personalization using recommendation ... dept. of cse, shri vishnu engineering college for women,...

International Journal of Computer Engineering and Applications, Volume XI, Issue IX, September 17, www.ijcea.com ISSN 2321-3469

V Lakshmi Chetana*, Dr. SP Syed Ibrahim, Dr. B Rama, Dr. P KiranSree 56

PERSONALIZATION USING RECOMMENDATION SYSTEMS:

STATE-OF-THE ART REVIEW

V Lakshmi Chetana*, Dr. SP Syed Ibrahim, Dr. B Rama, Dr. P KiranSree

Asst. Prof., DVR & Dr.HS MIC College of Technology, Kanchikacherla

Professor, School of ComputingScience and Engineering, VIT University-Chennai Campus,

Chennai

Asst. Prof., Computer Science Department, KakatiyaUniversity, Hanamkonda

Professor, Dept. of CSE, Shri Vishnu Engineering College for Women, Bhimavaram

*Corresponding author:[email protected]

ABSTRACT:

In the era of internet, the proliferation of services will be a great challenge to the users in selecting the best and optimal services that suit the user requirement. So there is a need to filter, prioritize and efficiently deliver relevant information according to the user requirement. Recommendation systems rescue the users by selecting the required content after searching through large volumes of services. Recommendation system predicts user responses to options using diverse prediction techniques like content-based system, collaborative filtering system and hybrid system. This paper, in detail, explores different characteristics, approaches, evaluation metrics and potentials of different prediction techniques in recommendation systems that endorse the quality of services for the users.

Keywords: Content-based Recommendation system, Collaborative Recommendation system,

Hybrid Recommendation system, Evaluating Recommendation system, Evaluation metrics, Similarity measures, Recommendation systems.

http://www.ugc.ac.in/journallist/subjectwisejurnallist.aspx?tid=MjMyMTM0Njk=&&did=U2VhcmNoIGJ5IElTU04=

PERSONALIZATION USING RECOMMENDATION SYSTEMS: STATE-OF-THE ART REVIEW

V Lakshmi Chetana, Dr.SP Syed Ibrahim, Dr.B Rama, Dr.P KiranSree 57

[1] INTRODUCTION

Increase in the growth of available information online has created the problems of information

overload and paradox of choice. Most of the e-commerce sites uses personalization and

prioritization which provides content and services tailored to individuals based on knowledge

about their preferences and behavior. Recommendation systems are the important tools to solve

the challenges of information overload and paradox of choice.

The objective of the recommendation system is to create meaningful predictions to the

users for products or items of their interest. There are multitude ways to find the interests of

the users like tags, reviews, likes and number of visits for a page or items or blogs.It is the

process of finding what the customer needs, either data or an item [1].Some of the example

recommender systems are Ringo, Pandora, Last.fm for music , Netflix,Sceneami.com for

movies , YouTube, IMDb for videos, LinkedIn for jobs, Facebook, Twitter for friends, Amazon

for books, OKCUPID for dating and ACM, Springer, IEEE, ScienceDirect etc for published

papers.

The point that differentiates the recommender system with search engines is,

recommender systems try to present relevant content that the user did not necessarily search

for and which they have not even heard about. The goal of the recommendation system is to

make legitimate recommendations for the user which the user has not seen or to provide users

with item lists ranked in accordance with the estimated preferences. The recommendation

systems are mostly commonly used in two scenarios. First, when there are a very large number

of items available for products from Amazon, Movies from Netflix, Music from Pandora, news

items on Google news etc, it becomes difficult for the user to find what they want. Searching

is helpful when the users know what they are looking for.In this situation, the recommender

system recommends items that the user may not know about or hard-to-find or non-hit items.

The total sale of such items is called long-tail. [Figure-1] represents the long-tail phenomenon.

Figure: 1. Long-tail Phenomenon



Nomenclatures

𝑟𝑥,𝑖 Rating given by user x on item i

𝑟𝑦,𝑖 Rating given by user y on item i.

�̅�𝑥 Average ratings given by user x

�̅�𝑦 Average ratings given by user y

I Set of Set of all items in User-Item space

�̅�𝑖 Average ratings on item i given by all users

�̅�𝑗 Average ratings on item j given by all users

U Set of all users in User-Item space

𝑟𝑢,𝑖 Rating given by user u for item i

𝑟𝑢,𝑗 Rating given by user u for item j

𝑃𝑎,𝑖 Predicted rating for active user a on item i

𝑤𝑎,𝑥 Similarity coefficient between a and x

Abbreviations

CBRS Content-based Recommendation System

CF Collaborative Filtering

CFRS Collaborative Filtering based Recommendation System

SVD Singular Value Decomposition

MF Matrix Factorization

PCA Principle Component Analysis

LSI Latent Semantic Indexing

ALS Alternate Least Squares

SVM Support Vector Machines

MAE Mean Absolute Error

RMSE Rooted Mean Squared Error

NMAE Normalized Mean Absolute Error

ROC Receiver Operating Characteristic

The y-axis represents the popularity i.e., the number of times a product is chosen and the x-

axis represents products ordered according to their popularity. Online sites provide the entire

range of products that are most popular and the rest of the tail.Second, as personal taste plays

an important role in selecting an item, recommendation systems often utilizes the interests of a

group which helps in discovering items based on the similarities.

This paper reviews various techniques used in recommendation systems and is organized

as follows - The second section describes the steps involved in recommender systems like

information gathering, learning and recommendation. Section three explains about various

filtering techniques used to produce recommendations followed by various evaluation metrics

in section four.

