A Survey on Recommendation System based on Knowledge Graph and Machine Learning

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 06 | Jun 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 427
A Survey on Recommendation System based on Knowledge
Graph and Machine Learning
Shailesh D.Kalkar1, Prof. Pramila M.Chawan2
1M. Tech Student, Dept of Computer Engineering, and IT, VJTI College, Mumbai, Maharashtra, India
2Associate Professor, Dept of Computer Engineering, and IT, VJTI College, Mumbai, Maharashtra, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - The work of recommendation system is to
guess the thought process of user and to predict the
interests of the users. This system can provide suitable
information to the user based on the needs of the user
while taking care of the interests of the user. For giving
better recommendation the data need to be analyzed more
effectively. There are various recommendation systems that
had been built using different approaches. And till today the
research in such system is popular as now days there are
many OTT platforms increasing day by day, many shopping,
travel etc. websites are increasing who want to improve
their recommendations to the users rapidly. So, the main
goal of this paper is to provide the overview of various
approaches related to the recommendations system as well
as to do the comparative study of them based on certain
features. After reviewing various papers we observed there
are many recommendation systems were built which were
majorly based on traditional methods. However, nowadays
recommendation system based on knowledge graph have
catches the attention of researchers and the industries as
they are capable of solving various performance and
information sparsity related problems and gives better
recommendations as compared to previous
approaches.Machine learning is used along with knowledge
graph to increase the performance of the system. Also, we
will study various proposed algorithms in the papers which
have used knowledge graph for better recommendation.We
have also given brief idea about our proposed
system.Finally, we will suggest different future paths of
research in the domain of recommendation system.
Key Words: Knowledge graph, Machine Learning,
Recommendation system, Technologies, Approaches.
1.INTRODUCTION
Recommendation system are used to recommend various
services to the user based on the activity and the
needs of the user. Recommendation system is widely used
in e-commerce[4], entertainment, e-learning, search
engines etc domains to provide better results to the user.
There are certain approaches which are used for building
recommendation system like content based filtering,
collaborative filtering[22],hybrid filtering, context filtering
methods. So, in this paper we are going to discuss about
those methods in details. Also, due to fast development
of internet, the size of data in each of the application
have increased drastically. Hence, it becomes difficult for
the user to choose their interests related items or services
from huge amount of data.
The recommendation algorithm is the main part of rec-
ommendation system. Collaboration filtering recom-
mendation gives recommendations based on the
preferences of other similar users for the product features
whereas in content based recommendation, this model
uses the item’s content features. Collaborative filtering is
widely used as compared to content based filtering
because they are efficient in capturing the preference of
user and can be easily applied in multiple scenarios. But
in content based filtering lot of effort is required for
extracting the features. Also, there is context based
recommendation in which as per the contextual in-
formation of the user like location, time etc, we recommend
services to him/her. In this model, we retrieve patterns
from the website or application based on the user’s past
interaction with the system and then we provide
recommendations to the user. Collaborative filtering
suffers from problems like data sparsity and cold start
problems. To solve these issues, hybrid recommendation is
proposed which merges the content level similarity and
interaction level similarity. In this model other extra
information is also explored like item reviews, item
attributes etc. In recent years, introduction of knowledge
graphin recommendation system have gained the attention
of researchers and industries. A knowledge graph is a kind
of heterogeneous graph in which nodes indicate entities
and edges indicate the relationship between the entities.
For better understanding the mutual relationship
between the items we can map items and their
attributes in the knowledge graph. Also, we can map user
information into the knowledge graph, this creates the
relationship between users and items and can we can
capture user preference more accurately. In this survey
paper we’re also going to do survey of graph based
recommendation system using knowledge graph. The
paper proposed by C.Liu[16] discussed that how one can
include the knowledge graph embedding in the
recommendation system. And the paper proposed by Z.
Sun[17] explains that how knowledge graph can work as
additional information for recommendationsystem, but the
categorization of recommendation approaches are not fine
grained in this paper.

