Comparing Machine Learning Algorithms: A Graph Theory 

Approach for Improving Accuracy 

1G. Keerthi        2,*M. Siva Parvathi     3D. Sujatha     4N.V. Muthu Lakshmi 

1,2Department of Applied Mathematics, Sri Padmavati MahilaVisvavidyalayam, Tirupati, Andhra Pradesh, India 
3Institute of Pharmaceutical Technology, Sri Padmavati MahilaVisvavidyalayam, Tirupati, Andhra Pradesh, India 

4Department of Computer Science, Sri Padmavati MahilaVisvavidyalayam, Tirupati, Andhra Pradesh, India 

 

Corresponding Author: M.Siva Parvathi 

Article History: Received 20.09.2024     Revised: 24.10.2024 Accepted: 30.11.2024 

Abstract 

Graph theory provides a robust framework for modelling complex relationships in medical data, 

enhancing classification accuracy through relational learning. Unlike traditional machine 

learning (ML) models that treat data points independently, graph-based approaches support 

structural dependencies to improve feature representation. This study explores the application of 

Graph Attention Networks (GAT), GraphSAGE, Graph Convolutional Network(GCN) and 

Graph Isomorphism Network (GIN) for breast cancer classification, employing k-nearest 

neighbor (KNN) graphs to construct a structured dataset where nodes represent patients and 

edges capture feature similarities. The effectiveness of the graph-based approaches is evaluated 

against traditional ML classifiers, including Decision Trees, Random Forest, LightGBM, and 

XGBoost. Experimental results indicate that GCN, GIN, GAT and GraphSAGE are beat 

conventional methods, with GCN, GAT and GraphSAGE achieved 100% test-accuracy and with 

GIN achieved 99.42%, to confirm the percentage of accuracy, authors conducted extensive 

experiments, including robustness testing by reducing KNN connections, introducing noise, and 

shuffling train-test splits. Results demonstrate that graph-based models are significantly best than 

traditional ML models, and these graph-based models maintain same classification accuracy 

while maintaining stability under robustness tests. The findings confirm that graph-based 

learning provides a scalable, interpretable, and highly accurate alternative for medical 

classification tasks, proving its effectiveness in distinguishing between benign and malignant 

tumors. 

Keywords: Breast Cancer Classification, Graph Neural Networks, Graph Attention Networks, 

GraphSAGE, Graph Convolutional Network, Graph Isomorphism Network, K-Nearest 

Neighbors. 

2020 Mathematics Subject Classification: 05C90, 05C62, 68T07, 90C35 

1. Introduction 

Graph theory is a powerful mathematical approach used to model relationships between data 

points, making it highly effective in medical data analysis. Unlike traditional machine learning 

Communications on Applied Nonlinear Analysis
ISSN: 1074-133X
Vol 31 No. 8s (2024)

https://internationalpubls.com 1059



(ML) models, which treat each data point as independent, graph-based methods capture the 

inherent connections between similar cases, allowing for a more structured and meaningful 

representation of complex datasets. In medical diagnosis, particularly cancer classification, 

patient data is often interrelated, as certain biological markers and genetic features exhibit strong 

dependencies, ignoring these relationships can limit the predictive power of conventional ML 

models. 

Graph-based learning approaches offer a solution by structuring data as a graph, where nodes 

represent patients and edges signify feature-based similarities. This structure enables relational 

learning, allowing models to incorporate both individual attributes and the connections between 

patients with similar characteristics. Unlike conventional ML techniques, which focus solely on 

numerical patterns, graph-based models can leverage the structural dependencies in data, making 

them particularly effective for challenging classification problems, such as distinguishing 

between benign and malignant tumours. This structure enables relational learning, allowing the 

model to incorporate not only individual patient characteristics but also the relationships 

betweensimilar cases. Recent advancements in Graph Neural Networks (GNNs), including 

Graph Convolutional Network (GCN), Graph Isomorphism Network (GIN), Graph Attention 

Networks (GAT), and GraphSAGEhave shown remarkable improvements in classification tasks 

by effectively leveraging these relationships. 