[1.1] STEPS IN RECOMMENDER SYSTEMS

The steps that are involved in recommender systems are

1. Information gathering

2. Learning

3. Recommendations



[1.1.1] INFORMATION GATHERING STEP

In this phase, user preferences or interest is collected from different sources like user

profiles,reviews,tags,shares,likes,newsgroup and so on.Reddit.com is a website which collects

votes for an item or newsgroup.Recommender systems depend on different types of input

extracted through.

Explicit feedback :

The system generates an interface for the user to provide ratings or feedback for an item to

improvise itself. The accuracy of the recommender system relies on the quantity of the ratings

given by the user.But the main problem with this method is, the user may not be everytime

ready to give complete information and apt information.

Implicit feedback :

The system implicitly derives the user preferences or interests by monitoring the previous

actions of users likes purchase history, navigation history, a number of visits to a web page,

number of shares and likes in social networking sites, content of the email and so on.[2].It

reduces burden on the user but the problem with it is, it gives less accuracy

Hybrid feedback :

To improve the performance of the system the pros of both explicit and implicit feedback is

combined. Here the system takes explicit feedback only when he is interested in expressing.

[1.1.2] LEARNING STEP

It uses different machine learning algorithms to filter user interests collected in the first step to

predict interests of user's choice. The different learning algorithms that are used are Bayesian,

Decision Tree, Neighborhood-based , Neural Network, Rule learning, Ensemble, Gradient –

Descent based, Kernal methods, Clustering, Associative classification, Lazy learning,

Regularization methods, Topic Independent Scoring Algorithm[3].

[1.1.3] RECOMMENDATION STEP

It predicts what sort of items the user prefers. These predictions are done either by the system

observed activities or by analyzing the data collected in the information collection step. Any

of the learning algorithms may be applied to analyze the input data. [Figure- 2]exhibits the

steps involved in the recommendation process.

Figure: 2. Recommendation Process Steps

Information

Collection

Step

Learning

Step

Recommendation /Prediction

Step

Feedback



[2] APPROACHES FOR RECOMMENDATION SYSTEMS

Recommendation systems should be efficient in predicting the items as per the user's interests

accurately.The different techniques that are used in filtering the recommendations are shown

in [Figure-3]

Figure:3.Approaches for Recommender Systems

[2.1] CONTENT-BASED APPROACH

Content is the foundation for building recommender applications. These recommendations are

based on item description rather than similarity of other users.There are various contents

available, such as wikis, polls, articles, blogs, classification terms, photos and videos. Metadata

can be extracted by analyzing content. The procedure consists of creating tokens, normalizing

the tokens, approving them, stemming them, and detecting phrases. [1].The major idea of this

approach is to recommend items to customer X similar to the user profile or previous items

purchased or liked. [Figure-4] explains how the content-based recommendations are

generated. The recommender system generates the user and item profiles from the web, where

user profile consists of information regarding their ratings, likes etc and the item profile

consists of a complete description of the items like actors, music director, a genre of the movie

etc. The recommender systems take these as inputs and generate recommendations.

The advantages of content-based recommender systems are – information of other users

is not required, ready to predict items to users with unique tastes.Ready to recommend new and

hard to find items. Some of the disadvantages are - finding the appropriate features is hard,

over specialization, cold start problem for new users, never recommends items outside user's

content profile and scalability.

Recommender systems

Content - based filteringCollaborative filtering

Memory based approach

User based Item based

Model based

approach

Hybrid approach[Memory based+Model based]

Hybrid filtering[Content-based+Collaborative filtering]



[2.2] COLLABORATIVE FILTERING APPROACH

The word “Collaborative Filtering” was first coined by one of the developers of Tapestry[4].

In this approach, recommendations are drawn based on the interests of the crowd. The vital

assumption of collaborative filtering is that if two users A and B rate ‘n’ items similarly or have

similar behaviors then they rate or act on various items similarly[4].This technique uses a data

structure called “User-Item Matrix”.It consists of ratings given by the users for different items

where these ratings usually scale from 1 through 5.Table 1 shows a simple movie database

with some rows (called as records) that describes five movies and some columns (which are as

attributes, fields, characteristics or variables ) that describe the properties like Genre of the

movie and its year of release.

Table 1 Movie Database

ID Movie Name Genre Year of

release

1 Terminator Action, Science Fiction 1984

2 Matrix Action, Science Fiction 1999

3 Final Destination Horror, Thriller 2000

4 Lord of the Rings Adventure, Drama, Fantasy 2001

5 Avatar Action,Adventure,Fantasy 2009

User-Item Matrix which consists of ratings given by different users for the movies mentioned

in the movie database. Table 2 depicts the User-Item Matrix for the movie database.

Generates

Web

Documents

User

profiles

Item

Profiles

CBRS Recommendations

Figure: 4. Architecture of Content based Recommendation System



Table 2 User-Item Matrix

Movie Terminator Lord of Rings Avatar Matrix

U1 3 2

U2 5 3 3

U3 3 ? 2

U4 2.5 3.5

U5 4.5 4 3.5

U6 3 2

Collaborative filtering is classified into two categories: Memory-based and Model-based.

[2.2.1] MEMORY-BASED COLLABORATIVE FILTERING

In this technique, complete or part of the user-item matrix is used to generate recommendations.

This technique is also called as Neighborhood- based CF.Memory based CF is further

classified into user-based CF and item-based CF.In user-based CF, the similarity is calculated

between the users by differentiating their ratings on the same item. In the item-based CF, the

similarity is calculated between items rather than users[2]. To generate recommendations, the

neighbors of the active users are found by calculating the similarity measure and a weighted

aggregate of those user’s ratings are computed. [Figure-5] describes the architecture of

memory-based collaborative filtering.

