Volume 8, Issue 3, March – 2023 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Message Passing-Based Prediction of Unlabelled

Node Embedding Using Graph Neural Network
1 2
Girish L Dept of CSE SIET, Raviprakash M L Dept of CSE KIT,
Tumkur, Karnataka, India Tiptur, Karnataka, India

Abstract:- Graph neural network are a part of deep  An Approach Used For Node Classification Are: The
learning methods created to perform presumption on node embedding gives us the
data described by graphs. Graph neural network is a  The data about the unlabelled node
neutral network that can straight away be applied to  The data about the neighbouring node
graphs. It provides a agreeable way for node level,  And the citation between them.
edge level and graph level prediction tasks. Moreover,
most GNN models do not account for long distance By this directive we can predict the unlabelled
relationships in graphs and instead simply aggregate nodes. So in order to get the node embeddings, we pass the
data from short distances (e.g., 1-hop neighbours) in node information through the message passage technique is
each round. In this paper work, we carry out node used. Node classification is one of the techniques in graph
classification using graphs which can be put into large neural network.
graphs comprise of labelled and unlabelled nodes. Here
we can predict the node embeddings of the unlabelled Graph Neural Networks (GNNs) are a class of deep
node by using an approach called message passing. For learning
executing this, we took Cora dataset, provided a overview
of it, builded our graph neural network, splitted the data
to test and train data, trained it and finally visualised
the output.

I. INTRODUCTION

Node classification is applied to large graphs include

la- belled and unlabelled nodes.So the assignment is to
predict the node embeddings for the unlabelled node based
on the neighbouring nodes details and the link between
them [1].

Graph Neural Network can carry out node Fig 1 Node Classification
classification, link prediction, graph classification
Implementation of node classification are Citation Methods generated to execute inference on data
networks/citation graph is a directed graph that represent the demonstrated by graphs. Graph Neural Network is a
citations reddit posts-internet community where people can neural network that can straight away can be put in to
post contents on numerous topics like politics,cooking etc. graphs [2]. It come up witha satisfactory way for node level,
youtube videos etc. Link edge-level prediction - approximate edge level, and graph level prediction tasks. Graph
the probability of links among nodes in a graph. Graph-level consisting of nodes, edges and neural network having
prediction - predicts single value for the entire graph. distinct layers like input layer , hidden layer, and output
layer together how to feed the graph to neural network is the
Problems faced in node classification is predicting the GNN, which is shown in Fig 2 and the area which deals with
node embeddings i.e the task here is to determine the the study of interaction of graph and neural network is
unlabelled nodes by involving the labels of their called geometric deep learning. The intuition of GNN
neighbours. Generally, problems of this type are trained in a their neighbors and connections are described naturally.
semi-supervised way (for this we import planetoid which is
a graph constructed semisupervised learning method, by
incorporating labelled and unlabelled data through the
training )with only a part of the graph being labeled.

IJISRT23MAR095 www.ijisrt.com 136

Volume 8, Issue 3, March – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig 2 Graph Neural Network