1.1 Organization of the paper
In heading 2, we discuss about various concepts used and
foundations of knowledge graph required in the
recommendation system.
In heading 3, we discuss the basic background required for
most of the recommendation system.
In heading 4, we discuss various standard recommendation
approaches used in the recommendation system.
In heading 5, we discuss about various knowledge graph
based recommendation systems and understand the
different approaches used in it.
In heading 6, we discuss about comparative study of
literature review.
In heading 7, we discuss about our proposed system.
In heading 8, we discuss about potential research paths in
the recommendation system.
In heading 9, we conclude our survey paper and also
mention the references used for this paper.
2. VARIOUS CONCEPTS
- Recommendation Systems
The work of recommendation system is to predict and
recommend the unobserved items or services to the user.It
can be done in the following manner like- first the system
learns the representation of target user ui and candidate
item vj. Second, for modelling the preference of ui and vjit
learns from the scoring function which is defined as f( ui
x vj gives y cap i,j).Third, We sort the preference scores for
items to generate recommendations for theuser. There are
many surveys available related to the recommendation
system with different emphasis. This system is majorly
classified into three categories like content based,
collaborative filtering based and hybrid recommendation
system. Amongst these categories collaborative filtering
recommendation is the most popular one. However,due to
development of various deep learning methods the
architecture of recommendation system have changed
drastically. The paper proposed by S.Zhang,[18] explored
how various deep learning techniques are adopted in the
current recommendation systems.
- Knowledge Graph
Nowadays knowledge graph is emerging as a abstraction
tool for organizing world’s structured knowledgepresent in
the internet. It is also used for integrating information
gathered from multiple data sources. It also plays a major
role in machine learning as a method for incorporating
world knowledge and also as a target knowledge
representation for the extracted knowledge. It also
explains what it has learned. It’s definition is like :-Given
entities(E),relations(R) and knowledge graph is a directed
multi relational graph(G) that comprises of triples
(S(Subject),P(property),O(Object)).Each edge in the
knowledge graph is of the form(head entity, relation, tail
entity).It is like an instance of a Heterogeneous Information
network. Knowledge graph is created and applied in many
domains like question answering system, recommendation
system, search engines etc. There are two types of
knowledge graph like :- First, Item Knowledge graph-In
this, item and its related entities (like item attributes) act a
nodes. And edges can represent item’s attribute
relations(like category, brand etc.) or user related
relations(like co-buy, co-view etc.).Second, User-item
knowledge graph- In this, items users and their related
entities acts as nodes. Inspite of items related rela-tions
present in the item knowledge graph, relationships
between user and item are also included in user-item
knowledge graph(like mention, buy, click etc.).
- MetaPath
It is defined as path P=A0(R1)A1(R2). (Rk)Ak defined on
the graph of network G=(A,R),it defines a new composite
relation as R1R2. . . ..Rk between entities A0 andAk, where Ai
A and Ri R for i=0,. . . ,k.It is used to extract connectivity
features between two nodes in the graph.
- Metagraph
It is somewhat similar to metapath, it’s like another metas-
tructure which connects two entities in a graph. The major
difference between metapath and metagraph is that meta-
path defines only one relation sequence whereas meta-
graph is a combination of various metapaths. Metagraph
can express more structural information between the
entities as compared to metapath.
- Knowledge Graph Embedding
Knowledge graph is embedded into low dimensional
vector space[19] using knowledge graph embedding. After
applying embedding on the knowledge graph each com-
ponent of graph(like entity and relation) is represented
by a dimensional vector. This dimensional vector still
preserves the properties of graph and we can quantify
these properties semantically or by higher order proximity
in the graph. For understanding various knowledge graph
embedding algorithms we can refer to [20][21].
3. BACKGROUND
For the purpose of recommending appropriate items to the
users as per needs of the user, this system collects as well
as process the useful information about the users and the
items.

3.1 Item Profiles
The details about the items need to be considered for
better recommendation. So, these details about the items
can be obtained from the dataset from standard websites
like kaggle etc. or we can extract the information about
the items from the web using web crawler etc. and
then do preprocessing of the data. The data extracted
about the items from web can be in structured, semi-
structured or unstructured form. So, preprocessing will be
required for the semi structured or unstructured data.
Once the data about the items is ready then we can then
use it for building recommendation system. If we are using
knowledge graph based approach then we use this data to
create item based entities in the knowledge graph.
3.2 User Profiles
The recommendation system fully the utilizes the interac-
tions of the user with the system or the service like user’s
past purchase history ,user’s activity etc. in order to give
better recommendations to him or her. There are various
ways by which we collect information about the users
like:-
(i) Explicit way Feedback- In this method , user is
involved to get information about him/her. The user can
asked to fill certain forms or asked about their
preferences related to the different services. The
information about the user can be asked in various forms
like numerical scale(like rate item or service in scale of 1
to 5),binary scale(like product good or bad),ordinal
scale(like user chooses from the list of feeling which ex-
plains its inclination towards a particular
service),descriptive way(like reviews in the form of text
about particular service).Ifthe reviews in the descriptive
way are utilized properly then it can help us to
understand the preferences of the user related to
various items. There are certain aspects of reviews like
considering the contextual information, multi-view of a
particular review,comparison based reviews etc. The
methods of extracting useful information from the user
review is usedin various papers.
(ii) Implicit way Feedback- In this kind of method user is
not involved in the process of gathering data about
him/her. Such methods does the proper analysis of the
user purchasing history like how many times user
bought a particular product, his rating for the product,
how much time the user spent on the website while
seeing the product etc. We can also use already available
datasets for such methods.
(iii) Hybrid way Feedback- In this method, both the
abovemethods can be used to gather proper information
about the user. In this approach both the above methods
can fulfill the drawbacks of one another. For better
understanding of the user’s behavior with the system we
can try to use implicit data as validation for the explicit
data provided by the user.
4.VARIOUS STANDARD RECOMMENDATION
APPROACHES
4.1 Content Based Recommendation
This kind of recommendation system uses the data
provided by the user in implicit way(like clicking on a
product etc.) or explicit way(like ratings or reviews
etc.).Based on these details user profile is generated and
used to give certain recommendations to the user. This
system also recommend items similar to the items which
were positively reviewed or rated/liked by the user in the
past. The user and item profiles consist of various features
of items and users respectively. For example - The user
profile can have attributes like userid, user review, user
rating, user purchase etc. The item profile can have
attributes like movieid, movie actors, moved I rectors etc.
If the user likes the comedy movies, so those comedy
movies which are not liked previously by the user will also
be recommended to user. There are main steps of
content based filtering like:-(i)first, the item attributes
are extracted to create item profile for all the items.(ii)
Then user profiles for each active users are
generated.(iii) Then user profile is compared with the item
profile.(iv) Then items are recommended to the user
which are new to the user and which matches the user
profile more.
There are certain approaches in this method like:-
(a) Preference based on product ranking: This approach is
used when items are described using different attributes.
In this method, user’s preference is denoted by (v1,. . . . .
. , vn, w1,. . . , wn), where vi is the value function which a
user specifies for a particular attribute ai and wi is the
relative importance(i.e.weight)of the attribute ai .Then
the utility of each product is calculated by multiplying
each vi with each wi and then doing the sum of all. So,
the product with high utility value are classified and
then recommended to the user.
(b) Exploiting terms on reviews for recommendation
purpose: It’s like a index based approach in which each
user is classified by the textual content of the reviews.
The term based user profiles are created by fetching
keywords from the user reviews and then weights are
assigned to the each extracted keywords by using TF-IDF
technique. The weight indicates how important is each
keyword for the user.
(c) Exploiting context from reviews for recommendation
purpose: In this approach context is extracted from
the textual information of user’s current scenario and
the features which are important to him. The utility
score for an item i of a user u is calculated as below:-