Figure : 5. Architecture for memory based Collaborative filtering



SIMILARITY MEASURES

There are many similarity computation methods like Jaccard coefficient, Kendall's coefficient,

Manhattan distance, Spearman’s correlation, Cosine, Adjusted Cosine, Euclidian distance,

constrained Correlation, Mean Squared Difference etc among which Pearson's correlation

coefficient and cosine –based similarity are most commonly used.

Pearson’s correlation coefficient is a statistical approach which is used to

measure to which extent the two variables are linearly related. For user-based CF, the Pearson’s

correlation coefficient between user x and user y is given as (1)

𝑤𝑥,𝑦 =∑ (𝑟𝑥,𝑖−�̅�𝑥)(𝑟𝑦,𝑖−�̅�𝑦)𝑖∈𝐼

√∑ (𝑟𝑥,𝑖−�̅�𝑥)2

𝑖∈𝐼 √∑ (𝑟𝑦,𝑖−�̅�𝑦)2

𝑖∈𝐼

(1)

where 𝑟𝑥,𝑖 is the rating given by user x on item i . 𝑟𝑦,𝑖is the rating given by user y on item i.

�̅�𝑥is the average ratings given by user x. �̅�𝑦 is the average ratings given by user y and I is the

set of all items in User-Item space.

For item – based CF, the Pearson’s correlation coefficient will be

𝑤𝑖,𝑗 =∑ (𝑟𝑥,𝑖−�̅�𝑖)(𝑟𝑥,𝑗−�̅�𝑗)𝑥∈𝑈

√∑ (𝑟𝑥,𝑖−�̅�𝑖)2𝑥∈𝑈 √∑ (𝑟𝑥,𝑗−�̅�𝑗)2

𝑥∈𝑈

(2)

where𝑟𝑥,𝑖 is the rating given by user x on item i . 𝑟𝑦,𝑖 is the rating given by user y on item i. �̅�𝑖is

the average ratings on item i given by all users. �̅�𝑗is the average ratings on item j given by all

users and U is the set of all users in User-Item space.

The other important similarity metric which is most commonly used is cosine

similarity.It is a linear algebraic technique widely used in the fields of text mining and

information retrieval. Documents are represented as vectors(i.e., instead of users or items,

documents are used and instead of ratings, a frequency of words in the document bare used) .

It measures the similarity based on the angle formed between the vectors. For User –based

CF,the similarity between users x and y is defined as

𝑤𝑥,𝑦 = cos(�⃗�, �⃗�) =𝑥 . �⃗⃗�

|𝑥|∗|�⃗⃗�|=

∑ 𝑟𝑥,𝑖𝑟𝑦,𝑖𝑖∈𝐼

√∑ 𝑟𝑥,𝑖2

𝑖∈𝐼 √∑ 𝑟𝑦,𝑖2

𝑖∈𝐼

(3)

where 𝑟𝑥,𝑖 is the rating given by user x for item i , 𝑟𝑦,𝑖 is the rating given by user y for item i

and I is the list of all items in the User-Item space.

For Item–based CF ,the similarity between users i and j is defined as

𝑤𝑖,𝑗 = cos(𝑖, 𝑗) =𝑖 . 𝑗

|𝑖|∗|𝑗|=

∑ 𝑟𝑢,𝑖𝑟𝑢,𝑗𝑢∈𝑈

√∑ 𝑟𝑢,𝑖2

𝑢∈𝑈 √∑ 𝑟𝑢,𝑗2

𝑖∈𝐼

(4)

where𝑟𝑢,𝑖 is the rating given by user u for item i , 𝑟𝑢,𝑗 is the rating given by user u for item j

and U is the list of all users in the User-Item space.

The important step in collaborative filtering is to obtain recommendations.So

after calculating the similarity, nearest neighbors of the active user are chosen and a weighted

aggregate of their ratings is used to generate predictions [4].The weighted aggregate of other

ratings is given as follows.

𝑃𝑎,𝑖 = 𝑟�̅� +∑ (𝑟𝑥,𝑖−𝑟𝑥̅̅ ̅𝑥𝜖𝑈 ).𝑤𝑎,𝑥

∑ |𝑤𝑎,𝑥|𝑥𝜖𝑈 (5)



where 𝑃𝑎,𝑖 is the predicted rating for the active user an on item i , 𝑟𝑥,𝑖 is the rating given by x

on item i , 𝑤𝑎,𝑥 is the similarity coefficient between a and x ,𝑟�̅� is the average of ratings given

by user x.

For example, consider the user-item matrix in Table 2 and we want to find whether

Avatar movie can be recommended to user 3.So first we calculate similarity between user 3

and other users and is as follows w(1,2)=0.403072, w(1,3)=0.323575, w(2,3)=-0.41523,

w(3,4)=-0.2325, w(3,5)=0.518545, w(3,6)=0.382336 etc.,From the above values,the highest

similarity is between user 3 and user 5 Now we calculate the predicted rating for user 3 using

weighted aggregate of other ratings and is given as follows

P(3,3) = 𝑟3̅ +(𝑟2,3−𝑟2̅̅̅).𝑤3,2+(𝑟4,3−𝑟4̅̅̅).𝑤3,4+(𝑟5,3−𝑟5̅̅̅).𝑤3,5

|𝑤3,2|+|𝑤3,4|+|𝑤3,5|

=2+(3−3.66667).(−0.41′5213)+3.5−3).(−0.2325)+(4−3).(0.518545)

|−0.415213|+|−0.2325|+|0.518545|

=2+ 0.679116

1.166275+6

=2.058229

[2.2.2] MODEL BASED COLLABORATIVE FILTERING

The performance of the memory-based technique decreases when the user-item matrix

becomes extremely sparse, which is very common on e-commerce sites.In model-based

collaborative filtering, the model is generated from a sample of larger databases using machine

learning and data mining algorithms.This pre-computed model is used to generate

recommendations similar to memory-based CF.The model based collaborative filtering

resolves the major problem of recommender systems called sparsity by employing

dimensionality reduction techniques.Some of the dimensionality reduction techniques are