To understand this consider an example that if we

remove the connections and neighbors throughout the node,
then the node will drop all its information. Therefore, the Fig 3 Cora Data Set
neighbors of a node and connections to neighbors explain
the concept of the node. The outcome of the work is that for each of the node,
we will obtain a prediction. We will use the class with
 There are two types of GNNs that are mostly dominant: highest probability and this will be our predicted class.
Graph Convolutional Network (GCN) and Graph Auto-
Encoder Network. Let us understand the two below: In the upcoming chapters we can get familiarised with
related work of the node classification and gist of the related
 Graph Convolutional Network: papers, system architecture of our model finally the
(GCNs)[3] refer to the procedure of applying a experimental details including results.
spatially moving filter on the nodes of a graph that
incorperate embeddings or data relevant to each node to II. RELATED WORK
get a characteristic representation of each node. To
incorperate input from wider neighborhoods, various In this chapter, we look upon machine learning, types
convolutional layers, related to a standard CNN, can be of machine learning, graph neural network, applications of
stacked. GNN used in node classification as GCN (Graph graph neural network, node classification and related
Convulation Network)layer is the layer which carry out this papers.
message passing. We use 2 message passing layers in cora
dataset To analyse node embeddings we use message A. Machine learning:
passing layers. GCNs execute similar operations where the Machine learning is one of the part of artificial
model grasps the features by inspecting neighboring nodes. Intelligence which can detect an output by combining data
with statistical tools.[8] A machine learns from the
 Graph Auto-Encoder Network: previous data to give extact results. Machine learning is
Auto-encoders are neural networks that integrate two closely belongs to data mining and Bayesian predictive
modeling. The machine will receive data as input and it
networks:[3] an encoder that downsamples the input by
transferring it through convolutional filters to provide a will use an algorithm to formulate answers. A typical
machine learning tasks are to give a recommendation.
compact characteristic representation of the image, and a
decoder that takes the encoder’s interpretation as input and For those who have a Netflix account, all suggestions of
tries to reconstruct the input based on it. Graph auto- movies or series are based on the user’s historical data.
encoders attempt to comprehend a visual illustration of the Technical companies are using unsupervised learning to
graph and then re-construct the graph using the decoder. improve the user experience with personalizing
They may be used to learn graph embeddings and hence recommendation.
anticipate embeddings for unseen nodes and categorize
newer nodes into existing categories inside the network. Simulated intelligence is the frontal cortex where all
the learning steps occurs. The way the machine learns is
 Dataset In Node Classification: fun- damentally as old as individual. Individuals gain from
The Cora Dataset [6] consists of 2708 scientific their experience. The more we know, the more viably we
can an- ticipate. By this assumption, when we face any dark
publications. The dictionary consists of 1433 unique
situation, the likelihood of accomplishment is lower than the
words, Fig 3 is the cora dataset representing 7 labels
Nodes are the Publications which are papers, Books, etc. known situation. Machines are also pre-arranged something
Edges are the citations. Node Features are the word almost identical. To make an exact assumption, the machine
vectors.7 Labels in this dataset are publication type e.g. sees a model. Right when we give the machine a comparable
Neural Networks, Rule Learning, Reinforcement Learning, model, it will in general be figure out the yield. Regardless,
Probabilistic Methods. like a human, if its feed an in the past unnoticeable model,
the machine encounters issues to expect the result. The

IJISRT23MAR095 www.ijisrt.com 137

Volume 8, Issue 3, March – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
essential objective of AI is the learning stage and derivation collection. We know the sex of all of our individual, it must
stage. Above all, the machine will learn through the be male or female. The necessary focus of the classifier
revelation ofthe models. This divulgence is made us to offer will be to give a probability of being a male or a female
significant thanks to the data. One fundamental piece of the (i.e., the imprint) considering the information (i.e., features
data scientist is to pick mindfully which data to provide we have collected). We can utilize the new information to
for the machine. Fig 4 shows the learning stage and The make a forwardness.
machine uses some excessive estimations to chip away at
reality and change this exposure into a model. Subsequently, For instance, we have actually got new data from a dull
the learning stage is used to depict the data and wrapped up individual, and we need to know whether it is a male or
it into a model Inferring is female. Acknowledge the classifier predicts male = 70
percent, it guesses that the calculation ensures at 70 percent
that this individual is a male, and 30 percent is a female.The
name can in like way be of no under two classes. The above
Machine learning model has just two classes, yet expecting
a classifier needs to anticipate object, it has so many of
classes (e.g., glass,pen, table, shoes, and so on everything
watches out for a class).

Fig 4 Learning Phase  Regression:

When the yield is rigid worth, the errand is a
another state in machine learning where the model is built, relapse. For instance, a monetary examiner might have to
it is possible to test how heavy it is on never-seen-before predict the worth of a stock dependent on a scope of
data. The new data are changed into a features vector, go component like value, past stock exhibitions,
through the model and give a assumption. This is all the macroeconomics file. The framework will be prepared to
colorful part of machine learning. There is no need to update measure the cost of the stocks with the most in-
the rules or train again the model. we can use the model conceivable blunder.
previously trainedto make inference on new data.
 Unsupervised learning :
In this unsupervised learning, an algorithm surveys
input data without being given an explicit output variable
(e.g., explores customer demographic data to predict the
patterns) we can use it when you do not know how to
categorize the given data, and we want the algorithm to find
out the patterns and classify the data.