Utility(u,i)=a*predicted Rating (u,i)+(1-a)*context score
(u,i)Here a is a constant value which indicates the weight
of predicted rating. In above formula predicting rating is
calculated using standard item based collaborative
approach and context score indicates the importance of
item i to the target user’s context.
4.2 Collaborative filtering Based
Recommendation
In this kind of recommendation system, we try to find the
similar users which matches the current users interest and
recommend those items to current user which similar kind
of users had liked. In this approach, we never use the
attributes of items for the recommendation purpose.
Instead of that we try to classify the users into various
clusters of similar types and then recommend to each of
the user as per the preference of the cluster. There are
various type of collaborative filtering approaches :-
(i) User Based: In this approach, those items are recom-
mended to the target user which he/she may like,
according to the ratings or reviews given by the other
users who have similar kind of interest as that of a target
user. The major steps of the user based collaborative
filtering can be mentioned as below:- (a)First of all we
need to have user- item rating data (b)Then we need to
create user-user similaritymatrix using pearson correlation
as it is most famous and widely used method for finding
similarity in collaborative filtering. Similarity between two
users is found out by pearson’s correlation with values of -
1,0,1.-1 means two users are having interests completely
opposite, 0 means there is no relation between the
interests of the users and 1 means two users are having
exactly same kind of interests. (c)Then find the similar
users to target user with the help of similarity matrix.
(d)Then we need to generate candidate items for
recommending purpose. (e)Then we should rank the
candidate items as per their prediction. (f)Finally we need
to filter possible candidates items and show it to the target
user. Drawbacks:- It might be possible that there can be
more number of users than items which can lead to much
larger user similarity matrices which can result in various
performance and memory issues on the larger dataset. So,
to deal with such issues we need to use parallelisation
techniques.
(ii) Item Based: In this approach, we try to find out the
relationship between the items(like user who bought P,
also bought Q).Also, in this approach we predict the new
rating with the help of other items ratings given by the
user. Let us understand this approach with simple
example, lets say there are three users(A,B and C) and four
fruits(like grapes, strawberry, water melon, orange).User
A likesgrapes, water melon and orange. User B likes grapes
and watermelon and User C likes watermelon. So, here
grapes are watermelon are similar to each other as they
had been liked by the user A and B. So, if I want to
recommend more fruits to the user C and as I know
already know that he likes watermelon then it is mostly
likely that C might like grapes also as grape is similar to
watermelon. User based collaborative filtering have
certain drawbacks so to deal with it item based
collaborative techniques analyze the user-item matrix and
finds out the relationships between various items.Sobased
on the identified linear relationships between items it
gives the recommendation.
(iii) Matrix Factorization: In the matrix factorization, we
multiply two different entities to get the latent features.
With the help of user rating on the shop items(i.e. input),
we want to know how will the user rate the items so that
user can get the recommendation on the basis of
prediction. For example, we’ve customer rank table of 6
users and 6 movies, rating are the values from 1 to 5.As
every user will not rate every movie so there can be
missing values in the matrix and this results in a sparse
matrix. So, in place of missing entries we put zero and
filled values are given for the multiplication. Using this
method, lets say we have scenario like user 3 didn’t give
rating to the movie 3.So,we’d like to know whether
user 3 liked the movie 3 or not.So, using this method
we can discover other similar kin d of users with
similar kind of preferences of user 3 by taking the ratings
from the users of similar preferences to the movie 3 and
predict that whether user 3 will like movie or not. In this
we get the prediction rating of item by doing the dot
product of matrix p and q where p indicate the
relationship between user and features and q indicate the
relationship between an item and the features. R=PxQT.
There are various matrix factorization techniques like
SVD(Singular value decomposition),NMF(Non Negative
Matrix factorization),PCA(Principal Component Analysis)
which are used for finding latent factors from explicit
users feedback. Matrix factorization models are becoming
more efficient these days as in this we can consider the
text, time and social links which helps us to understand
the user behaviour in a better way. There are certain
problems which occurs in the collaborative filtering
approach like :- a)Earlier rate problem -this happens when
new user is using the system and has not rated that
much items that we can recommend items to them.
b)Sparsity problem - it happens when there is much little
information available to make appropriate prediction.
c)Cold start problem-it is situation when doing
prediction becomes difficult due users or items
added in the system are new.
4.3 Context Based Recommendation
The recommendation using context based
recommendation have mainly three approaches like pre
filtering, post filtering and contextual modelling. In the pre
filtering approach, we try to reduce the information of user