In this study, we investigate the potential of graph-based learning models for breast cancer 

classification, comparing their performance to conventional ML classifiers such as Decision 

Trees, Random Forest, LightGBM, and XGBoost. A graph forK-Nearest Neighbours (KNN) is 

used to establish connections between patients with similar medical profiles, enabling a more 

structured representation of the dataset. To ensure reliability, we conducted robustness testing by 

varying KNN connectivity, introducing feature noise, and shuffling the train-test split. 

Experimental results show that GCN, GIN, GAT and GraphSAGE outperform traditional ML 

models. Even under robustness testing graph-based models maintained same accuracies. These 

findings highlight the potential of graph-based learning in medical diagnosis, demonstrating that 

incorporating structural dependencies enhances model accuracy and interpretability. This 

research underscores the importance of integrating graph theory with ML techniques to develop 

more reliable and efficient diagnostic models for medical applications. 

2. Literature Review 

Graph-based learning has gained significant attention due to its ability to model relationships and 

dependencies between data points. Traditional machine learning models often treat data as 

independent entities, missing crucial structural information. In contrast, graph-based models use 

connectivity patterns, making them highly effective in domains such as social networks, fraud 

detection, and healthcare. These models leverage graph structures to improve classification 

accuracy, link prediction, and representation learning. 

The foundation of Graph Neural Networks (GNNs) was introduced by Scarselli et al. [1], who 

proposed a framework to extend traditional neural networks to graph-structured data. Their study 

introduced a recursive approach where node representations were updated iteratively based on 

Communications on Applied Nonlinear Analysis
ISSN: 1074-133X
Vol 31 No. 8s (2024)

https://internationalpubls.com
1060



neighbouring nodes, allowing for effective learning from complex graph relationships. This 

pioneering work paved the way for more advanced graph-based architectures that have since 

been widely applied in various domains. 

One of the most influential advancements in this field was the introduction of Graph 

Convolutional Networks (GCNs) by Kipf and Welling [2]. Their work extended traditional 

convolutional neural networks (CNNs) to operate on graph-structured data using spectral 

methods. The key innovation in GCNs was their ability to aggregate information from 

neighbouring nodes in a computationally efficient manner. Their study demonstrated that GCNs 

outperform traditional models in semi-supervised classification tasks, particularly in social 

networks and citation graphs. By applying convolutional operations to graph data, they improved 

node classification accuracy and enhanced feature learning from graph structures. 

Building upon the limitations of early GNNs, Graph Isomorphism Networks (GIN) were 

introduced by Xu et al. [3] to address the challenge of distinguishing graph structures that appear 

similar but have different underlying properties. Their study proposed a more powerful 

aggregation function compared to GCNs. 

Another significant contribution came from Velicković et al. [4], who introduced Graph 

Attention Networks (GAT), incorporating self-attention mechanisms into graph-based learning. 

Unlike GCNs and GIN, which treat all neighbouring nodes equally, GAT assigns different 

attention weights to each neighbour, allowing the model to focus on the most relevant 

connections. This innovation led to improved performance in semi-supervised classification 

tasks, as it better captured the varying importance of different nodes. 

In parallel, Hamilton et al. [5] developed GraphSAGE, an inductive learning framework 

designed to generate embeddings for unseen nodes. Unlike GCNs, which require the entire graph 

during training, GraphSAGE learns node representations by sampling and aggregating 

information from neighbouring nodes. This approach is particularly beneficial for dynamic 

graphs, such as evolving social networks and healthcare records, where new nodes frequently 

appear. Their study demonstrated that GraphSAGE generalizes well to unseen data, making it 

ideal for large-scale applications requiring real-time learning. 

Beyond these foundational models, researchers have explored the application of GNNs in 

healthcare and other fields. Zhang et al. [6] applied GNNs to personalized healthcare by 

developing a Heterogeneous Graph Neural Network (HGNN) that model’s relationships between 

medical entities such as patients, diseases, and treatments. Their findings showed that 

incorporating heterogeneous graph structures significantly enhances diagnostic accuracy. 