Singular Value Decomposition(SVD)[44-45],Matrix Factorization(MF)[46-48],Principal

Component Analysis(PCA) ,Latent Semantic Indexing(LSI),among which MF is rapidly

used.Dimensionality reduction helps in improving the performance of the recommender

systems.Widely used model-based techniques are Association rule mining[11],Bayesian

classifiers[5-13],Decision trees[9,11,12],Clustering[8][21],Latent factor models(MF using

SVD, ALS, Regularization)[27].Some of the model-based CF techniques are based on the bio-

inspired approaches like Neural networks[9,12,14-16], Fuzzy systems[17], Genetic

algorithms[18-20].Slope One Method deals with sparsity and cold start problems. It is used to

predict the vacant ratings wherever necessary[26].Classification algorithms like Association

rule mining, Decision trees, Naïve Bayesian classification, SVM etc are used if the user ratings

are categorical.If the user-item matrix is of multi-class data then it is converted to binary class

for simplicity. Regression models are used if the user ratings are numerical. [Figure-6]

describes the architecture of Model-based Collaborative filtering.

[2.2.3] HYBRID COLLABORATIVE FILTERING

This technique is used to improve the prediction accuracy by combining some of the advantages

of both neighborhood based and model based collaborative filtering techniques.

Neighbourhood based approach does not perform well with sparse data in the user-item matrix

and it also suffers from scalability i.e., if the user-item matrix consists of more number of items

or users, then the entire computation has to be done to generate a single recommendation. A



model-based approach is extremely complex to generate a model if the dataset consists of

multiple parameters. The construction of the model takes a large amount of time and machine

and user perspective are different, so if the assumptions of the model do not fit the data, results

in wrong predictions.To avoid these problems hybrid collaborative filtering generates

algorithms using models that are fast and easy to calculate[40].

The recommendations generated by this approach are more accurate compared to

memory based and model based collaborative filtering algorithms.ParitoshNagarnaik et al.[28]

Uses a hybrid collaborative filtering approach to generate web page recommendations using

pattern discover algorithms such as clustering and association rule mining.K. Yu, A.

Schwaighofer et al. [29]combined memory based and model based techniques to generate

predictions based on the user profiles and posterior distribution of user ratings.D. M. Pennock,

E. Horvitz et al.[30] introduced Personality Diagnosis which selects the active user uniformly

at random and calculates same personality type(neighbors for the active user) by adding

Gaussian noise to his or her ratings.

[2.3]

HYBRID RECOMMENDER SYSTEMS

To overcome the individual disadvantages of Content-based RS and Collaborative RS, both are

combined together to form a new Recommender system called Hybrid Recommender system.

It uses the advantages of both the systems to optimize the predictions. There are multiple ways

to build a hybrid recommender system. They are

1. Both the approaches are individually implemented and the results are combined

in generating the recommendations.

2. Content-based techniques are used in Collaborative recommender systems.

3. Collaborative techniques are used in Content –based Recommender systems.

4. Both are combined together as a model.

[Table 3] depicts different combinations of to build a hybrid model.

The different hybridization methods are

Weighted Hybrid

It integrates the results of each technique and generates new recommendations. It linearly

combines the predicted ratings. Initially, the predicted ratings are considered equal and

gradually weights are added during the process of generating Top-N recommendations. P-

Figure:6.Architecture for Model-based Collaborative filtering



Tango[31] is an example of weighted hybrid recommender system which combines Content-

based RS and Collaborative RS.For further examples refer (Pazzani1999), (Towle&Quinn

2000).The advantage of this techniques is that all the benefits of the recommendation systems

are used during the process of generating recommendations.

Cascade Hybrid

It follows an iterative approach in generating the recommendations.Initially, one

recommendation technique is used to generate a coarse list of predictions and again a second

recommendation technique is used on this coarse list of predictions to generate a finer

recommendations.The advantage of this technique is, it is optimal and tolerant to noise as low-

priority recommendations are filtered at the first iteration and in the second iteration, the only

high-level coarse list of predictions are processed to generate the final recommendations.Its

disadvantage is, it adds complexity. EntreeC[13]-a restaurant recommender system, Fab are the

examples of cascade hybridization.

Table 3 Different ways to build a hybrid model

SNo Approach

1

2

3

4

Switching Hybrid

It uses the strategy of switching in between the recommendation techniques based on

some criteria. DailyLearner is an example of switching technique which uses content and

collaborative techniques where content-based recommendation technique is employed first and

collaborative recommendation technique is implemented later when the result drawn from

content based RS are with low confidence. The advantage of this approach is it uses all the

benefits of recommendation techniques used.The disadvantage of this technique is the criterion

has to be determined on which switching has to take place.It results in another level of

parameterization.

CBRS CFRS Recommendations

CFRS CBRS Recommendations

CBRS

CFRS Recommendations

MODEL

CBRS CFRS

Recommendations



Mixed Hybrid

It is a combination of more than one recommendation systems.The recommendations are drawn

individually by each technique and the results are combined to generate final recommendations.

Each item has multiple recommendations associated with it from different

recommendationtechniques[2].PTV[32] – A television show recommender system,

PickAFlick[34], ProfBuilder[33] are some of the examples of mixed hybrid recommendation

techniques.

Feature combination Hybrid

It is another hybrid approach which combines content and collaborative filters. The predicted

rankings generated by the collaborative technique is taken as an extra feature and on this

augmented dataset content based filter is employed to generate recommendations. Pipper is an

example of feature combination hybrid technique which uses ratings of collaborative filters as

a feature in the content-based system for movie recommendations[35].GroupLens is another

example. The feature that is combined with the data set may not be always collaborative filters.

