Correlation, to score
1 1 1 1 1 1 1 1 1

semantic music similarity measure to generate 1 1 1 1 1 1 features based on their information contents with respect to the
1 1 1 1 1 1 1 1 1 1

recommendations. The third approach creates a probabilistic 1 1 1 1 1 1 1 prediction task. Similar to wrapper methods, filter methods are 1 1 1 1 1 1 1 1 1

model of the user's preference in the semantic domain.

1 1 1 1 1 1 1 1 1 also independent of the algorithm in use. However, they do not
1 1 1 1 1 1 1 1 1 1 1

require training many models and therefore scale well for large
1 1 1 1 1 1 1 1 1 1

Approach – 3 1 1
datasets. Yet, filter methods cannot be naturally extended to 1 1 1 1 1 1 1 1 1

recommender systems, in which the prediction target varies and 1 1 1 1 1 1 1 1

Collaborative filtering - 1 1

1depends both on the user's history and on the item under 1 1 1 1 1 1 1 1 1 1 1

Collaborative systems predict what you like based on what 1 1 1 1 1 1 1 1 1

consideration. This work proposes a framework and algorithms 1 1 1 1 1 1 1 1

other similar users have liked in the past.

1 1 1 1 1 1 1
to address the above difficulties[14].
1 1 1 1 1

3) Embedded methods are a family of algorithms in which the

1 1 1 1 1 1 1 1 1 1 1

It is a method of making automatic predictions (filtering) about

1 1 1 1 1 1 1 1 1 1 feature selection is performed in the course of the training
1 1 1 1 1 1 1 1 1 1

the interests of a user by collecting preferences or taste

1 1 1 1 1 1 1 1 1 1 phase. Unlike wrapper methods, they are not based on cross-
1 1 1 1 1 1 1 1 1

information from many users (collaborating). The underlying 1 1 1 1 1 1 1 validation and therefore scale with the size of the data. 1 1 1 1 1 1 1 1 1 1

assumption of the collaborative filtering approach is that if a 1 1 1 1 1 1 1 1 1 1 However, since the feature selection is an inherent property of 1 1 1 1 1 1 1 1 1 1

person A has the same opinion as a person B on an issue, A is

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 the algorithm, an embedded method is tightly coupled with the
1 1 1 1 1 1 1 1 1 1

more likely to have B's opinion on a different issue than that of

1 1 1 1 1 1 1 1 1 1 1 1 1 specific model: If the recommendation algorithm is replaced,
1 1 1 1 1 1 1 1

a randomly chosen person.

1 1 1 features selection needs to be revisited[15]. 1 1 1 1 1 1

Collaborative Filtering is the process of filtering for 1 1 1 1 1 1 1 1

Techniques of Content Based Approach 1 1 1 1 1

information or patterns using techniques involving 1 1 1 1 1 1

TF-IDF : The terms that occur frequently in one document 1 1 1 1 1 1 1 1 1 1

collaboration among multiple agents, viewpoints, data sources, 1 1 1 1 1 1 1

(TF=term-frequency), but rarely in the rest of the corpus (IDF = 1 1 1 1 1 1 1 1 1 1

etc. Applications of collaborative filtering typically involve

1 1 1 1 1 1 1
1inverse-document-frequency), are more likely to be relevant to 1 1 1 1 1 1 1 1

very large data sets. 1 1 1 1

the topic of the document. In addition, normalizing the

1 1 1 1 1 1 1 1 1

resulting weight vectors prevent longer documents from having 1 1 1 1 1 1 1

V. CONTENT BASED APPROACH 1 1 1

1a better chance of retrieval.