Similarly, Wang and Wang [7] investigated the use of GNNs in electronic health records, 

demonstrating that structured patient relationships improve early disease detection and risk 

assessment. 

The effectiveness of graph-based learning has also been explored in fake news detection and 

education analytics. Mahmud et al. [8] conducted a comparative study of GNNs versus 

traditional machine learning models for fake news detection, revealing that GNNs outperform 

conventional classifiers by leveraging relationships between articles and sources. Likewise, 

Communications on Applied Nonlinear Analysis
ISSN: 1074-133X
Vol 31 No. 8s (2024)

https://internationalpubls.com
1061



Wang et al. [9] introduced a graph-based ensemble learning method for predicting student 

performance, showing that using graph representations leads to more accurate predictions than 

standalone ML models. 

A broader discussion on graph-based learning was presented by Shaila and Varsha [10], who 

provided a comprehensive review of various graph-based machine learning approaches. Their 

study highlighted how GNNs, including GCN, GIN, GAT, and GraphSAGE, are transforming 

fields such as recommendation systems, fraud detection, and healthcare analytics by capturing 

complex structural dependencies. 

Our study builds upon these previous works by applying GIN, GCN, GAT, and GraphSAGE to 

breast cancer classification, demonstrating that graph-based models outperform traditional 

classifiers. We construct a k-NN-based graph to improve feature learning and classification 

accuracy. Furthermore, we conduct robustness testing, validating that graph-based learning 

remains effective under reduced KNN connections, data noise, and varying train-test splits. 

These results support the potential of graph-based deep learning in medical diagnosis and are 

consistent with previous research. 

3. Methods and Materials 

In this section, the details of the data collection and methodology used were discussed. 

 

3.1 Dataset 

The Kaggle cancer dataset, compiled by Erdem Taha [11], consists of 570 samples and 33 

columns, where one column represents the diagnosis (malignant or benign), and the remaining 

32 numerical features are extracted from image-based measurements of cell nuclei. These 

features include radius, texture, perimeter, area, smoothness, compactness, and other shape-

related metrics. In this dataset, the diagnosis column serves as the dependent variable, while the 

32 numerical features act as independent variables. To prepare the data for machine learning 

classification, the dataset is split into training and testing sets using a 70-30 ratio. This results in 

399 samples (70%) for training and 171 samples (30%) for testing. The splitting process ensures 

a balanced approach to model evaluation, allowing the machine learning algorithm to learn from 

a sufficient number of samples while also being tested on unseen data. This statistical 

arrangement facilitates effective classification and analysis of breast cancer biopsies.  

3.2 Methods  

In this research, the authors compare traditional machine learning classifiers with graph-based 

learning models for cancer classification. Conventional ML models operate on tabular data and 

rely on predefined features, whereas graph-based models capture the structural relationships 

between data points, leveraging graph representations to enhance classification accuracy. By 

integrating graph learning techniques (GCN, GIN, GraphSAGE, and GAT) we constructed a k-

nearest neighbor (KNN) graph where nodes represent patients and edges capture feature 

similarities. This graph structure enables deep learning models to effectively propagate 

information and generate meaningful representations that improve classification outcomes. 

Communications on Applied Nonlinear Analysis
ISSN: 1074-133X
Vol 31 No. 8s (2024)

https://internationalpubls.com 1062



Graph Convolutional Networks (GCNs) work by aggregating information from a node’s 

neighbors to learn meaningful representations. The key mathematical operation in GCNs is based 

on spectral graph theory, where the graph structure is represented using an adjacency matrix (A) 

and a degree matrix (D).The fundamental formula for a single GCN layer is: 

𝐻(𝑙+1) = 𝜎(�̂�−1/2�̂��̂�−1/2𝐻(𝑙)𝑊(𝑙)) 

Where 𝐻(𝑙) is the feature matrix at layer 𝑙, with each row representing a node’s feature,  �̂� = 𝐴 +
𝐼 is the adjacency matrix with self-loops added, �̂� is the degree matrix of �̂�, which helps 

normalize the aggregation, 𝑊(𝑙) is the learnable weight matrix for layer 𝑙, and 𝜎 is an activation 

function. 