item matrix using context before applying any
recommendation algorithm. In the post filtering approach,
we try to reduce the information which is obtain after
applying recommendation algorithm using the context.
And in the contextual modelling approach, context is
included in the recommendation system. In the pre and
post filtering method, we generally use the existing
recommendation algorithms. But in case of context
modelling we can modify the existing recommendation
methods.
4.4 Demographic Recommendation Approach
In this approach work on the idea that users having similar
demographic attributes(like gender, age, city, job etc.) can
have similar kinds of trends in the future. For example you
might have noticed that when we go to different state
youtube based on your current location starts
recommending whatever popular videos are trending in
the state thinking that you might also like. This approach
focuses more on the demographic location of the user
rather than its evaluation history. As this approach does
not takes into consideration the user’s preference so it may
not always be the case that it will recommend items or
services to the user in the accurate manner.
4.5 Hybrid Filtering Approach
In this approach, we combine collaborative and content
based methods to overcome the drawbacks of both the
approaches. There are various ways in which we can
implement hybrid filtering like:- (i)After implementing
both content and collaborative filtering approach
separately then we can combine the prediction logic of
both methods. (ii)Using collaborative properties in content
based approach or vice-versa (iii)By modelling two or
more approaches together like content and collaborative
approach together.
4.6 Knowledge based Filtering Approach
This method suggest the items to the user as per how the
item satisfy the user preferences using domain knowledge.
The recommendation system using knowledge graph
should use majorly three kinds of knowledge like
knowledge about items, knowledge about users and
knowledge based on relationship between the item and the
user needs. knowledge graph provides extra information
which can be used to solve problems where were present
in content and collaborative filtering based approaches. It
uses domain knowledge for giving recommendations to the
user. However, there are also drawbacks of knowledge
based recommendation is that for creating such systems
we need to have skills related to knowledge engineering.
Semantic relationships present in the knowledge graph
can be utilized to improve the accuracy of recommendation
system and can increase the recommended items diversity.
The paper proposed by Zhang[22] suggest a method in
which collaborative filtering with implicit comments are
used and the interaction between the users and items are
learned by using knowledge graph embedding method.
knowledge graph based approaches have advantage over
traditional approaches. In this survey, we’ll analyze and
focus on various applications developed using knowledge
graph and various knowledge based technologies used for
recommendation system.
5. RECOMMENDATION SYSTEMS CREATED OVER
KNOWLEDGE GRAPHS
When Google introduced knowledge graph its purpose was
to improve the search engine’ s searching capability
and to improve the user experience. Knowledge graph
actually provides machine-readable data organized in the
form of graph. This graph data interconnects and describes
the entities. As we know that the data in a knowledge
graph can be accessed via web and can be consumed
automatically so because of these characteristics many
applications havebeen created using knowledge graph like
question answering system, recommendation systems etc.
Knowledge graph based recommendation system fully
utilizes the connections between the entities representing
the items, users and the interactions between them. These
connections can be direct or indirect. The relationships
present in the knowledge graph acts as a additional
information and this information can help us to deduce
inference between the entities to explore new
relationships. According to the paper proposed by Liu[23]
and Zou[24],the recent knowledge graph based
recommendation approaches can be divided majorly into
four types like linked open data based, knowledge graph
embeddding based ,ontologies based recommendation and
the path based recommendation.
a) Ontology based recommendation: In this method,
ontology is used for modelling the knowledge using
the information like users and their context, items
information and the information about domain. While
defining ontology we create various semantics and
structures which help us to create some rules for
generating recommendations to the user. So, the items
which fulfills the rules as per the user’s preference will
shown as recommendation to the user. The drawback of
this method is that it is time consuming process and it
also requires little bit of expertise.
b) Linked open data based recommendation: As rich
semantic information can be extracted from linked open
data so it can help us to find the similar attributes among
the items to be recommended. This method helps us to
overcome the problem of data sparsity. However ,there
are some recommendation systems which depends on
outside data so this kind of data can affect the
recommendation results.