1 1 1 1

Any Systems implementing a content-based recommendation

1 1 1 1 1 1

approach analyze a set of documents and/or descriptions of 1 1 1 1 1 1 1 1 1

NAÏVE BAYES: Naıve Bayes is a probabilistic approach to

1 1 1 1 1 1 1 1 1

items previously rated by a user, and build a model or profile of

1 1 1 1 1 1 1 1 1 1 1 1 1

inductive learning, and belongs to the general class of Bayesian 1 1 1 1 1 1 1 1 1

user interests based on the features of the objects rated by that

1 1 1 1 1 1 1 1 1 1 1 1
1classifiers. These approaches generate a probabilistic model 1 1 1 1 1 1 1

user. The recommendation process basically consists in

1 1 1 1 1 1 1
based on previously observed data.
1 1 1 1

matching up the attributes of the user profile against the 1 1 1 1 1 1 1 1 1 1

attributes of a content object. The result is a relevance judgment 1 1 1 1 1 1 1 1 1 1 Merits and Demerits of Content Based approach1 1 1 1 1 1 1

1that represents the user’s level of interest in that object. If a

1 1 1 1 1 1 1 1 1 1 1 1 The approval of the content-based recommendation paradigm
1 1 1 1 1 1 1

profile correctly reflects user preferences, it is of tremendous

1 1 1 1 1 1 1 1 1 has several advantages:
1 1 1

advantage for the effectiveness of an information access 1 1 1 1 1 1 1 1 USER INDEPENDENCE - Content-based recommenders
1 1 1 1 1

process[11]. 1 exploit solely ratings provided by the active user to build her
1 1 1 1 1 1 1 1 1 1 1

Methods for Content Based Feature Selection[12] 1 1 1 1 1 1 own profile. Instead, collaborative filtering methods need
1 1 1 1 1 1 1

1) Wrapper methods evaluate different subsets of features by

1 1 1 1 1 1 1 1 1 ratings from other users in order to find the “nearest neighbors”
1 1 1 1 1 1 1 1 1 1 1

training a model for each subset and then evaluating each 1 1 1 1 1 1 1 1 1 1 of the active user[11].
1 1 1 1

subset's contribution on a validation dataset. As the number of

1 1 1 1 1 1 1 1 1 1 TRANSPARENCY - Explanations on how the recommender 1 1 1 1 1 1 1

all possible subsets is factorial in the number of features,

1 1 1 1 1 1 1 1 1 1 system works can be provided by explicitly listing content
1 1 1 1 1 1 1 1 1

different heuristics are used to choose “promising" subsets 1 1 1 1 1 1 1 1 features or descriptions that caused an item to occur in the list
1 1 1 1 1 1 1 1 1 1 1 1

INTERNATIONAL JOURNAL OF RESEARCH IN ELECTRONICS AND COMPUTER ENGINEERING

A UNIT OF I2OR 2030 | P a g e
 IJRECE VOL. 7 ISSUE 2 (APRIL- JUNE 2019) ISSN: 2393-9028 (PRINT) | ISSN: 2348-2281 (ONLINE)
of recommendations. Those features are indicators to consult in
1 1 1 1 1 1 1 1 scheme; however, historical data is needed for this 1 1 1 1 1 1 1 1

1order to decide whether to trust a recommendation[11].

1 1 1 1 1 1 1 1 approach[13]. 1

NEW ITEM - Content-based recommenders are capable of

1 1 1 1 1 1 1 1 Mixed Hybrid. In this approach, recommendations for each
1 1 1 1 1 1 1 1

recommending items not yet rated by any user. As a 1 1 1 1 1 1 1 1 1 1 source are ranked, and then the top-n are picked from each
1 1 1 1 1 1 1 1 1 1 1

consequence, they do not suffer from the first-rater problem, 1 1 1 1 1 1 1 1 1 source, one recommendation at a time by alternating the
1 1 1 1 1 1 1 1 1

which affects collaborative recommenders which rely solely on

1 1 1 1 1 1 1 sources. This approach only considers relative position in a
1 1 1 1 1 1 1 1 1

1users‟ preferences to make recommendations. Therefore, until

1 1 1 1 1 1 1 ranked list and does not include individual recommendation
1 1 1 1 1 1 1 1

the new item is rated by a substantial number of users, the

1 1 1 1 1 1 1 1 1 1 1 1 scores. In cases where a recommendation is produced by
1 1 1 1 1 1 1 1 1

system would not be able to recommend it[11].