Fig 5 Inference from the Model C. Graph Neural Networks:

Graph neural network (GNN)[9] is a class of neural
Fig 5 shows that there is no need to update the rules or net- works for processing data represented by graph data
train again the model. You can use the model previously structures. They were well known by their use in supervised
trained to make inference on new data. learning on properties of various molecules. Charts or the
graphs are atany place around us. Our relational association
B. Types of machine learning: is a model for outline of people and relations. The roads you
Machine learning can be classified into two kinds of take to go from guide c toward point d build up a chart. The
learn- ing tasks: Supervised and Unsupervised. There are associations that partner this site page to others structure a
many otheralgorithms. graph. Right when your manager pays you, your portion
goes through an outline of money related foundations.
 Supervised learning: Basically, whatever is solidified of associated substances
A calculation utilizes preparing information and can be called as a graph. Graphs are very incredible
evaluation from people to become familiar with the inventions to visualize relationship between people, items,
association of given contributions to a given yield. For and thoughts. Also past envisioning information, not with
instance, an individual can utilize promoting cost and standing, diagram can moreover be for the most part
climate figure as information to figure the deals of jars. We brilliant wellspring of data to plan AI models for absurd
can utilize directed realizing when the yield information is undertakers. Graph neural networks (GNN) are a type of
known. This calculation will going to envision new machine learning algorithm that can extract essential
information. There are two types of supervised learning: informa- tion from graphs and make useful assumptions.
With graphs becoming more prevalent and richer with
 Classification: information, and artificial neural networks becoming more
Imagine we really want to expect the gender popular and capable, Graph neural network have become a
orientation of a person for a business. We will start by powerful tool for many important applications.
assem- bling the data on the stature, weight, work, pay,
purchasing holder, etc from our singular informational