Feature augmented Hybrid

Feature augmented hybrids are superior to feature combination hybrids. It employs any

technique to predict ratings or classify the dataset and this output is taken as input to the

recommendation process. Libra is an example of feature augmentation which generates book

recommendations using content-based system based on the data found in Amazon.com using a

naive Bayes text classifier[36].

Meta–level Hybrid

It is another type of combining the recommendation systems. The model generated by the first

system is taken as input to the other. The difference between the feature augmentation and meta

level is, in feature augmentation, the feature generated by the model is taken as input to the

second whereas, in metal level hybrid, the entire model is taken as input to the second. Fab,

LaboUr are the examples of meta-level hybrids.

[Table 4] describes the various approaches used by the recommendation systems.

[3] EVALUATION METRICS

To evaluate the performance of the recommendation systems and accuracy of the

recommendations generated, a myriad evaluation metrics are used. Herlockeret.al[37]., broadly

classified the evaluation metrics into predictive metrics and recommendation accuracy metrics.

[3.1] Predictive metrics

It measures the difference between the predicted ratings generated by the system and the real

ratings.The most popular prediction metrics are Mean Absolute Error(MAE), Rooted Mean

Squared Error(RMSE), Normalized Mean Absolute Error(NMAE) etc. Predictive metrics are

also called as statistical accuracy metrics.

MAE: It is the most common and popularly used metric. It is the absolute difference between

the predicted ratings and real ratings. It is defined as follows

𝑀𝐴𝐸 =∑ |𝑝𝑖−𝑟𝑖|𝑁

𝑖=1

𝑁 (6)

where N is the total number of ratings of all the users on the item i, 𝑝𝑖 is the predicted rating

and 𝑟𝑖 is the real rating. Lower the MAE indicates that the recommendation engine is accurately



predicting the user ratings.(𝑝𝑖 − 𝑟𝑖)represents the error in prediction.

RMSE: It has become more popular after its usage in Netflix Prize Competition. Willmott et

al.[38] claimed thatRMSE is more appropriate to represent model performance than the MAE

when the error distribution is expected to be Gaussian[39].It is defined as follows

𝑅𝑀𝑆𝐸 = √∑ (𝑝𝑖−𝑟𝑖)2𝑁𝑖=1

𝑁 (7)

Table 4 Classification of Recommendation Systems

Recommendation

approaches Description Limitations

Content based RS

Recommendations are generated based

on user and item features

ColdStart

Problem

Over

Specialization

Coll

abora

tive

filt

erin

g R

S Memory-based

Collaborative

Filtering

Rec

om

men

dat

ions

are

gen

erat

ed b

ased

on t

he

use

r pas

t his

tory

It uses statistical methods to find the

similar users and items.It is also

called as Neighborhood-based CF.

It is further categorized into User-

based CF and Item-based CF

ColdStart

Problem

Sparsity

Gray Sheep

Shilling Attacks

Overfitting

Scalability

Synonymity

Trust

Model-based

Collaborative

Filtering

It uses machine learning and data

mining algorithms to generate a

model. This model is responsible for

generating the recommendations.

Hybrid

Collaborative

Filtering

It combines both memory based and

model based approaches.

Hy

bri

d R

S

Weighted

Hybridization

Rec

om

men

dat

ions

are

gen

erat

ed

by

mer

gin

g

conte

nt

bas

ed a

nd

co

llab

ora

tive

filt

erin

g t

echniq

ues

The results of several

recommendation techniques are

integrated to generate a final

recommendation. Weights are

added linearly to individual results

of each RS to draw a single

recommendation.

Ex

pen

siv

e

Co

mp

lex

to

im

ple

men

t

Switching

Hybridization

The system switches between the

recommendation techniques based

on some criteria.

Cascade

Hybridization

It is an iterative procedure where

one recommendation technique

refines the output of another

recommendation technique.

Mixed

Hybridization

Recommendations generated from

individual techniques are mixed to

generate the final recommendation.



where N is the total number of ratings of all the users on item i ,𝑝𝑖 is the predicted rating and

𝑟𝑖 is the real rating. This metric puts more emphasis on larger absolute error because,on a 5-

point scale, a 1-point error may not be perceptible by the user(items rated with either 4 or 5

points are good recommendations) whereas with a 4-point error, the algorithm could be

recommending bad items[40] and the lower the RMSE is, the better the recommendation

accuracy.

[3.2] Recommendation accuracy metrics

These are further classified into classification accuracy metrics and rank accuracy

metrics. These metrics are also called as decision support accuracy metrics.

[3.2.1] Classification accuracy metrics

Classification accuracy measures how well the recommendation system differentiates

relevant(good) items from non-relevant(bad) items.Some of the classification accuracy metrics

are precision, recall, F-measure.

Precision: It is the ratio of relevant recommended items to the total recommended items.It is

defined as follows.

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =𝑅𝑒𝑙𝑒𝑣𝑎𝑛𝑡𝑅𝑒𝑐𝑜𝑚𝑚𝑒𝑛𝑑𝑎𝑡𝑖𝑜𝑛𝑠

𝑇𝑜𝑡𝑎𝑙𝑅𝑒𝑐𝑜𝑚𝑚𝑒𝑛𝑑𝑎𝑡𝑖𝑜𝑛𝑠 (8)

Recall: It is the ratio of relevant recommended items to the total relevant items. It is defined as

follows

𝑅𝑒𝑐𝑎𝑙𝑙 =𝑅𝑒𝑙𝑒𝑣𝑎𝑛𝑡𝑅𝑒𝑐𝑜𝑚𝑚𝑒𝑛𝑑𝑎𝑡𝑖𝑜𝑛𝑠

𝑇𝑜𝑡𝑎𝑙𝑅𝑒𝑙𝑒𝑣𝑎𝑛𝑡𝐼𝑡𝑒𝑚𝑠 (9)

It is preferable to have high precision and recall, but both the metrics are inversely

proportional i.e., when precision increases, recall decreases and vice versa[40].So most of the

systems use another measure called f-measure.