c) Path based recommendation: This method is very
normal way of using knowledge graph in the
recommendation system. In this method we try to
explore different patterns of relationships between
entities in a knowledge graph and in this way it helps to
give better recommendation to the user. This method
actually depends on hardcoded metapaths and they are
difficult to optimize. The drawback of this method is
that when entities and relationships are not from
the same domain then in that case it becomes
difficult to design metapaths for it[24].
d) Embedding based recommendation: This method is
more popular because it provides an easy and efficient
wayfor generating recommendations[16].In this method,
we canconvert the knowledge graph by using knowledge
graph based embeddings then the recommendation
model can fully utilize the learned entities and
relationships embedding to produce better
results[24].The main goal of this method is to make
the processing of knowledge graph easier while
preserving its structure. Knowledge graph embedding
is used to enhance the information of users and
items and then it is used for calculating the similarity
between users and items[16].Some of the examples of
such embedding models are TransE, TransD, TransH,
TransR etc. Introduction of knowledge graph embedding
in recommendation system includes using of traditional
recommendation algorithm. In the paper[16], as per the
relationship between recommendation algorithm and
knowledge graph embedding there are majorly two ways
for doing these tasks:-independently learning and jointly
learning.
6. LITERATURE REVIEW
Table : Comparative Study of Literature Survey
Author Pros Cons Methodology
Used
Haithem
Mezni [1]
Able to give
multi-
relational
representati-
on of
contextual
data related
to users and
services,
which the
previous
approaches
like matrix
and tensor
based failed
to give.
1) Their
model does
not work for
the
uncertain
factors like
incomplete
context
information,
missing
reviews or
feedback.
2) Their
system’s
average
accuracy is
almost half
i.e 50.52
1) For finding
similarity
between the
context of two
or more users
and services
used concept of
subgraph-aware
proximity.
2) Content and
collaborative
filtering
percent
Sihang Hu
[2]
Their
recommend-
ation
algorithm
performs
better than
SP rank
algorithm
which is
novel hybrid
recommend-
ation
algorithm to
compute top
N
recommend-
ation.
Their model
mainly
focuses on
point of
interest
which are
obtained
from user
review.But
other
contextual
information
can also
improve
recommend-
ation.
Content based
filtering and
directed graph
Tiantian
He [3]
Their
method
combines
graph
clustering
and
multiview
learning so
as to
perform the
task of
clustering in
the
multiview
featured
graph
Their model
is based on
unsupervise
-d learning
so it does
not have
support for
vertex
embedding
for
attributed
graph as it is
used in
supervised
learning
model
Attributed
graphs and
contextual
correlation
which preserves
multiview
features based
graph clustering
Cairong
Yan [4]
1) Their
method
provides
different
recommend-
ations for
active and
inactive
users
respectively.
2) Their
paper claims
to solve the
cold start
problem and
sparsity
problem
which was
problem in
previous
recommend-
ation
systems.
1) More
parameters
need to be
considered
for
recommend-
ation other
than
userid,itemi
d, itemtags ,
season,st-
yle,gender
and
customer
score for
improving
recommend-
ation.
2) Approach
is based on
knowledge
graph but
does not go
in detail
between the
entities
Collaborative
filtering,data
augmentation
for improving
quality of
data,factorizatio
-n machine
model for high
accuracy of
recommendatio-
n,knowledge
graph

Xiaoliang
Fan [5]
1) Their
approach
fully utilizes
the concept
of spatial
and
temporal
correlation
which
previous
methods
failed to do.
2) They have
used
weighted
rating factor
to improve
the
effectiveness
of
recommend-
ation
1) Other
contextual
information
like social
context etc.
of user is not
considered.
2)
Performan-
ce of the
system can
be improved
by using
knowledge
graph
Temporal and
spatial
correlation,
weighted rating
effect for
similarity
computation.
Hao Wang
[6]
Their
proposed
clustering
algorithm
converges
fast.
1) Time
complexity
of their
algorithm is
square.
2) For large
amount of
data their
clustering
algorithm
works
slower
Multiview
clustering
Weizhi Ma
[7]
Their model
performs
good in case
of noisy item
knowledge
graph that is
created by
linking item
names to
related
entities
Creating
rules for
large
number of
items can
make the
system
complex and
can affect
recommen-
dation
Jointly learning
rules,random
walk,
collaborative
filtering and
knowledge
graph
Zhiheng
Wu [8]
Used the
concept of
user
reputations
for
validating
the
reputation of
item or
service
Their
method
gives more
reputation
value to real
user and
less value to
malicious
user.So, if
this value is
same for
both types
of users then
their system
fails to
distinguish
User reputation
calculation,
collaborative
filtering and
context
different
users.
Lan Zhang
[9]
Considered
user-event
interactions,
event-event
interactions
for better
recommend-
ation
For complex
events their
model may
work
slower.
User behavior
analysis and
knowledge
graph
Zhixue
Jiang [14]
Created
efficient
question
answering
system using
concept like
information
extraction
and
knowledge
fusion
1)
Construction
and retrieval
speed of
knowledge
graph is not
efficient. 2)
Multi round
dialogue and
complex
knowledge
reasoning
need to be
improved.
Natural
language
processing,
knowledge
graph, Rule
based matching
techniques and
string matching
algorithms are
used for
classifying and
querying the
questions.
7. PROPOSED SYSTEM
7.1 Problem Statement
“To implement the recommendation system using context
aware services, Dilated RNN based on the knowledge
graph using machine learning”
7.2 Problem Elaboration
Nowadays recommendation systems are used in many
domains and the challenge is to provide better
recommendation to the user when we have large amount
of data. So, We will be building recommendation system
using knowledge graph as it have many advantages over
traditional approaches. Also,we will be using machine
learning for improving the performance of knowledge
graph based recommendation system.
7.3 Proposed Methodology
Dilated RNN(Recurrent Neural Networks) is the latest
embedding algorithm that can be used in the knowledge
graph for improving the performance of the system. We
will be comparing performance of three algorithms
Dilated RNN(Recurrent Neural Networks),Dilated
CNN(Convolutional Neural Networks) and LSTM(Long
Short Term Memory).Out of these algorithms Dilated RNN
gives better accuracy as compared to others. So, Machine
learning algorithms play an important role in improving
the performance of the system. We have used Yelp dataset
for five domains (Movie, Travel, Health, Shopping and
Restaurant) for creating recommendation systems. We