IJISRT23MAR095 www.ijisrt.com 138

Volume 8, Issue 3, March – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Types of GNN:  Graph Clustering:
Two types of GNNs are Mostly Dominant: It refers to data clustering in the form of graphs. On
graph data, there are two unique types of clustering. Vertex
Graph Convolutional Network (GCN) and Graph clustering tries in organizing the network nodes into clusters
Auto- Encoder Network (GAEN) Let us understand them of highly linked areas based on edge weights or edge
in detail: distances. The second type of graph clustering considers
graphs to be the objects to be grouped and groups them
 Graph Convolutional Network: based on similarity.
GCNs calls to the process of applying a spatially
moving filter on the nodes of a graph that include  Graph Visualization:
embeddings or data related to each node to get a feature It is a part of math and software engineering that joins
representation of each node. To include input from wider mathematical chart hypothesis with data representation. It is
neighborhoods,various convolutional layers, similar to a centered around the visual portrayal of charts, which shows
standard CNN, can be put together. constructions and irregularities in the data and helps the
client in understanding the diagrams. There are many
 Graph Auto-Encoder Network: applications in GNN in which we have implemented Node
Auto-encoders are neural networks that integrate two classification in this work.
networks:an encoder that down- samples the contribution by
moving it through convolutional channels to give a smaller  Node Classification:
element portrayal of the picture, and a decoder that accepts Node Classification is a very com- mon machine
the encoder’s translation as information and attempts to learning task to try in graphs: training a model to learn in
recreate the information dependent on it.Chart auto-encoders which class a node belongs. GDS will train the supervised
aim to appreciate a visual layout of the diagram and machine learning models based on node properties and
afterward re-build the chart by using the decoder. They characteristics (features) in our graph to forecast what class
might be utilized to learn chart embeddings and accordingly an unknown node or future node would belong to. Node
expect embeddings for concealed hubs and classify fresher Classification can be used likely together with pre-
hubs into existing classifications inside the organization. processing algorithms. Actually, Node Classification models
are utilized to figure the non-existing hub property
 Applications of GNN: dependent on the other hub properties. The non-existing hub
We can see so many real life applications of Graphical properties address the class, and is referenced to as the
Neural Networks like recommender systems, natural objective property. The characterized hub properties are
sciences, posts prediction, etc. The listed below are some of utilized as info highlights.
the applications of GNN:
The preparation cycle follows this means: The info
 Graph Classi- Fication: chart will part into two sections: the train diagram and the
The essential point here is to segment the outline into test chart. The train chart is again separated into various
various social events. It resembles picture request, except conformity overlap, each folds comprises of a train part and
for the genuine developments to the graph space. Outline an approval part. Each model competitor is prepared on
portrayal has a wide extent of vocations, from identifying each train partand determined on the separate approval part.
whether or not a protein is a synthetic in bioinformatics to The preparation interaction utilizes a strategic relapse
requesting articles in NLP or casual local area assessment calculation, and the computation utilizes the predefined
and investigation. measurements. The main measurement is the essential
measurement. The model with the most elevated normal
 Node Classification: score whose as indicated by the essential measurement will
The primary point here is to partition the diagram into win the preparation. The triumphant model will be
different gatherings. It’s like picture order, with the quirk of then be retrained on the whole train diagram. The
the objective movements to the diagram space. Diagram triumphant model is assessed on the train chart just as
characterization has a wide scope of employ- ments, from the test diagram.
recognizing whether or not a protein is a chemical in
bioinformatics to ordering articles in NLP or informal The triumphant model is retrained on the whole unique
community examination and exploration. chart. What’s more, At last, the triumphant model will be
enlisted in the Model Catalog. Prepared models may then be
 Link Prediction: utilized to anticipate the worth of the designated property
In this case, the algorithm must comprehend the link that is of before hand inconspicuous and obscure hubs. Not
between entities in graphs and attempt to forecast if two with standing this expectation of the class for every hub, the
entities are connected. It is critical in social networks to anticipated likelihood for the each class may like wise be
infer social connections or to recommend potential buddies there. The request for the probabilities match the request for
to users. It’s also been used to solve recommender system the classes enrolled in the model.
issues and forecast criminal connections.

