https://doi.org/10.31449/inf.v49i17.7595                                             Informatica 49 (2025) 11–24   11 
Innovative Application of Intelligent Mechanical Manufacturing 
Based on Self-Supervised Learning and Graph Neural Network 
Fusion Optimization 
Hui Wang
1,2
 
1
School of Intelligent Manufacturing, Henan Polytechnic, Zhengzhou, 450046, China 
2
Henan Province Engineering Research Center of Numerical Control Application Technology, Zhengzhou, 450046, 
China 
E-mail: wodealetai_888@163.com 
Keywords: self-supervised learning, graph neural networks, pre-training, intelligent manufacturing, optimization 
models 
Received: November 14, 2024 
In the current context of industrial automation, with the rapid development of the Internet, artificial 
intelligence and other technologies, intelligent mechanical manufacturing technology has received 
widespread attention. The study builds an intelligent manufacturing optimization model using methods 
based on self-supervised learning and graph neural networks to address the issue of insufficient 
efficiency and accuracy in mechanical manufacturing. Based on the traditional manufacturing method, 
the comparison self-supervised learning model is utilized to mine the available data and monitor the 
quality of mechanical equipment by constructing pre-training tasks to improve the efficiency of 
equipment production. Then it is fused with graph neural networks to design intelligent manufacturing 
optimization model based on self-supervised learning and graph neural networks intelligent 
manufacturing. The experimental results indicated that the comprehensive performance and practical 
application of the research model were good on both. Compared to the three performance types of 
multi-behavior models on different datasets, the model improved by 10.8%, 8.5%, and 6.3%, 
respectively. The noise robustness of the research model improved by 5.36% and 6.27% compared to the 
NGCF and EHCF models, respectively. The performance of the research model reached its optimal 
state when the selected parameters were β=0.01, K=30, and L=2. The results demonstrate that the 
designed optimization model based on self-supervised learning and graph neural networks intelligent 
manufacturing has advantages in terms of operational efficiency and index performance. The research 
results are of great significance for intelligent mechanical manufacturing. 
Povzetek: V članku je opisan inovativen optimizacijski model inteligentne mehanske proizvodnje na 
osnovi samonadzorovanega učenja in grafovskih nevronskih mrež (GNN), ki izboljša učinkovitost in 
robustnost proizvodnje (do 10,8 %). 
 
1  Introduction 
Today, with the continuous development of information 
network and artificial intelligence (AI), automation 
technology is silently penetrating and changing people's 
way of life at an unprecedented speed. Intelligent 
mechanical manufacturing, as a necessary path for 
industrial automation, has been increasing in market 
demand, and the traditional manufacturing methods are 
no longer able to meet the demand [1-2]. This 
phenomenon not only originates from the surge of 
market demand for product personalization and 
customization, but also because of the increasingly 
fierce competition in the global market, which requires 
enterprises to be able to respond to market changes more 
quickly and realize product iteration and upgrade [3]. 
With its strong automatic recognition capability, 
machine learning (ML) approaches have drawn the 
attention of several researchers in recent years who are 
attempting to predict possible equipment failures of 
machinery. However, the existing ML techniques are 
still insufficient in terms of recognition efficiency and 
accuracy, and the wide range of applications cannot 
meet the basic requirements of intelligent mechanical 
manufacturing [4]. The primary concerns have centered 
on the utilization of self-supervised learning (SSL) and 
graph neural networks (GNNs) technologies to enhance 
the accuracy and efficiency of intelligent 
recommendation systems in mechanical manufacturing 
processes, with a focus on improving anomaly detection 
and fault prediction. SSL is one of the most classical 
techniques in ML technology. By analyzing the data 
generated during the operation of mechanical equipment 
to identify abnormal patterns, the effect of quality 
detection can be achieved. Moreover, GNN, as a tool for 
information aggregation (IA) and transfer, can capture 
the complex relationship between users and equipment 
for the recommendation of relevant mechanical parts [5]. 
In this context, to improve the efficiency and accuracy 
of mechanical manufacturing, the study attempts to 