1 1 1 1 1 1 1 1 multiple context sources (i.e. was previously picked from 1 1 1 1 1 1 1 1

content-based systems have several shortcomings: 1 1 1 1 1 another source) the algorithm simply selects the next 1 1 1 1 1 1 1 1

LIMITED CONTENT ANALYSIS - Content-based 1 1 1 1 1 recommendation from the ranked list for that source[13]. 1 1 1 1 1 1 1 1

techniques have a natural limit in the number and type of

1 1 1 1 1 1 1 1 1 1 1 Cross-Source Hybrid. This approach strongly favors 1 1 1 1 1 1

features that are associated, whether automatically or manually,

1 1 1 1 1 1 1 recommendations that appear in more than one source. Is is 1 1 1 1 1 1 1 1 1 1

1with the objects they recommend.

1 1 1 1 1 believed that if a recommendation is generated from more than
1 1 1 1 1 1 1 1 1 1

OVER-SPECIALIZATION - Content-based recommenders 1 1 one context source / algorithm, i.e. by both collaborative
1 1 1 1 1 1 1 1 1

have no inherent method for finding something unexpected. Filtering (Facebook) and content-based recommendation 1 1 1 1 1

The system suggests items whose scores are high when (Wikipedia), then it should be considered more important. To 1 1 1 1 1 1 1 1 1

matched against the user profile; hence the user is going to be compute a final recommendation set, the weighted hybrid 1 1 1 1 1 1 1 1

recommended items similar to those already rated. This approach is first applied, then each recommendation's weight is 1 1 1 1 1 1 1 1 1

drawback is also called serendipity problem to highlight the multiplied by the number of sources in which it appeared. The 1 1 1 1 1 1 1 1 1 1 1

tendency of the content-based systems to produce following equation describes the the cross-source hybrid 1 1 1 1 1 1 1

recommendations with a limited degree of novelty. approach: 1

NEW USER - Enough ratings have to be collected before a Wreci =Σsj2S(Wreci;sj _Wsj ) *| Sreci |
1 1 1 1 1 1 1

content-based recommender system can really understand user where jSreci j is the number of context sources
1 1 1 1 1 1 1 1 1

preferences and provide accurate recommendations. recommendation i was generated by (i.e. 1, 2, or 3)[13]. 1 1 1 1 1 1 1 1 1 1

Therefore, when few ratings are available, as for a new user, How Hybrid Approach Works:
1 1 1

the system will not be able to provide reliable In a Movie Recommender system, the content based part of the
1 1 1 1 1 1 1 1 1 1 1

recommendations[11]. movie recommender is based on a naive Bayesian text

1 1 1 1 1 1 1 1 1

classification method. The classifier creates a naive Bayesian 1 1 1 1 1 1 1 1

VI. HYBRID APPROACH 1 1

model for every user, based on the content of the movies the
1 1 1 1 1 1 1 1 1 1 1 1

Traditional recommender system techniques such as 1 1 1 1 1 1

user has rated. 1 1

collaborative filtering (CF), content-based, and knowledge- 1 1 1 1 1

based filtering, each have unique strengths and limitations. For

1 1 1 1 1 1 1 1 1 Issue with Hybrid Approach
1 1 1 1

example, CF suffers from sparsity and cold start problems,

1 1 1 1 1 1 1 1 1 Reliable Integration: The first problem is to reflect the 1 1 1 1 1 1 1 1 1

while content-based approaches suffer from narrowness and

1 1 1 1 1 1 1 collaborative and content-based data when making 1 1 1 1 1 1

require descriptions. However, a hybrid approach can use one

1 1 1 1 1 1 1 1 1 recommendations. An easy solution is to use collaborative and 1 1 1 1 1 1 1 1 1

approach to make predictions where the other fails, resulting in

1 1 1 1 1 1 1 1 1 1 content-based methods in parallel or in cascade. However, such 1 1 1 1 1 1 1 1

a more robust recommender System[1][13].