Graph Isomorphism Networks (GIN) are designed to effectively capture graph structures by 

applying a powerful aggregation function to node features. The mathematical formulation of 

GIN can be written as: 

ℎ𝑣
(𝑘)

= 𝑀𝐿𝑃(𝑘)((1+∈) . ℎ𝑣
(𝑘−1)

+ ∑ ℎ𝑢
(𝑘−1)

𝑢∈𝑁(𝑣)

 

Where ℎ𝑣
(𝑘)

represents the node embedding at layer 𝑘, 𝑀𝐿𝑃(𝑘)is a multi-layer perceptron that 

transforms the aggregated features,∈is a learnable parameter that adjusts the influence of the 

node's previous state, 𝑁(𝑣)denotes the set of neighboring nodes of node 𝑣 and the sum 

operation(+) ensures that the model captures structural information efficiently. 

GraphSAGE learns node embeddings by aggregating information from a node’s neighbourhood 

using an inductive learning approach. Unlike transductive methods that require the entire graph 

during training, GraphSAGE generalizes to unseen nodes by iteratively updating representations 

based on sampled neighbours. The feature update process follows the equation  

ℎ𝑣
(𝑘)

= 𝜎 (𝑊(𝑘). 𝑓 ({ℎ𝑢
(𝑘−1)

, ∀𝑢 ∈ 𝑁(𝑣)})) 

where ℎ𝑣
(𝑘)

is the feature embedding of node 𝒗 at layer 𝑘, 𝑊(𝑘) is the learnable weight matrix, 𝜎 

is an activation function, and 𝑁(𝑣) denotes the neighbours of node 𝑣. Where aggregation 

function denoted as 𝑓 and it can take different forms, such as mean pooling, Long Short-Term 

Memory (LSTM)-based aggregation, or max pooling, ensuring adaptability in different graph 

structures. This methodology allows GraphSAGE to retain localized feature interactions while 

capturing hierarchical information, leading to improved classification accuracy. 

Graph Attention Networks (GAT) extend traditional graph convolution techniques by 

introducing an attention mechanism that assigns different importance weights to neighbouring 

nodes. Unlike conventional aggregation methods that treat all neighbours equally, GAT 

dynamically determines the contribution of each node through a self-attention mechanism. The 

node embedding update follows 

ℎ𝑣
(𝑘)

= 𝜎 ( ∑ 𝛼𝑣𝑢
(𝑘)

𝑊(𝑘)ℎ𝑢
(𝑘−1)

𝑢∈𝑁(𝑣)

) 

Communications on Applied Nonlinear Analysis
ISSN: 1074-133X
Vol 31 No. 8s (2024)

https://internationalpubls.com
1063



where 𝛼𝑣𝑢 is the attention coefficient computed as 

𝛼𝑣𝑢 =
e(ℒℛ(𝑎𝑇[𝑊ℎ𝑣‖𝑊ℎ𝑢]))

∑ 𝑒(ℒℛ(𝑎𝑇[𝑊ℎ𝑣‖𝑊ℎ𝑗]))
𝑗∈𝑁(𝑣)

 , 

where 𝑎 is a learnable attention vector,𝑊 is a learnable weight matrix applied to the node 

features, || denotes concatenation, 𝑁(𝑖) is the set of neighbors of node 𝑖, ℒℛ 

representsLeakyReLUit is a variant of the Rectified Linear Unit (ReLU or ℛ) activation function 

used in neural networks. The ℒℛ function is defined as: 𝑓(𝑥) = {
𝑥 , 𝑖𝑓 𝑥 > 0

𝛼𝑥 , 𝑖𝑓 𝑥 ≤ 0
 where, 𝑥 is the 

input value and 𝛼is a small leak factor (e.g., 0.01), which determines how much negative values 

are allowed to pass through instead of being completely zeroed out. The attention mechanism 

enhances learning by selectively focusing on the most relevant neighbors, improving node 

classification by preserving critical relationships within the graph. 