12   Informatica 49 (2025) 11–24 H. Wang 
innovatively design an optimization model based on 
SSL and GNN intelligent manufacturing. The novelty of 
the research lies in the introduction of heterogeneous 
graphs and meta-paths to combine SSL with GNN to 
innovatively improve anomaly detection and failure 
prediction capabilities in intelligent mechanical 
manufacturing. This integration not only improves the 
model performance, but also enhances the adaptability 
and generalizability to complex industrial scenarios. It 
aims to provide a technical foundation for intelligent 
mechanical manufacturing that can improve production 
efficiency and quality. 
2  Related works 
With the intelligent and automated development of the 
manufacturing industry, the traditional manufacturing 
methods have been unable to meet the production needs, 
while intelligent mechanical manufacturing has become 
an important trend in the development of the 
manufacturing industry. As a result, numerous 
academics have studied intelligent mechanical 
manufacturing techniques in related fields. He H 
combines deep reinforcement learning with graph neural 
networks (GNN) to propose an intelligent network 
traffic scheduling algorithm to solve traffic scheduling 
problems in large-scale dynamic network environments. 
This algorithm leverages the decision-making 
capabilities of deep reinforcement learning and the 
advantages of GNNs in processing graph structured data 
to achieve an efficient decision-making process from 
macro strategy formulation to micro-operation execution 
through a hierarchical reinforcement learning 
framework. The experimental results show that the 
algorithm exhibits excellent robustness and 
generalization ability [6]. Sharma T et al. constructed 
fuzzy ensemble technology based on mathematical 
analysis of trust scores and deep neural networks to 
promote the research of automatic ALSC. The proposed 
method can correct the misclassification of basic deep 
learners through reward and punishment strategies. The 
experimental results conducted on five benchmark 
datasets show that this method has advantages over 
component-based deep neural networks and other 
important benchmarks [7]. Yao X et al. proposed a 
blockchain-based value-added industrial meta-universe 
solution for intelligent manufacturing based on 
cyber-physical systems. The study used intelligent 
manufacturing social technology, enhanced intelligent 
manufacturing architecture and framework to achieve 
high resource utilization. The results indicated that the 
proposed scheme was intelligently resilient and 
sustainable [8]. Aldrini J et al. proposed intelligent fault 
diagnosis and self-healing conceptual model architecture 
for intelligent manufacturing system to overcome the 
performance and safety issues of intelligent 
manufacturing. The approach improved the performance 
of intelligent manufacturing systems through fault 
detection as well as self-healing fault-tolerant strategies. 
Experimental results indicated that the proposed 
structure had a good robustness [9]. To facilitate 
intelligent decision-making in dynamic scheduling and 
reconfiguration, Yang S. suggested a mathematical 
model of intelligent scheduling based on deep 
reinforcement learning. According to test results, the 
suggested model has good scheduling efficiency [10]. 
Intelligent mechanical manufacturing often introduces 
artificial networks and ML methods for research, and 
several scholars have conducted research on SSL and 
GNN. Krishnan R proposed self-supervised methods 
and models for medicine and healthcare in response to 
the problem of medical AI development. The study 
investigated SSL's benefits and drawbacks while 
creating models with multimodal datasets. The 
suggested approach could hasten the development of 
medical AI, according to the findings [11]. Rani V et al. 
proposed SSL detection for image processing problems. 
The study used unstructured and unlabeled data to 
develop an AI system to perform downstream computer 
vision tasks. The findings indicated that this method had 
good applicability [12]. Kumar P incorporated contrast 
learning into SSL to investigate the conditions of 
application in different domains. The findings indicated 
that this method has positive significance for 
complementary contrast learning [13]. To address the 
issue of information overload, Wu S. et al. presented a 
recommendation model based on GNN. The method 
provided information categorization based on the type of 
information used and the recommendation task, and 
extracted content from the interaction information. 
Numerous experiments indicated that the proposed 
method had a good stability [14]. Zhou Y proposed a 
new classification method to overcome the limitations of 
GNN. The study categorized the GNN into the 
corresponding categories after showing the full picture 
of the GNN. The test results indicated the superiority of 
the proposed approach in categorization [15]. The 
summary and analysis of existing research methods are 
shown in Table 1. 
 