F-measure: Itis a combination(harmonic mean) of precision and recall and is defined as

follows.

𝐹 − 𝑚𝑒𝑎𝑠𝑢𝑟𝑒 =2∗𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛∗𝑟𝑒𝑐𝑎𝑙𝑙

𝑝𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑟𝑒𝑐𝑎𝑙𝑙 (10)

The best f-score is 1 and the worst is 0.

[3.2.2] Rank accuracy metrics

When the number of recommended items are large, then the users usually prefer the top-

Feature-

Combination

Features drawn from different

algorithms are augmented and taken

as input for a recommendation

algorithm.

Feature-Augmented

It is superior to feature combination

hybrid. Features generated from one

hybrid is taken as input to the

second.

Meta-level

Hybridization

It is a combination of two hybrid

methods.



N recommendations from the list. This creates more errors than to the items placed at the last

of the recommendation list. Rank accuracy metrics considers this situation and it measures the

ability of the recommendation algorithm to produce the recommendation list that matches with,

how the user would have ordered the same list of items. For example, if the top ten list of items

generated by the recommendation algorithm is relevant to the user and if the user does not feel

that the top item in the list is good(relevant), then the rank accuracy metric score for that

specific algorithm is low. Some of the rank accuracy metrics are ROC, half-life utility (HL)

and discounted cumulative gain(DCG).

ROC: Swets[41,42] introduced ROC metric to the information retrieval community, which was

very popular in signal detection theory. It is an alternative to precision and recall. It is a

graphical representation of what extent the recommendation algorithm discriminates

good(relevant) items with bad(irrelevant) items. It is also called as Relative Operating

Characteristic.The area below the ROCcurve is called as Swet’s A measure, which is used as a

single metric of the system's ability to discriminate between good and bad items, independent

of the search length[37].

Half – lifeUtility : It was proposed by Breese et al.[43]. It evaluates the utility of an ordered

list of recommendations by the user. It is based on the fact that, the utility of the user is greater

for the beginning list of items and it exponentially decreases going deeper into the list. It is

defined as follows

𝐻𝐹 =∑ (𝑟𝑎𝑖−𝑑,0)𝑖

2(𝑖−1)(𝛼−1) (11)

where𝑟𝑎𝑖 𝑖s the rating of the user a on item i of the recommended list, d is the default rating,

and α is the half-life. The half-life is the rank of the item on the list such that there is a 50%

chance that the user will view that item[37].

Discounted Cumulative gain(DCG): It is defined as

𝐷𝐶𝐺 = 𝑟𝑎𝑖 + ∑𝑟𝑎𝑖

log2(𝑖)𝑘𝑖=2 (12)

where 𝑟𝑎𝑖 represents the rating of the user a on item i of the recommended list, d is the default

rating, and α is the half-life. The half-life is the rank of the item on the list such that there is a

50% chance that the user will view that item[37] and k is the rank of the evaluated item.

[4] CONCLUSION

The review of various approaches like content-based, collaborative and hybrid alleviates

the problem of information overload and opens new opportunities of retrieving personalized

information to the users. This paper discussed the steps involved in recommender systems, the

approaches for recommendation systems and highlighted their strengths and challenges with

diverse kind of hybridization strategies used to improve their performances. Various learning

algorithms used in generating recommendation models and evaluation metrics used in

measuring the quality and performance of recommendation algorithms were discussed. This

review will serve as a blue print to the researchers in improving the state of the art

recommendation techniques besides developing new approaches.



REFERENCES

[1] G. Suganeshwari and S.P. Syed Ibrahim,”A Survey on Filtering collaborative Based

Recommendation System", March 2016 DOI: 10.1007/978-3-319-30348-2_42.

[2] F.O Isinkaye, Y.O.Folajimi, B.A.Ojokoh," Recommendation systems: Principles, methods, and

evaluation”, Egyptian Informatics Journal, pp. 261-273(2015)

[3] Ivens Portugal, PauloAlencar, DonaldCowan,"The use of Machine Learning Algorithms in

Recommender Systems: A Systematic Review”,

[4] Xiaoyuan Su and TaghiM.Khoshgoftaar," A survey of Collaborative Filtering Techniques",

Hindawi Publication Corporation-Advances in Artificial Intelligence, Vol 2009, Article ID

421425.

[5] S.B. Cho, J.H. Hong, M.H. Park, Location-based recommendation system using Bayesian user’s

preference model in mobile devices, Lecture Notes in Computer Science 4611 (2007) 1130–1139.

[6] Wang, Y., Chan, S. C. F., & Ngai, G. (2012, December). Applicability of demographic

recommender system to tourist attractions: a case study on trip advisor. In Proceedings of the The

2012 IEEE/WIC/ACM International Joint Conferences on Web Intelligence and Intelligent Agent

Technology-Volume 03 (pp. 97-101). IEEE Computer Society.

[7] Seric, L., Jukic, M., &Braovic, M. (2013, May). Intelligent traffic recommender system. In

Information & Communication Technology Electronics & Microelectronics (MIPRO), 2013 36th

International Convention on (pp. 1064-1068). IEEE.

[8] Ericson, K., &Pallickara, S. (2013). On the performance of high dimensional data clustering and

classification algorithms. Future Generation Computer Systems,29(4), 1024-1034.