1 1 1 1 1an approach has drawbacks. Although Meta recommender
1 1 1 1 1 1 1

systems have been proposed to select a recommender system

1 1 1 1 1 1 1 1 1

Types of Hybrid 1 1 1
among conventional ones on the basis of certain quality
1 1 1 1 1 1 1 1 1

Weighted Hybrid. In this approach, a score for each 1 1 1 1 1 1 1 1 1

measures the disadvantages of the selected system are 1 1 1 1 1 1 1 1

recommended item is simply the weighted sum of the 1 1 1 1 1 1 1 1 1

inherited. Moreover, the heuristics-based integration dealt with 1 1 1 1 1 1 1

recommendation scores for each source. Weights for each 1 1 1 1 1 1 1 1

in other studies lacks a principled justification[5].
1 1 1 1 1 1 1

context source are user-configurable through interactive

1 1 1 1 1 1
Efficient Calculation: The second problem, which has been 1 1 1 1 1 1 1 1

sliders. Automatically optimizing the set of weights for each

1 1 1 1 1 1 1 1 1
scarcely dealt with, is to efficiently adapt a recommender
1 1 1 1 1 1 1 1 1

context source is desirable, but not trivial. Empirical

1 1 1 1 1 1 1 1
system according to the increase in rating scores and users. An
1 1 1 1 1 1 1 1 1 1 1

bootstrapping can be used to calculate an optimal weighting 1 1 1 1 1 1 1 1 1

INTERNATIONAL JOURNAL OF RESEARCH IN ELECTRONICS AND COMPUTER ENGINEERING

A UNIT OF I2OR 2031 | P a g e
 IJRECE VOL. 7 ISSUE 2 (APRIL- JUNE 2019) ISSN: 2393-9028 (PRINT) | ISSN: 2348-2281 (ONLINE)
easy solution is to take a memory-based approach, which is
1 1 1 1 1 1 1 1 1 1 Sample recommendation
originally free from this problem because the whole data is
1 1 1 1 1 1 1 1 1 1

always used to make recommendations. However, these results

1 1 1 1 1 1 1 1
We 1 will 1 call 1 it 1 at 1 the 1 end 1 using 1 the 1 model 1
in the late responses tried to overcome this disadvantage by
1 1 1 1 1 1 1 1 1 1
data 1 and 1 a 1 list 1 of 1 3 1 or 1 more 1 random 1 user 1 ids
using a probabilistic method in a pure collaborative filtering
1 1 1 1 1 1 1 1 1
1 as 1 the 1 parameters.
context. On the other hand, proposed an efficient method that
1 1 1 1 1 1 1 1 1 1

Each 1 user 1 has 1 movies 1 that 1 they 1 listed 1 as 1 well 1

incrementally trains an aspect model used for model-based 1 1 1 1 1 1 1 1

as 1 movies 1 that 1 our 1 system 1 has 1 recommended 1 for

collaborative filtering. To our knowledge, there are no studies
1 1 1 1 1 1 1 1 1

1 them.
on incremental adaptation of hybrid recommender systems. It
1 1 1 1 1 1 1 1

need to carefully design hybrid architecture while considering

1 1 1 1 1 1 1 1
So 1 basically 1 recommendation 1 algorithms 1 help 1 us 1
whether the previous prominent methods can be applied or
1 1 1 1 1 1 1 1 1
make 1 decisions 1 by 1 learning 1 our 1 preferences.
not[5]. 1

RESULT 1 AND 1 DISCUSSION After entering the final step, we will get 2 results-
 It will print the top three known positive movies that
Popularity Based- the user has picked.
We 1 create 1 an 1 instance 1 of 1 popularity 1 based  It will print the top three recommended movies that
recommender 1 class 1 and 1 feed 1 it 1 with 1 our 1 training 1 our model predicts which is required.
data. 1 The 1 code 1 below 1 achieves 1 the following 1 goal:
1 based 1 on 1 the 1 popularity 1 of 1 each 1 song, 1 create 1 a
1 recommender 1 that 1 accept 1 a 1 user_id 1 as 1 input 1 and
1 out 1 a 1 list 1 of 1 recommended 1 song 1 of 1 that 1 user
pm 1 = 1 Recommenders.popularity_recommender_py()
pm.create(train_data, 1 'user_id', 1 'song')
#user 1 for 1 the 1 popularity 1 model 1 to 1 make 1 some 1
prediction
user_id 1 = 1 users[5]
pm.recommend(user_id)
VII. CONCLUSION 1