IJISRT23MAR095 www.ijisrt.com 139

Volume 8, Issue 3, March – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
D. Gap Analysis: non-structural framework, and other frameworks, then they
In this section, we have referred some papers related to given a detailed review for applications in each scenario.
node classification using GNN and we have provided
overview of it. Inneke Mayachita et al in their work[7], training graph
convolutional networks on node classification task text Clas-
Sakthi Kumar et al in their paper [2], performed node sification approach problem work well if there are enough
classi- fication tasks using spectral and non-linear labeled but unfortunately, in real world cases, labeling data
approaches providing a outline representing the network. might be expensive. They took the dataset-CORA reference
The models which are trained using network-dependent network dataset. Implementation of Graph Convolution Net-
node features may not theorize to other networks.And the works (GCN) by making use of Spektral API, which is
methods like Graph At- tention Networks (GAT) are not a Python library for diagram profound learning dependent on
able to consider higher-order neighborhood information Tensorflow. They performed Semi-Supervised Node
without rising the number of neural network layers. To Classifi- cation using CORA dataset. In the above parts we
overcome this restriction, they have used kernel propagation have gone over CORA reference network dataset.
graph neural network (KP-GNN). It has a pre-handling step Conclusion from their work is that the conventional machine
by which one can the get the high-request node information. learning procedure to perform document classification, for
This has a technique called Hadamard- product attention instance CORA dataset, is to use supervised text
mechanism which reduces the number of hidden parameters classification approach. Graph Con- volutional Networks
by half therefore balancing the losses during training and (GCNs) is an alternative semi-supervised procedure to solve
validation. This allowed GNNs not to be dependent of node this problem by seeing the documents as a network of
features when they are unavailable/missing and while related papers. Utilizing just 20 marked nodes for each class,
available it increases the performance of GNNs because in GCNs beat Fully-Connected Neural Networks in this paper.
this model node features can be replaced with higher-order
structural features without loss of classification accuracy. Ziwei Zhang et alin their survey[4] on the topic deep
Cora, citeseer and PubMed networks in conjunction with learning on graphs reviewed the various types of deep
OGB networks are utilized in their work and hyper learning methods on graphs. They gave outlines of the
boundaries, for example, dropout rate and weight rot are techniques such as graph recurrent neural networks, graph
setted. convolutional networks, graph autoencoders, graph
reinforcement learning, and graph adversarial
 After Preparing and Approval they Got the methods.Graph RNNs which can cap- ture looping and
Accompanying Outcomes: subsequent patterns of graphs. They are seperated into two
 Attention mechanism used helped them prevent model categories- node-level RNNs and graph- level RNNs.Graph
fitting over (training loss is remarkably lower than RNN can also be combined with other constructions ,such as
validation loss). By utilizing AKP-GCN they summed GCNs or GAEs.By using GCN , useful node features can be
up hole among approval and testing correctnesses is learned to solve many node-focused tasks and node
minor. information needs to be combined to perform graph-level
 In Cora and PubMed network, structural features when tasks.The GAE along with its variations have been widely
combined with node features can improve node applied in unsupervised learning tasks and are suitable for
classification. learning node representations for graphs.Reinforcement
 This approach on ogb-things organization, which is a learning (RL) is known to be good and acceptable at
business association can additionally foster the node learning from feedbacks.DeepPath and MINERVA which
portrayal accuracy of GCN by generally 3.5 percent. are the rein- forcement learning method are both adopted RL
for knowledge graph (KG) reasoning.They found that
Jie Zhou Ganqu Cui et al in their paper [5] worked DeepPath targeted at finding the most information by giving
on the four open problems indicating the major challenges path between two target nodes and MINERVA found the
and future research directions of graph neural networks, right answer node given a question node and a relation.So by
including robustness, interpretability, pretraining and this survey they showed that deep learning on graphs is a
complex structure modeling. They have taken dataset- promising and booming research field that brings exciting
ImageNet and it contains 3000 to 20,000 nodes. In the and challenging oppurtunities.
new occasions, graph neural networks have become
amazing and viable instruments for AI assignments in Keyulu Xu et al in their survey[10] on the topic how
diagram zone. This progress owes to advances in expressive powerful are graph neural network have explained the graph
power, model elasticity, and training algorithms. In this neural networks and even they have showed how Graph
survey, they conducted a extensive review of graph neural Neural Net- works (GNNs) are an effective framework for
networks. For GNN models, they have introduced its representating learning of graphs. At the point when we
alternatives categorized by computation modules, graph investigate the over all survey of this paper , we can
types, and training types. Moreover, they also summarized find how GNNs follow a neighborhood aggregation
several general frameworks and introduced several scheme, where the representa- tion vector of a node is
theoretical analysis. In terms of application cytology, they computed by recursively combining and transforming
divided the GNN applications into structural framework, representation vectors of its neighboring nodes.Numerous
GNN variations have been proposed and have accomplished

IJISRT23MAR095 www.ijisrt.com 140

Volume 8, Issue 3, March – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
progressed results on both node and graph classification The below fig 6 represents the flow of the model.
techniques.

However, despite GNNs transforming graph

representation learning, there is limited understanding of
their representa- tional properties and limitations. They
presented a theoreti- cal framework for analyzing the strong
power of GNNs to capture different graph structures.Their
results characterized the inequitable power of popular GNN
variants, such as Graph Convolutional Networks and
Graph SAGE, and showed that they cannot learn to
distinguish certain simple graph structures.Then they
developed a effortless architecture that is likely the most
expressive among the class of GNNs and is equally as
powerful as the Weisfeiler Lehman graph isomor- phism
test.Then, at that point, they tentatively approve their
hypothetical discoveries on number of graph classification
benchmarks, and exhibit that their model acheived state-of-
the-art(advanced) performance.