[9] Felden, C., &Chamoni, P. (2007, January). Recommender systems based on an active data

warehouse with text documents. In System Sciences, 2007. HICSS2007. 40th Annual Hawaii

International Conference on (pp. 168a-168a). IEEE.

[10] Gorodetsky, V., Samoylov, V., &Serebryakov, S. (2010, August). Ontology–based context

dependent personalization technology. In Web Intelligence andIntelligent Agent Technology (WI-

IAT), 2010 IEEE/WIC/ACM International Conference on (Vol. 3, pp. 278-283). IEEE.

[11] Lucas, J. P., Segrera, S., & Moreno, M. N. (2012). Making use of associative classifiers in order

to alleviate typical drawbacks in recommender systems. ExpertSystems with Applications, 39(1),

1273-1283.

[12] Marović, M., Mihoković, M., Mikša, M., Pribil, S., & Tus, A. (2011, May).Automatic movie

ratings prediction using machine learning. In MIPRO, 2011Proceedings of the 34th International

Convention (pp. 1640-1645). IEEE.

[13] Burke, R.: Hybrid recommender systems: survey and experiments. User Model. User-Adap.

Inter. 12(4), 331–370 (2002)

[14] Alvarez, S. A., Ruiz, C., Kawato, T., &Kogel, W. (2011). Neural expert networks for faster

combined collaborative and content-based recommendation. Journal ofComputational Methods

in Sciences and Engineering, 11(4), 161-172.

[15] Martineau, J. C., Cheng, D., &Finin, T. (2013). TISA: topic independence scoring algorithm. In

Machine Learning and Data Mining in Pattern Recognition(pp. 555-570). Springer Berlin

Heidelberg.



[16] Ingoo, H., Kyong, J.O., Tae, H.R.: The collaborative filtering recommendation based on

SOMcluster-indexing CBR. Expert Syst. Appl. 25, 413–423 (2003).

[17] R.R. Yager, Fuzzy logic methods in recommender systems, Fuzzy Sets andSystems 136 (2)

(2003) 133–149

[18] R. Garcı´ a, X. Amatriain, Weighted content based methods for recommending connections in

online social networks, in Proceedings of the 2010 ACMconference on Recommender Systems,

2010, pp. 68–71.

[19] L.Q. Gao, C. Li, Hybrid personalized recommended model based on genetic algorithm, in

International Conference on Wireless Communication, Networks and Mobile Computing, 2008,

pp. 9215–9218.

[20] Qian Wang,Xianhu Yuan,Min Sun : Collaborative filtering recommendation algorithm based on

hybrid user model ,in Proceedings of the Seventh International Conference on fuzzy Systems and

Discovery,2010,pp.1985-1990

[21] Ericson.K & Pallickara.S ,On the performance of distributed clustering algorithm in file and

streaming processing systems.,In Utility and Cloud Computing(UCC),2011 Fourth IEEE

International ,pp.33-40,IEEE.

[22] Arenas-Garcia.J,Meng.A,Petersen.K.B,Lehn-Schioler.T,Hansen.L.K & Larsen.J (2007

August),Unveiling music structure via PLSA similarity fusion.In Machine Learning for Signal

Processing,2007 IEEE Workshop .(pp.419-424).IEEE.

[23] Schelter.S,Boden.C,Schenck.M,Alexander roy.A & Markl.v(2013 October),Distributed Matrix

factorization with mapreduce using a series of broadcast joins.,In proceedings of the 7 th ACM

Conference on the Recommendation Systems(pp.281-284),ACM.

[24] Takacs.G,Pilaszy.I,Nemeth.B & Tikk.D(2009).Scalable collaborative filtering approaches for

large recommendation systems.The Journal of Machine Learning Research,10,623-656.

[25] Tewari .N.C,Koduvely.H.M,Guha.S,Yadav.A & David.G(2013 October),Mapreduce

implementation of Variational Bayesian Probabilistic Matrix Factorization algorithm in Big

Data,2013 IEEE International Conference (pp.145-152),IEEE.

[26] Pu WANG, HongWuYE: A Personalized Recommendation Algorithm Combining Slope One

Scheme andUser Based Collaborative Filtering.In2009 International Conference on Industrial and

Information Systems (pp.152-154)

[27] Robert Bell Yehuda Koren and Chris Volinsky. Matrix factorization techniques for recommender

systems. In IEEE Computer, volume 42 (8), pages30–37, 2009

[28] ParitoshNagarnaik and Prof.A.Thomas: Survey on Recommendation System Methods.In IEEE

SPONSORED 2ND INTERNATIONAL CONFERENCE ON ELECTRONICS AND

COMMUNICATION SYSTEM(ICECS 2015)

[29] K. Yu, A. Schwaighofer, V. Tresp, X. Xu, and H.-P. Kriegel, “Probabilistic memory-based

collaborative filtering,” IEEE Transactions on Knowledge and Data Engineering, vol. 16, no. 1,

pp. 56–69, 2004.

[30] D. M. Pennock, E. Horvitz, S. Lawrence, and C. L. Giles, “Collaborative filtering by personality

diagnosis: a hybrid memory- and model-based approach,” in Proceedings of the 16th Conference

on Uncertainty in Artificial Intelligence (UAI '00), pp. 473–480, 2000.

[31] Claypool M, Gokhale A, Miranda T, Murnikov P, NetesD, Sartin M. Combining content- based

and collaborative filters in an online newspaper. In: Proceedings of ACM SIGIR workshop on

recommender systems: algorithms and evaluation, Berkeley, California; 1999.



[32] Smyth B, Cotter P. A personalized TV listings service for the digital TV age. J Knowledge-Based

Systems 2000;13(2–3):53–9.

[33] Wasfi AM. Collecting user access patterns for building user profiles and collaborative filtering.