Table 5.1: Popularity Based Output

Several recommendation systems have been anticipated are
1 1 1 1 1 1 1

based on collaborative filtering, content based filtering and

1 1 1 1 1 1 1 1

hybrid recommendation methods and so far most of them have

1 1 1 1 1 1 1 1 1 1

been able to resolve the problems while providing improved

1 1 1 1 1 1 1 1 1

recommendations. However, due to information explosion, it is 1 1 1 1 1 1 1 1

required to work on this research area to explore and provide

1 1 1 1 1 1 1 1 1 1 1

new methods that can provide recommendation in a wide range

1 1 1 1 1 1 1 1 1 1

of applications while considering the quality and privacy

1 1 1 1 1 1 1 1

aspects. Thus, the current recommendation system needs

1 1 1 1 1 1 1

enhancement for present and future requirements of better 1 1 1 1 1 1 1 1

recommendation qualities. 1

References 1

Hybrid system- [1] Arenas-García, J., Meng, A., Petersen, K. B., Lehn-
1 1 1 1 1 1 1 1

Schioler, T., Hansen, L. K., & Larsen, J. (2007, August).

1 1 1 1 1 1 1 1 1 1

Unveiling music structure via plsa similarity fusion. In

1 1 1 1 1 1 1 1

INTERNATIONAL JOURNAL OF RESEARCH IN ELECTRONICS AND COMPUTER ENGINEERING

A UNIT OF I2OR 2032 | P a g e
 IJRECE VOL. 7 ISSUE 2 (APRIL- JUNE 2019) ISSN: 2393-9028 (PRINT) | ISSN: 2348-2281 (ONLINE)
Machine Learning for Signal Processing, 2015 IEEE 1 1 1 1 1 1 1 [15] Meenakshi Sharma and Sandeep Mann, “A Survey of
1 1 1 1 1 1 1 1 1

Workshop on (pp. 419-424). IEEE. 1 1 1 1 1 Recommender Systems: Approaches and Limitations” 1 1 1 1 1

,ICAECE-2013. 1

[2] Greg Linden, Brent Smith, and Jeremy York,”Amazon.com

1 1 1 1 1 1 1 1

[16] Royi Ronen, Noam Koenigstein, Elad Ziklik and Nir

1 1 1 1 1 1 1 1 1

Recommendations Item-to-Item Collaborative Filtering”, IEEE 1 1 1 1

Nice,” Selecting Content-Based Features for Collaborative

1 1 1 1 1 1

Computer Society 2003.

1 1 1 1

Filtering Recommenders”, ACM 2013

1 1 1 1

[3] Jun Wang, Arjen P. de Vries, Marcel J.T. Reinders,”

1 1 1 1 1 1 1 1 1 1

[17] Hendrik Drachsler, Hans G.K. Hummeland Rob Koper,

1 1 1 1 1 1 1 1

Unifying Userbased and Itembased Collaborative Filtering 1 1 1 1 1 1

“Personal recommender systems for learners in lifelong

1 1 1 1 1 1 1

Approaches by Similarity Fusion”, 2006 ACM. 1 1 1 1 1 1

learning networks: the requirements,techniques and model”.

1 1 1 1 1 1

[4] Panagiotis Symeonidis *, Alexandros Nanopoulos,

1 1 1 1 1 1

[18] Mohammad Yahya H. Al-Shamri, Nagi H. Al-Ashwal,”

1 1 1 1 1 1 1 1

Apostolos N. Papadopoulos,Yannis Manolopoulos,” 1 1 1 1

Fuzzy-Weighted Similarity Measures for Memory-Based 1 1 1 1 1

Collaborative recommender systems: Combining effectiveness 1 1 1 1 1

Collaborative Recommender Systems”, Journal of Intelligent

1 1 1 1 1 1

and efficiency”, 2007 Elsevier Ltd.