III. SYSTEM MODEL

 System Architecture:

 Description of each module:

 Data collection: Fig 6 System Methodology
The dataset we put forward for this project has been
taken from Cora. But this data set is in raw format. The data In our model, we normalized the features using torch
set consists of 2708 scientific publications consisting of 7 geometric’s transform functions.
labels or classes .The initial step is to convert raw data into
processed data which is done by normalization process,  Splitting data:
since the raw data collected have multiple attributes but only The train-test split strategy is utilized to appraise the
some of those attributes are needed for the prediction. exhibition of Machine Learning calculations when they are
utilized to make expectations on information not used to
 Data Processing: prepare the model. The strategy includes taking a dataset
Information or data preprocessing in Machine Learning and partitioning it into two subsets.Train information and
is a significant advance that works on the nature of Test informat ion. Tr Data: Used to fit the machine learning
information to advance the extraction of significant model. Test Data: Used to evaluate the machine learning
perception from the information. Information model.The objective is to check the display of the machine
preprocessing in Machine Learning alludes to strategy of learning model on new datar model , the imprints are
getting ready (cleaning and putting together) the crude encoded as numeric worth between 0-6. There are 20 centers
information to make it appropriate for a structure and for each class which contain twofold test, train and
preparing Machine Learning models. In basic words, endorsement masks(they let use in on which center point can
information preprocessing in Machine Learning is an be used for which task). Dropout is simply applied in the
information mining method that changes crude information readiness step, but not such a huge amount for expectations
into a reasonable and decipherable configuration. It is the [21].
initial step denoting the commencement of the inter- action
in machine learning.Typically, certifiable information is  Model Building and Training:
deficient, conflicting, incorrect (contains blunders), and Pytorch Geometric which is a mathematical profound
needs explicit trait esteems/trends.This is the place where learning library is utilized for building diagram network.
information preprocessing enters the situation it assists with Pytorch Geometric permits us to utilize distinctive chart
cleaning, design and sort out the crude information, in neural organization layers inside Py- torch module. Graph
thisway making it ready for Machine Learning models. Convolutional layer(GCN) is the layer which performs
message passing. We have 2 Message Passing Layers since
we have utilized 2 message passing advances and one Linear
yield layer which yields the likelihood of 7 classes.

By applying dropout we ensured no information is lost

when performing prediction. Rectified Linear Archictecture
work applied for better execution ,it yields the information if

IJISRT23MAR095 www.ijisrt.com 141

Volume 8, Issue 3, March – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
the information is positive or it gives zero if the information Figure 7 is the result of the experiment.
is negative. The softmax work is utilized as the enactment
work in the yield layer of neural organization models that
anticipate a multinomial likelihood appropriation.

 Prediction:
The output of the model are 7 probabilities, one for
each class. For each of the node, we can get the prediction.
Here, 95 percent predicted that the node is class 2 and
small prediction that its class 0, 1, 5 and for others the
prediction is 0. The class with highest probability will be
chosen as our predicted class. Here it is class 2.

IV. EXPERIMENTAL DETAILS

 Experimental Setup:
In this section, we briefly explained the libraries
which we have used in our experiment. Torch is an open-
source AI library, a logical figuring framework. Torch nn
help us in mak- ing and preparing of the neural network. Fig7 Result Table
Torch.nn.functional incorporates relu (rectified direct
initiation work for neurons). GCNConv-diagram is the
convolutional layer which performs message passing.
Seaborn is the information representation and exploratory
information analysis. Numpy is the mathe- matical python
library for performing numerical procedure on arrays.
Planetoid is the chart based semi-regulated learning method.
transforms is used for performing normalisation on data.
Initially we install the Pytorch Geometric which is a
geometric deep learning library used to build graph
networks then loading the cora dataset for performing the
node classi- fication.

 Dataset:
Overview of Cora dataset: The number of graphs
are 1, number of features are 1433, number of classes are 7, Fig 8 Loss Visualization in Graph
Data(x=[2708, 1433], edgeindex=[2, 10556], y=[2708],
train- mask is 2708, valmask is 2708, testmask is 2708),
number of nodes are 2708, number of edges are 10556,
number of training nodes are 140, training node label rate
is 0.05 and itis undirected.