In: Proceedings of the 1999international conference on intelligent user, Redondo Beach,

CA;1999. p. 57–64.

[34] Burke R, Hammond K, Young B. The FindMe approach to assisted browsing. IEEE Expert

1997;12(4):32–40.

[35] Basu C, Hirsh H, Cohen W. Recommendation as classification: using social and content-based

information in recommendation.In: Proceedings of the 15th national conference on artificial

intelligence, Madison, WI; 1998. p. 714–20.

[36] Mooney RJ, Roy L. Content-based book recommending using learning for text categorization.

In: Proceedings of the fifth ACMconference on digital libraries. ACM; 2000. p. 195–204.

[37] HERLOCKER, J. L., KONSTAN, J. A., TERVEEN, L. G., AND RIEDL, J. T. 2004. Evaluating

collaborative filtering recommender systems. ACM Trans. Inform. Syst. 22, 1, 5–53.

[38] Willmott, C. J., Matsuura, K., and Robeson, S. M.: Ambiguitiesinherent in sums-of-squares-

based error statistics, Atmos. Environ.,43, 749–752, 2009.

[39] T.Chai, R.R Draxler : Root mean square error (RMSE) or mean absolute error (MAE)? –

Arguments against avoiding RMSE in the literature, Geo-scientific Model Development.,2014.

[40] Cacheda, F., Carneiro, V., Fernandez, D., and Formoso, V. 2011. Comparison of collaborative

filtering algorithms: Limitations of current techniques and proposals for scalable, high-

performance recommendersystems. ACM Trans. Web 5, 1, Article 2 (February 2011), 33 pages.

[41] SWETS, J. A. 1963. Information retrieval systems. Science 141, 245–250.

[42] SWETS, J. A. 1969. Effectiveness of information retrieval methods. Amer. Doc. 20, 72–89.

[43] BREESE, J. S., HECKERMAN, D., AND KADIE, C. 1998. Empirical analysis of predictive

algorithms for collaborative filtering. In Proceedings of the 14th Conference on Uncertainty in

Artificial Intelligence(UAI-98). G. F. Cooper, and S. Moral, Eds. Morgan-Kaufmann, San

Francisco, Calif.,43–52.

[44] M.G. Vozalis, K.G. Margaritis, Using SVD and demographic data for the enhancement of

generalized collaborative filtering, Information Sciences 177(2007) 3017–3037.

[45] S. Zhang, W. Wang, J. Ford, F. Makedon, Using singular value decomposition approximation for

collaborative filtering, in IEEE International Conference on E-Commerce Technology, 2005, pp.

1–8.

[46] Y. Koren, R. Bell, CH. Volinsky, Matrix factorization techniques for recommender systems, IEEE

Computer 42 (8) (2009) 42–49.

[47] X. Luo, Y. Xia, Q. Zhu, Incremental collaborative filtering recommender based on regularized

matrix factorization, Knowledge-Based Systems 27 (2012)271–280

[48] X. Luo, Y. Xia, Q. Zhu, Applying the learning rate adaptation to the matrix factorization based

collaborative filtering, Knowledge-Based Systems 37(2013) 154–164.



AUTHORS INTRODUCTION

V Lakshmi Chetana has obtained her MCA degree from Acharya Nagarjuna Uni versity and

M.Tech in Computer Science and Engineering from JNTU Kakinada with distinctions. She is now

working as Assistant Professor in the department of CSE at DVR & Dr. HS MIC College of

Technology,Vijayawada. She has 8yrs of Experience in teaching. She has more 10 international Journal

and Conference papers. Her areas of interest in research are Big Data Analytics, Data Science,Artificial

Intelligence and Machine Learning.

Dr. S.P SYED IBRAHIM has received his B.E(CSE) degree from Bharathidasan University ,

M.E(CSE) degree from Anna University and PhD in CSE from Anna University .He is now working as

a professor in the School of Computing Sciences and Engineering at VIT University,Chennai Campus

and also coordinates Data Analytics Research Group.His research interests includes Data Mining,Big

Data Analytics.He has published more than 40 International Journal and CDonference papers and

completed five Industrial consultancy worl related to Big Data Analytics.

Dr.B Rama has obtained her PhD from Sri Padmavathi Mahila Viswa Vidyalayam, Tirupati.She

has 5 years industrial and 8 years of teaching experience .She is now working as HoD in the Department

of Informatics/ Computer Science, University college Kakatiya University, Warangal .She is the

recipient of best paper award for the paper “ A Data mining perspective for forest fire prediction”

presented in International Conference. She has more than 45 international journal and IEEE and

Springer conference papers. She organized various workshops and delivered many guest lectures.Her

research areas of interest are Data Mining,Artificial Intelligence and Big Data Analytics.

Dr P Kiran Sree has obtained his B.Tech degree in Computer Science & Engineering from

JNTU Hyderabad and M.E in Computer Science & Engineering from Anna University with

distinctions. He has obtained his Ph.D in Artificial Intelligence from Jawaharlal Nehru Technological

University-Hyderabad. His interests include Parallel Algorithms, Artificial Intelligence, and

Bioinformatics. His bibliography was listed Listed in Marquis Who’s Who in the World, 28 Edition

(2012), U.S.A. He is the recipient of Bharat Excellence Award from Dr GV Krishna Murthy, Former

Election Commissioner of India in 2013. He worked as Principal I/C of the N.B.K.R. Institute of

Science & Technology, the second oldest private engineering college of the state for two years

(04/12/2009-02/12/2011). He is the editor in chief of international journal of Parallel and Cloud

Computing Research (PCCR). He has published 52 research papers in international journals and

conferences. He is recognized supervisor in the stream of Artificial Intelligence for guiding Ph.D

scholars at VIT and KL universities.