Table 1: Summary and analysis of currently available research methods 
Methodologies 
Performance 
indicators 
Gaps Limitations Unique requirement 
He H [6] 
Network traffic 
scheduling efficiency 
Traffic scheduling 
problem in 
large-scale dynamic 
network environment 
The efficiency and 
robustness of 
traditional algorithms 
in handling 
large-scale dynamic 
networks 
Improve scheduling 
efficiency and 
robustness 
Sharma T et al. [7] Automatic ALSC Integrated Error classification Improve the 

Innovative Application of Intelligent Mechanical Manufacturing Based… Informatica 49 (2025) 11–24 13 
performance technology of 
mathematical 
analysis based on 
trust scores and deep 
neural networks 
of basic deep 
learners 
accuracy and 
robustness of 
automatic ALSC 
Yao X et al [8] 
Speed of 
redeployment of 
resources 
Slow to respond to 
market changes 
Lack of real time 
Real-time resource 
allocation 
Aldrini J et al [9] Self-healing time 
Slow response of 
self-healing systems 
Limited ability to 
self-heal 
Rapid self-healing 
ability 
Yang S. [10] 
Movement control 
accuracy 
Poor adaptation of 
highly dynamic 
scheduling 
Inadequate 
adaptation 
Dynamic Scheduling 
Adaptation 
Krishnan R [11] 
generalization 
capability 
Inadequate ability to 
generalize medical 
data 
Limitations of 
multimodal data 
processing 
Multimodal data 
processing 
Rani V et al [12] Outlier detection rate 
Poor handling of 
image processing 
outliers 
lack of robustness Enhanced robustness 
Kumar P [13] 
Cross-domain 
accuracy 
Insufficient 
flexibility in 
cross-cutting 
applications 
Poor cross-cutting 
adaptation 
Cross-cutting 
adaptability 
Wu S. et al. [14] processing speed 
Poor stability of 
massive data 
processing 
information overload 
Information 
processing efficiency 
Zhou Y [15] 
Classification 
accuracy 
Poor classification of 
complex network 
structures 
Insufficient 
classification 
accuracy 
Improvement of 
classification 
accuracy 
 
To summarize, although many experts and scholars have 
carried out related research on intelligent mechanical 
manufacturing, there has not been an approach that 
combines with SSL and GNN. In view of this, the study 
proposes an intelligent mechanical manufacturing model 
based on SSL and GNN. The study combines the data 
characteristics and technological advantages of 
real-world application scenarios to construct a 
supervised networked intelligent manufacturing 
innovation approach. This can improve production 
efficiency and quality and contribute to expanding the 
innovative applications of intelligent mechanical 
manufacturing. 
3 Intelligent mechanical 
manufacturing method based on SSL 
and GNN 
3.1 Intelligent mechanical manufacturing 
design based on SSL 
With the increasing demand for intelligent mechanical 
manufacturing, the traditional manufacturing methods 
are more and more incompetent in terms of efficiency, 
precision and flexibility [16]. Therefore, it is necessary 
to propose an innovative manufacturing program that 
can improve the production efficiency and product 
quality of mechanical manufacturing. The SSL 
technique has gained a lot of traction in ML in recent 
years. From a vast amount of unlabeled data, it  
 