For comparison, we also implemented traditional ML classifiers, including Decision Trees, 

Random Forest, AdaBoost, XGBoost, and LightGBM etcetera. Among these models, AdaBoost, 

and LightGBM demonstrated the highest accuracy due to its gradient boosting framework, which 

iteratively improves weak learners. 

The key factors that contributed to achieving 100% accuracy in this study include the structural 

information captured by the graph,graph-based models provided a deeper understanding of the 

dataset by representing relationships that traditional ML models could not capture. 

4 Results and Discussions 

The classification results obtained using graph-based models, particularly Graph Neural 

Networks (GNNs), were compared with traditional machine learning (ML) approaches to 

evaluate their effectiveness in distinguishing between benign and malignant tumour data. The 

dataset was first converted into a graph structure, where nodes represented individual tumour 

samples, and edges were created based on feature similarity. This transformation is visually 

represented in fig.1 illustrating the connectivity within the dataset 

Communications on Applied Nonlinear Analysis
ISSN: 1074-133X
Vol 31 No. 8s (2024)

https://internationalpubls.com
1064



 

Figure 1: Graph Representation of Cancer Data 

The images show how different graph neural network (GNN) modelsthat isGIN, GCN, 

GraphSAGE, GATand a general GNN. The authors separate benign (blue) and malignant (red) 

tumors based on learned embeddings. Each plot represents the 2D projection of node 

embeddings using t-SNE, highlighting how well each model distinguishes between the two 

tumor types. The GCN and general GNN models create a clear boundary between benign and 

malignant clusters, while GAT, GraphSAGE, and GIN also show good separation but with some 

overlap. The color bar on the right indicates the classification labels, where 0 represents benign 

tumors and 1 represents malignant tumors. These visualizations help in understanding how 

different GNN architectures learn feature representations for medical tumor classification. 

 
               Figure2: GIN Embeddings in 2D        Figure 3:GCN Embeddings in 2D 

 

 

Communications on Applied Nonlinear Analysis
ISSN: 1074-133X
Vol 31 No. 8s (2024)

https://internationalpubls.com
1065



 
                 Figure 4: GraphSAGEEmbeddings in 2D                                       Figure 5:GATEmbeddings in 2D 
 

 

To improve classification, GNN model was trained on the same dataset. The effectiveness of this 

approach is evident in fig.6, which demonstrates a clearer separation between the two classes 

compared to the previous embedding 

method.  

  

 

Figure 6: GNN Classification Boundaries 

The classification performance was further analysed through accuracy, Precision, Recall,F1-

Score, Confusion matrix and loss trends over training epochs. The model has achieved 100% 

accuracy for GCN, GAT and GraphSAGE, and 99.42% accuracy for GIN which indicates that 

this graph-based learning approach is perfectly classifying the dataset as shown below. 

Communications on Applied Nonlinear Analysis
ISSN: 1074-133X
Vol 31 No. 8s (2024)

https://internationalpubls.com
1066



 

The output shows the performance of a Graph Convolutional Network (GCN) in classifying 

benign and malignant tumors. The confusion matrix indicates that the model achieved perfect 

classification, correctly identifying all 107 benign and 64 malignant samples without any 

misclassifications. The performance metrics confirm this, with a test accuracy of 100%, along 

with precision, recall, and F1-score all being 1.000. The loss values remain low, and early 

stopping was applied at epoch 45, ensuring optimal training without overfitting. The runtime for 

the model was just 1 second, highlighting its efficiency. This result demonstrates that the GCN 

successfully learned meaningful representations for distinguishing between benign and 

malignant tumor samples. 

Communications on Applied Nonlinear Analysis
ISSN: 1074-133X
Vol 31 No. 8s (2024)

https://internationalpubls.com
1067



 

The confusion matrix and performance metrics illustrate the effectiveness of the Graph 

Isomorphism Network (GIN) in classifying benign and malignant tumors. The model achieved a 

high-test accuracy of 99.42%, with precision, recall, and F1-score all reaching approximately 

0.994. Out of 107 benign samples, the model classified all correctly, while for 64 malignant 

cases, it correctly identified 63 but misclassified 1 as benign. The training process stopped early 

at epoch 59, indicating stability in performance. Although slightly less accurate than GCN, GIN 

still demonstrates strong classification capability, efficiently distinguishing between benign and 

malignant tumors with minimal errors. 