will do pre-processing of the dataset. Then training and
testing of model is done by using three algorithm. And
then classification of dataset is done by each of three
algorithm. After that recommendation prediction part is
implemented and then we are comparing the accuracy of
three algorithms and the best algorithm is selected out of
three algorithm and then we will be creating knowledge
graph for each of the above mentioned domains.
Fig 7.3 Workflow Diagram
1. Data Collection
We will collect data from Yelp dataset for five domains
(Movie, Travel, Health, Shopping and Restaurant) for
building recommendation systems. Each domain dataset
will have relevant parameters required for it.
2. Training
Before starting the training of model we will first do pre-
processing of the dataset. For training the model we will
then split the dataset into training set and testing set. We
will be using Dilated RNN, Dilated CNN and LSTM. Dilated
RNN is also used here to convert the knowledge graph into
low dimensional vector space and also for reducing the
processing complexity of knowledge graph.
3. Testing
We will be testing our model on testing data and then we
will be evaluating the performance of the system using
certain metrics like accuracy, MRR(Mean Reciprocal
Rank),Recall,Precision,F1 measure etc.
8. FUTURE SCOPE
In this section we will discuss some of the famous research
approaches related to the recommendation system.
a) Dynamic Recommendation: We have seen that some
knowledge graph based recommendation system with
GNN or GCN architecture shows good performance but
the training process is quite time consuming. The
recommendation system using such models are called as
static preference recommendation. However, nowadays
recommendation system should also be deal with real
time interests because nowadays user’s preference can
get affected by social media or friends also. So, in that
scenario static preference model would not be
sufficient. However for the purpose of capturing
frequently changing user preference dynamic graph
network can be used. The paper proposed by W.song
[26] used this dynamic graph network for including long
and short term interests from friends. Along with this,
other side information can be used and knowledge
graph for dynamic recommendation can be prepared.
b) Multi-task Learning: One of the most important task
in creation of knowledge graph for recommendation
purpose is the link prediction in the graph. So, there is a
scope to improve the performance of graph based
recommendation system. Some of the indirect links
may get ignored because of user preference fact is
missing which can ultimately impact the
recommendation results. However, paper proposed by
W.Cao [27] showed that it is effective to train the
knowledge graph completion and recommendation
module together for giving better recommendation.
There are some other papers which have used multitask
learning by training together task of recommendation
module with knowledge graph embedding and item
relation regulation task.
c) Cross domain recommendation: For dealing with the
problem of long standing data sparsity in the
recommendation system this cross domain
recommendation is used. In this approach, we fully
utilize the richer information from a richer domain for
improving recommendation performance in a less richer