V. RESULTS

 Training and evaluation of the model:

Epoch indicates the number of passes of the entire
training set. Cross entrophy loss is a loss function used in
machine learning. The smaller the loss the better the
model. Here, the loss has decreased from 1.9461 to
1.3365.

 Loss visualization in graph:

As the epochs increases, we can observe that the loss Fig 9 Output of the Model
has been decreasing in the fig 8.
 Output of the Model:
 Test accuracy: In our experiment, 95 percent predicted that the node
Using the test function we can get the test accuracy.In is class 2 as the graph shown in Fig9 and small prediction
our experiment,we got 74 percent accuracy i.e. three-forth that its class 0, 1, 5 and for others the prediction is 0. For
of the predicted class is correct. every node we can get prediction in this way. The class with
highest probability will be chosen as our predicted class.
Here it is class 2.

IJISRT23MAR095 www.ijisrt.com 142

Volume 8, Issue 3, March – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Visualizing the Embeddings with TSNE: VI. CONCLUSION
TSNE is an un- supervised, non-linear technique
primarily used for data explo- ration and visualizing high- This paper work shows that using graph neural
dimensional data. It shows how the data is arranged in a networks, we can easily perform node level classification,
high-dimensional space. Here we fit the TSNE with our which tells us to which class the node belongs to. These
node embeddings. We reduce the dimension of node techniques will be useful in real world application like
embeddings to a 2 dimension as we cannot visualize our citation networks, youTube videos where user can get the
embeddings of dimension 16. For every 50 epochs, we recommendations based on the previously viewed content
showed how the node embeddings looks like. TSNE by user. Here, we performed node classification on Cora
Visualization: X-axis: 1st dimension of dimensionality- dataset, processed our model, trained it and got output which
reduced embedding Y-axis: 2nd dimension of indicates that 95 percent predicted that the node belongs to
dimensionality-reduced embedding Points = nodes class 2 and small prediction that it’s class 0, 1, 5 for other
(embeddings) Colors=classes/labels. We colored the class so classes, prediction is 0. The class with highest probability
that all nodes with the same class will have same color. will be chosen as our predicted class. Here, it is class 2.
Here in the Fig 10, we can see initially for epoch 0, Using TSNE, we visualised the node embeddings. Over
the time, GNN improved, we can see presence of clusters,
which allows us to easily predict classes for new nodes.

REFERENCES

[1]. Langston Nashold,Rayan Krishnan, ”using LSTM

and SARIMA Models to Forecast Cluster CPU
Usage”.published:16 Jul 2020.
[2]. Pranav T P, Charan S, Darshan M R, “Devops
Methods for Automa- tion of Server Management
using Ansible ”, International Journal of Advanced
Scientific Innovation, vol. 1, no. 2, pp. 7-13, May
2021.
[3]. Sak, Has¸im, Andrew Senior, and Franc¸oise
Beaufays. ”Long short- term memory based
recurrent neural network architectures for large
Fig 10 Initial Visualization of Node Embeddings vocabulary speech recognition.” arXiv preprint
arXiv:1402.1128 (2014).
we have node embeddings spreaded all over the
place. Over the time, GNN improved, we can see their [4]. Jitendra Kumar,Rimsha Goomer,Ashutosh Kumar
are some Singh,”Long Short Term Memory Recurrent Neural
Network(LSTM-RNN)Based Workload Forecasting
Model For Cloud Datacenters”.Published:7-8 Dec
2017.
[5]. Z. Asha Farhath,B.Arputhamary,Dr.L.Arockiam,”A
Servey On ARIMA Forecasting Using Time Series
Model”.Published:8 Aug 2016.
[6]. KUMAR Manoj,ANAND Madhu,”An Application
Of Time Series ARIMA Forecasting Model For
Predicting Sugarcane Production In India”.
[7]. Ratnadip Adhikari,R.K. Agarwal,”An Introductory
Study On Time Se- ries Modeling And
Forecasting”.DOI :10. 13140 /2.1. 2771.8084.
[8]. Bijeshdhyani,Manish Kumar,Poonam
Verma,Abhishik Jain,”Stock Mar- ket Forecasting
Technique Using ARIMA Model”.Published:6 Mar
Fig 11 Final Visualization of Node Embeddings 2020.
[9]. L. Girish and S. K. N. Rao,”Mathematical tools and
clusters i.e. classes with same node embeddings appear methods for analysis of SDN: A comprehensive
in the same area. So now we have perfect clustering which survey,” 2016 2nd International Conference on
will allow us easily to predict the classes for new data Contemporary Computing and Informatics (IC3I),
points(nodes) as shown in the Fig 11. 2016, pp. 718-724, doi:10.1109 /IC3I.2016.
7918055.
[10]. Anupama K C,Shivakumar B R,Nagaraja
R,”Resource Utilization Pre- diction in Cloud
Computing using Hybrid Model”.Vol 12.No.4,2021.