 

Communications on Applied Nonlinear Analysis
ISSN: 1074-133X
Vol 31 No. 8s (2024)

https://internationalpubls.com 1068



The GraphSAGE model demonstrated exceptional performance in classifying medical graph 

data, achieving 100% accuracy, precision, recall, and F1-score after only 22 epochs. As shown in 

the confusion matrix, the model correctly classified all benign (72 instances) and malignant (42 

instances) tumors without any false positives or false negatives. The early stopping mechanism 

ensured optimal training efficiency, preventing overfitting while maintaining perfect 

classification within zero seconds runtime. 

 

The Graph Attention Network (GAT) model exhibited outstanding performance in classifying 

medical graph data, achieving 100% accuracy, precision, recall, and F1-score by epoch 22. 

Initially, at epoch 0, the model had a loss of 0.6213 and an accuracy of 99.12%, which improved 

rapidly as training progressed. By epoch 10, the model reached perfect classification (100% 

accuracy), effectively distinguishing between benign (72 instances) and malignant (42 instances) 

tumors without any false positives or false negatives. The early stopping mechanism halted 

training at epoch 22, preventing overfitting while ensuring maximum performance. The 

confusion matrix confirms this, showing that all instances were correctly classified. The GAT 

model's ability to leverage attention-based message passing further supports its effectiveness in 

handling complex graph-structured medical data, making it a powerful tool for cancer detection 

and automated medical drug analysis. 

The authors tested the robustness by making small changes to the graph structure and node 

features and measured accuracy, precision, recall, and F1-score before and after these changes. 

After modifications also all four models (GCN, GIN, GAT, and GraphSAGE) performed well 

and maintained the same accuracy. The findings confirmed that GCN, GIN, GAT, and 

GraphSAGE can extract useful graph-based features while staying reliable even when the data 

changes slightly. The performance comparison of traditional ML classifiers and graph-based 

models is presented in the tables below:  

Communications on Applied Nonlinear Analysis
ISSN: 1074-133X
Vol 31 No. 8s (2024)

https://internationalpubls.com
1069



 

The test accuracy of various machine learning models indicates that collective methods such as 

AdaBoost and LightGBM achieved 98.8% accuracy, while XGBoost and Random Forest 

performed slightly lower. However, when transitioning to graph-based learning, the 

classification accuracy further improved, with GIN reaching 99.42% and GCN,GAT and 

GraphSAGE are achieving 100% accuracy, demonstrating the power of relational learning. 

Communications on Applied Nonlinear Analysis
ISSN: 1074-133X
Vol 31 No. 8s (2024)

https://internationalpubls.com 1070



These differences are visually represented in fig.7where this bar chart compares the accuracy of 

traditional machine learning models (in blue) and graph-based learning models (in red) which 

clearly illustrates the superior performance of graph-based models. 

Figure 7: Accuracy Comparison: Traditional ML versus Graph-Based Models 

Conclusion: 

This study shows that graph-based models better than traditional machine learning techniques in 

medical classification problems. Even while advanced machine learning models like AdaBoost 

and LightGBM obtained a high accuracy of 98.83% where graph-based models performed 

significantly better. Even the most advanced conventional ML models were outperforming by 

GIN, which obtained 99.42% accuracy, while GCN, GAT and GraphSAGE achieved 100% 

accuracy. 

The effect is further shown by the confusion matrices, which show that graph-based models 

achieved very reliable classification by consistently minimizing both false positives and false 

negatives. On the other hand, traditional machine learning algorithms were unable to achieve an 

appropriate balance between recall and precision, especially models that utilize linear decision 

boundaries (such as SGD, which had an accuracy of 85.38%). Even the best-performing ML 

models showed occasional misclassifications, whereas GNNs handled complex structural 

dependencies in the dataset with remarkable efficiency. 