domain. The paper proposed by Zhang[28],created a
matrix based method for cross domain recommendation.
Also in some of the approaches user item graph contains
only interaction relation and it does not consider any
other relationships between users and items. For
improving the cross domain recommendation we can
include different types of users and items information in
the graph.
d) Knowledge based enhanced language representation:
For the purpose of improving the performance of
various natural language processing tasks, nowadays
external knowledge is merged into the language
representation model. So, language and text
representation can be improved mutually. The paper
proposed by Chen[29] created a model for short text
classification and it utilizes previous knowledge from
knowledge graphs for enhancing the semantic
representation of the short texts.
9. CONCLUSION
In this survey paper, we have studied various methods
used for filtering in recommendation system and also
studied various traditional and recent knowledge graph
based recommendation approaches. We have also seen
approaches in which knowledge graph is used as additional
information for giving better recommendations to the user.
We observed that knowledge graph based
recommendation are good for better and explainable
recommendation.
In this paper, We have also introduced a method using
knowledge and machine learning which we are going to
experiment on existing knowledge graph based
recommendation systems. After that, we will get the model
having more average accuracy as compared to earlier
model which will be able to give better recommended
items or services to the user. With the help of this proposed
system ,we will try to create more efficient and accurate
model for the recommendation systems.
We hope that through this survey paper we have helped
the readers to understand various works done in the
recommendation field.
REFERENCES
[1] Haithem Mezni,Djamal Benslimane, and Ladjel
Bellatreche,“Context-aware Service Recommendation
based on Knowledge Graph Embedding”,2021 IEEE
Transactions on Knowledge and Data Engineering ,
DOI: 10.1109/TKDE.2021.3059506
[2] Sihang Hu, Zhiying Tu, Zhongjie Wang,“A POI-
Sensitive Knowledge Graph based Service
Recommendation Method”,2019 IEEE International
Conference on Services Computing (SCC), DOI:
10.1109/SCC.2019.00041
[3] Tiantian He ,Yang Liu,Tobey H. Ko,Keith C. C. Chan
and YewSoon Ong,“Contextual Correlation Preserving
Multiview Featured Graph Clustering”,2019 IEEE
Transaction on Cybernetics, DOI: 10.1109/
TCYB.2019.2926431
[4] Cairong Yan,Yizhou Chen,“Differentiated Fashion
Recommendation Using Knowledge Graph and Data
Augmentation”,2017 IEEE Access, DOI:
10.1109/ACCESS.2019.2928848
[5] Xiaoliang Fan, Yakun Hu, Zibin Zheng,Yujie Wang,
Wenbo Chen and Patrick Brezillon,“CASR-TSE:
Context-aware Web Services Recommendation for
Modeling Weighted Temporal-Spatial
Effectiveness”,2017 IEEE Transactions on Services
Computing , DOI: 10.1109/TSC.2017.2782793
[6] Hao Wang, Yan Yang, Bing Liu,“GMC: Graph-based
Multi-view Clustering”,2018 IEEE Transactions on
knowledge and data engineering , DOI:
10.1109/TKDE.2019.2903810
[7] Min Zhang,Weizhi Ma, Yue Cao , Woojeong Jin ,
Chenyang Wang , Xiang Ren,Yiqun Liu , Shaoping
Ma,“Jointly Learning Explainable Rules for
Recommendation with Knowledge Graph”,2019
ACM(Association for Computing Machinery) , DOI:
10.1145/3308558.3313607
[8] Zhiheng Wu,Jinglin Li,Qibo Sun,Ao Zhou,“Service
recommendation with context-aware user reputation
evaluation”,2017IEEE 7th International Symposium on
Cloud and Service Computing(SC2) , DOI:
10.1109/SC2.2017.47
[9] Lan Zhang,a Xiang Li,b Weihua Li,a Huali Zhou,a
Quan Baib,“Context-Aware Recommendation System
using Graph based Behaviours Analysis”,2021 Journal
of Systems Science and Systems Engineering, DOI:
10.1007/s11518-021-5499-z
[10] Hoang Long Nguyen1 , Dang Thinh Vu1 , Jason J.
Jung, “Knowledge graph fusion for smart systems: A
Survey”,2020 ScienceDirect, DOI: 10.1016/ j.inffus.
2020.03.014
[11] Chongyan Chen, Islam Akef Ebeid, Yi Bu, and Ying
Ding,“Coronavirus Knowledge Graph: A Case
Study”,2020 arXiv, DOI: 2007.10287
[12] Shuai Qianjun,Cheng Zhang, “Question Answering
System based on Knowledge Graph of Film
Culture”,2020 ICCST(International Conference on
Culture-Oriented Science and Technology), DOI:
10.1109/ICCST50977.2020.00035