1 1 1 1 1

Learning Systems and Applications, 2014.

1 1 1 1 1

[5] Kazuyoshi Yoshii, Masataka Goto, Kazunori Komatani,

1 1 1 1 1 1 1

[19] Michael J. Pazzani and Daniel Billsus,” Content-based

1 1 1 1 1 1 1 1

Tetsuya Ogata, Hiroshi G. Okuno,” An Efficient Hybrid Music

1 1 1 1 1 1 1 1 1

Recommendation Systems”. 1 1

Recommender System Using an Incrementally 1 1 1 1 1

[20] Robin van Meteren and Maarten van Someren,” Using

1 1 1 1 1 1 1 1 1

TrainableProbabilistic Generative Model”,IEEE 2008 1 1 1 1

Content-Based Filtering for Recommendation”. 1 1 1 1

[6] Akhmed Umyarov, Alexander Tuzhilin,”Improving

[21] Jaldert Rombouts, Tessa Verhoef,” A simp le hybrid
1 1 1 1 1

1 1 1 1 1 1 1 1 1

Collaborative Filtering Recommendations Using External

movie recommender system”.
1 1 1 1 1

1 1 1

Data”, 2008 IEEE. 1 1 1

[7] Zunping Cheng, Neil Hurley,” Effective Diverse and

1 1 1 1 1 1 1 1

Obfuscated Attacks on Model-based Recommender Systems” 1 1 1 1 1 1

2009 ACM. 1 1

[8] Hyeong-Joon Kwon, Tae-Hoon Lee, and Kwang-Seok

1 1 1 1 1 1 1

Hong,” Improved Memory-based Collaborative Filtering

1 1 1 1 1

Using Entropy-based Similarity Measures”, 2009 ACADEMY

1 1 1 1 1 1

PUBLISHER. 1

[9] Asela Gunawardana, Christopher Meek,” A Unified

1 1 1 1 1 1 1

Approach to Building Hybrid Recommender Systems” 200X 1 1 1 1 1 1 1

ACM 1

[10] Sergio Cleger-Tamayo, Juan M. Fern´andez-Luna, and

1 1 1 1 1 1 1

Juan F. Huete, ”A New Criteria for Selecting Neighborhood in

1 1 1 1 1 1 1 1 1 1

Memory-Based Recommender Systems*”,Springer-Verlag 1 1 1

Berlin Heidelberg 2011. 1 1 1

[11] Pasquale Lops, Marco de Gemmis and Giovanni

1 1 1 1 1 1 1 1

Semeraro,” Content-based Recommender Systems: State of the 1 1 1 1 1 1

Art and Trends”, Springer 2011

1 1 1 1 1 1

[12] Yoav Bergner,Stefan Dr¨oschlery, Gerd Kortemeyer, Saif

1 1 1 1 1 1 1

Rayyan, Daniel Seaton, and David E. Pritchard,” ModelBased

1 1 1 1 1 1 1 1

Collaborative Filtering Analysis of Student Response Data: 1 1 1 1 1 1 1

MachineLearning Item Response Theory”,2012. 1 1 1 1

[13] Svetlin Bostandjiev, John O‟Donovan, Tobias Höllerer,”

1 1 1 1 1 1 1

TasteWeights:A Visual Interactive Hybrid Recommender 1 1 1 1 1

System” 2012 ACM 1 1 1

[14] Jiankai Sun, Shuaiqiang Wang, Byron J. Gao, Jun Ma,”

1 1 1 1 1 1 1 1 1 1

Learning to Rank for Hybrid Recommendation”, 2012 ACM 1 1 1 1 1 1 1 1

INTERNATIONAL JOURNAL OF RESEARCH IN ELECTRONICS AND COMPUTER ENGINEERING

A UNIT OF I2OR 2033 | P a g e