These results highlight the ability of GNNs to capture complex relationships within medical 

datasets, making them a highly promising approach for real-world applications. Their efficacy 

lies on their capacity to represent complex structural dependencies in data, which are difficult for 

traditional machine learning models to capture. Future work will focus on exploring more 

advanced GNN architectures, optimizing computational efficiency, and expanding the study to 

larger and more diverse medical datasets to further validate these findings. 

Communications on Applied Nonlinear Analysis
ISSN: 1074-133X
Vol 31 No. 8s (2024)

https://internationalpubls.com 1071



References

[1] Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M., &Monfardini, G. (2009). The Graph 

Neural Network model. IEEE Transactions on Neural 

Networks.https://ieeexplore.ieee.org/document/4700287

[2] Kipf, T. N., & Welling, M. (2017). Semi-supervised classification with graph convolutional 

networks.International Conference on Learning Representations 

(ICLR).https://arxiv.org/pdf/1609.02907

[3] Xu, K., Hu, W., Leskovec, J., &Jegelka, S. (2019). How powerful are graph neural networks? 

ICLR.https://arxiv.org/abs/1810.00826

[4] Veličković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., &Bengio, Y. (2018). Graph 

attention networks.International Conference on Learning Representations 

(ICLR).https://arxiv.org/pdf/1710.10903

[5] Hamilton, W., Ying, R., &Leskovec, J. (2017). Inductive representation learning on large 

graphs.Advances in Neural Information Processing Systems (NeurIPS), 30. 

https://arxiv.org/pdf/1706.02216

[6] Zhang, C., Song, D., Huang, C., Swami, A., & Chawla, N. V. (2019). Heterogeneous graph 

neural network for personalized healthcare.Proceedings of the 25th ACM SIGKDD International 

Conference on Knowledge Discovery & Data Mining (pp. 1308–1316). ACM. 

https://dl.acm.org/doi/pdf/10.1145/3292500.3330961

[7] Wang, S., & Wang, S. (2023). Disease prediction using graph machine learning based 

on electronic health data.Healthcare, 11(7), 1031. https://www.mdpi.com/2227-

9032/11/7/1031

[8] Mahmud, F. B., Rayhan, M. Md. S., Shuvo, M. H., Sadia, I., &Morol, M. K. (2022). A 

comparative analysis of graph neural networks and commonly used machine learning algorithms 

on fake news detection.IEEE Access.https://arxiv.org/pdf/2203.14132

[9] Wang, Y., Ding, A., Guan, K., We, S., & Du, Y. (2022). Graph-based ensemble machine 

learning for student performance prediction.arXiv preprint.https://arxiv.org/pdf/2112.07893

[10]Shaila, S. J., & Varsha, H. S. (2024). Application of graph theory in machine learning. 

ResearchGate. 

https://www.researchgate.net/publication/379825214_APPLICATIONS_OF_GRAPH_THEOR 

Y_IN_MACHINE_LEARNING

[11] Erdem, T. (2023). Cancer data [Data set]. Kaggle. 

https://www.kaggle.com/datasets/erdemtaha/cancer-data

Communications on Applied Nonlinear Analysis
ISSN: 1074-133X
Vol 31 No. 8s (2024)

https://internationalpubls.com 1072

https://ieeexplore.ieee.org/document/4700287
https://arxiv.org/pdf/1609.02907
https://arxiv.org/abs/1810.00826
https://arxiv.org/pdf/1710.10903
https://arxiv.org/pdf/1706.02216
https://dl.acm.org/doi/pdf/10.1145/3292500.3330961
https://www.mdpi.com/2227-9032/11/7/1031
https://arxiv.org/pdf/2203.14132
https://arxiv.org/pdf/2112.07893
https://www.researchgate.net/publication/379825214_APPLICATIONS_OF_GRAPH_THEORY_IN_MACHINE_LEARNING
https://www.researchgate.net/publication/379825214_APPLICATIONS_OF_GRAPH_THEORY_IN_MACHINE_LEARNING
https://www.kaggle.com/datasets/erdemtaha/cancer-data

	Article History: Received 20.09.2024     Revised: 24.10.2024 Accepted: 30.11.2024
	Abstract
	2. Literature Review
	References