[13] Yan Jia, Yulu Qi, Huaijun Shang, Rong Jiang, Aiping
Li,“A Practical Approach to Constructing a Knowledge
Graph for Cybersecurity”,2018 ScienceDirect, DOI:
10.1016/j.eng.2018.01.004
[14] Zhixue jiang,Chengying chi,Yunyun zhan,“Research on
Medical Question Answering System Based on
Knowledge Graph”,2021 IEEE Access, DOI:
10.1109/ACCESS.2021.3055371
[15] Addi AitMlouk , Lili Jiang,“KBot: a Knowledge graph
based chatBot for natural language understanding
over linked data”,2020 IEEE Access ,DOI:
10.1109/ACCESS.2020.3016142
[16] C. Liu, L. Li, X. Yao, and L. Tang,“A survey of
recommendation algorithms based on knowledge
graph embedding”, 2019 MDPI, DOI:
10.3390/info12060232
[17] Z. Sun, Q.Guo, J. Yang, H. Fang, G. Guo, J. Zhang, and R.
Burke, “Research commentary on recommendations
with side information: A survey and research
directions,”,2019 ACM(Association for Computing
Machinery), DOI: 10.1016/j.elerap.2019.100879
[18] S. Zhang, L. Yao, A. Sun, and Y. Tay, “Deep learning
based recommender system: A survey and new
perspectives”,2019 ACM Computing Surveys, DOI:
10.48550/arXiv.1707.07435
[19] H. Cai, V. W. Zheng,K. C.-C. Chang, “A comprehensive
survey of graph embedding: Problems,techniques, and
applications”,”2018 IEEE Transactions paper on
Knowledge and Data Engineering, DOI:
10.48550/arXiv.1709.07604
[20] Q. Wang, Z. Mao, B. Wang, and L. Guo, “Knowledge
graph embedding: A survey of approaches and
applications”,2017 IEEE Transactions on Knowledge
and Data Engineering, DOI
:10.1109/TKDE.2017.2754499
[21] S. Ji, S. Pan, E. Cambria, P. Marttinen, and P. S. Yu,“A
survey on knowledge graphs: Representation,
acquisition and applications”, 2020 arXiv, DOI
:10.48550/arXiv.2002.00388
[22] Zhang, Y.; Wang, J.; Luo, J.,”Knowledge Graph
Embedding Based Collaborative Filtering”,2020 IEEE
Access, DOI: 10.1109/ACCESS.2020.3011105
[23] Chan Liu,Lun Li,Xiaolu Yao,Lin Tang, “A Survey of
Recommendation Algorithms Based on Knowledge
Graph Embedding”,2019 IEEE Internation Conference
on Computer Science and Educational Informatization,
DOI: 10.1109/CSEI47661.2019.8938875
[24] Xiaohan Zou,”A Survey on Application of Knowledge
Graph”,2020 4th International Conference on Control
Engineering and Artificial Intelligence, DOI:
10.1088/1742-6596/1487/1/012016
[25] Ayesha Ameen,”Knowledge based Recommendation
System in Semantic Web - A Survey”,2019
International Journal of Computer Applications, DOI:
10.5120/ijca2019918538
[26] Weiping Song, Zhiping Xiao, Yifan Wang,”Session-
based social recommendation via dynamic graph
attention Networks,2019 12th ACM International
Conference , DOI:10.1145/3289600.3290989
[27] Yixin Cao,Xiang Wang,Xiangnan He,” Unifying
Knowledge graph learning and
recommendation:Towards a better understanding of
user preferences”,2019 ACM, DOI : 10.1145/
3308558.3313705
[28] Qian Zhang,Jie Lu,Guangquan Zhang,”A cross-domain
recommender system with kernel-induced knowledge
transfer for overlapping entities”,2018 IEEE
transactions on neural networks and learning
systems, DOI:10.1109/TNNLS.2018.2875144
[29] Jindong Chen,Yizhou Hu,Jingping Liu,Yanghua
Xiao,”Deep Short text classification with knowledge
powered attention”,2019 Proceedings of the AAAI
Conference on Artificial Intelligence , DOI:
10.1609/aaai.v33i01.33016252
[30] Qingyu Guo, Fuzhen Zhuang, Chuan Qin, Hengshu
Zhu, Xing Xie, Hui Xiong, Qing He,”A survey on
Knowledge graph based Recommender Systems”,2020
arXiv, DOI: 10.48550/arXiv.2003.00911
[31] Khalid Haruna,Maizatul Akmar Ismail, Suhendroyono
Suhendroyono,Damiasih Damiasih,Adi Cilik
Pierewan,Haruna Chiroma,Tutut Herawan, “”Context-
Aware Recommender System: A Review of Recent
Developmental Process and Future Research
Direction”,2017, DOI: 10.3390/app7121211

BIOGRAPHIES
Shailesh D.Kalkar
MTech Computer Engineering,
VJTI Mumbai
University. She has 28 years of teaching experience
and has guided 80+ M. Tech. projects and 100+ B.
Tech. projects. She has published 134 papers in
International Journals, 20 papers in National
/International Conferences/Symposiums. She has
worked as an Organizing Committee member for 21
International Conferences and 5 AICTE/MHRD
sponsored Workshops/STTPs/FDPs. She has
participated in 14 National/International Conferences.
She has worked as NBA Coordinator of the Computer
Engineering Department of VJTI for 5 years. She had
written a proposal under TEQIP-I in June 2004 for
‘Creating Central Computing Facility at VJTI’. Rs. Eight
Crore were sanctioned by the World Bank under TEQIP-
I on this proposal. Central Computing Facility was set up at
VJTI through this fund which has played a key role in
improving the teaching learning process at VJTI.
Prof. Pramila M. Chawan is
working as an Associate
Professor in the Computer
Engineering Department of VJTI,
Mumbai. She has done her
B.E.(Computer Engineering) and
M.E.(Computer Engineering)
from VJTICollege of Engineering,
Mumbai

A Survey on Recommendation System based on Knowledge Graph and Machine Learning

Recommended

More Related Content

Similar to A Survey on Recommendation System based on Knowledge Graph and Machine Learning (20)

More from IRJET Journal (20)

Recently uploaded (20)

A Survey on Recommendation System based on Knowledge Graph and Machine Learning