IJISRT23MAR095 www.ijisrt.com 143

Volume 8, Issue 3, March – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
[11]. Khanum, Salma, and L. Girish. ”Meta heuristic
approach for task scheduling in cloud datacenter for
optimum performance.” International Journal of
Advanced Research in Computer Engineering &
Technology (IJARCET) 4.5 (2015): 2070-2074.
[12]. Girish, L.,Rao, S.K.N. Anomaly detection in cloud
environment using artificial intelligence techniques.
Computing (2021) .https //doi.org/10.1007/s00607-
021-00941-x
[13]. Prajwal S, Siddhartha M, Charan S. (2021).
DDos Detection and Mitigation SDN using support
vector machine. International Journal of Advanced
Scientific Innovation, 1(2), 26-31.
https://doi.org/10.5281/zenodo.4782280
[14]. Girish L, Rao SKN (2020) Quantifying sensitivity
and performance degradation of virtual machines
using machine learning. J Comput Theor Nanosci.
https://doi.org/10.1166/jctn.2020.9019
[15]. Rashmi TV, Prasanna MK, Girish L (2015) Load
balancing as a service in Openstack-Liberty. Int J
Sci Technol Res 4(8):70–73.
[16]. Ravi prakash M L, C S Pillai, Ananda babu J,
“Object Recognition using Lucas-Kanade Technique
and Support Vector Machine Based Classification in
Video Surveillance Systems”, International Journal
of Engineering and Advanced Technology (IJEAT)
ISSN: 2249 – 8958, Volume-9 Issue-1, October
2019.
[17]. Ravi Prakash M L , Chethan Chandra S Basavaraddi
, Ananda Babu J, “Face Recognition Using Hybrid
Neuro Fuzzy Network for Occluded Images”,
International Journal of Science and Research
(IJSR), 2018
[18]. MANOJ KUMAR D P. (2021). A Clustering and
Detection of Liver Disease in Indian Patient Using
Machine Learning Algorithms. SPAST Ab- stracts,
1(01). Retrieved from
https://spast.org/techrep/article/view/838
[19]. Chethan Chandra S Basavaraddi, Raviprakash M L,
Dr. C S Pillai, “Face Recognition from Feed
Forward Neural Network Using Occluded Images for
Automating the Surveillance Using Hybrid Neuro
Fuzzy Network”, International Journal of
Engineering Applied Sciences and Technology,
10.33564/ijeast.2020.v05i02.087, 2020, Vol 5 (2) pp.
508- 519
[20]. Raviprakash M L, C S Pillai, “Histogram
Specification and Kalman Filter with Euclidian
distance Technique for identification of Human in
the video Surveillance systems”, IJRAR March
2019, Volume 6, Issue 1.
[21]. Girish L, and Raviprakash M L. “Data Analytics in
SDN and NFV: Techniques and Challenges”.
International Journal of Advanced Scien- tific
Innovation, vol. 4, no. 8, Dec. 2022,
doi:10.5281/zenodo.7657569.

IJISRT23MAR095 www.ijisrt.com 144