 
automatically learns usable feature representations by 
using the properties of the data itself as supervised 
information. When compared to conventional 
supervised learning techniques, SSL shows a more 
potent and adaptable approach to data use. Its core 
advantage lies in getting rid of the reliance on external 
labels and utilizing the internal structure and properties 
of the data to automatically generate supervised signals 
[17]. This feature enables SSL to generate the potential 
to effectively mine large-scale data. Fig. 1 depicts the 
SSL model. 
In Fig. 1, the SSL model learns advanced 
representations or features of data by designing 
pre-training tasks. The pre-training task of SSL is 
designed as contrastive learning, which involves data 
encoding, positive and negative sample pair generation, 
and similarity comparison operations in sequence. These 
pre-training tasks help the model extract useful 
information from data without external labels, 
improving the model's ability to detect abnormal 
patterns in mechanical manufacturing processes. 
Therefore, it can be migrated to other related supervised 
learning tasks for efficient and accurate model training. 
The SSL algorithmic models can be categorized into 
three types: adversarial, generative, and comparative 
models. Since intelligent mechanical manufacturing 
requires comparative damage determination, the study 
mainly uses comparative SSL models for learning. Fig. 
2 depicts the contrastive model's structure. 
 

14   Informatica 49 (2025) 11–24 H. Wang 
 
Unlabeled data Network model Predecessors Annotate data Network model Predecessors
 
Fig. 1 SSL model architecture 
 
Input graph
Image 
enhancement
View 1
View 2
Encoder
Embedded 
representation 1
Embedded 
representation 2
Maximizing 
mutual 
information
 
Figure 2: Comparing the structure of self-supervised learning models 
 
In Fig. 2, the comparative model works by first 
encoding the input data efficiently by training the 
encoder. The encoded data are fed into the discriminator, 
and subsequently these encoded data are used to 
perform similarity comparison operations, contrasting 
losses to maximize mutual information. By analyzing 
the massive amount of data generated during operation, 
the SSL algorithm can automatically identify anomalous 
patterns and predict potential equipment failures. The 
original identification formula is shown in Equation (1). 
 
( , ) =
r r r r
G N M E
   (1) 
 
In Equation (1), G is the graph. N for the set of 
nodes. r is the type of relationship. 
r
G
 is the 
original data (OD) sample. 
r
N
 
is the set of nodes in 
the data sample. 
r
M
 represents the original mask 
vector, which is used to randomly discard the edges of 
the OD. 
r
E
 represents the set of edges in the OD 
sample. Next, the loss identification is performed as 
shown in Equation (2). 
 
( , ) =
S S S S
G N M E
   (2) 
 
In Equation (2), S is the loss generation. 
S
G
 is the 
processed loss samples. 
S
N
 represents the node set of 
loss samples. 
S
M
 represents the loss mask vector. 
S
E
 
represents the set of edges in the loss sample. Then, 
determine the positive probability of unlabeled data 
features by first calculating the cosine similarity of each 
data point, then calculating the sum of the similarities of 
all data points. Finally, the ratio of these two sums is the 
probability of positive samples, as shown in Equation 
(3). 
max(cos( , ))
+
=
rs
ii
y Soft e e
   (3) 
 
 
 
 
In Equation (3), 
+
r
i
y
 is the probability that the data 
i 
in the unlabeled sample set is positive for a given data 
feature. 
e 
is the embedding representation of the data 
in the unlabeled sample set. 
s
i
e
 is the feature 
embedding representation of the data. 
cos( )
 
is the 
cosine similarity. Then, it is necessary to obtain the data 
self-supervised signal (SsS) and define a set of SsSs that 
includes all cosine similarity values exceeding the 
threshold. As shown in Equation (4). 
 
 
( ), ~
++
=  −
rr
i k i r
P e k Top K y E G
   (4) 
 
In Equation (4), 
+
r
i
P
 is the set of SsSs for information 
i 
under data information interaction. 
k
e
 
is the 
K
 
data with the highest value after passing the confidence 
test. 
k 
represents all data of the confidence test. E is 
the set of edges. 
E
 
represents the set of edges in the 
data. These signals are used to train the model to learn 
useful feature representations from the structure of the 
data itself. Similarly, the feature-based SsS for data 
under data information interaction can be obtained as in 
Equation (5). 
 
 
( ), ~
++
=  −
ss
i k i s
P e k Top K y E G
   (5) 
 
In Equation (5), 
+
s
i
P
 represents the set of SsSs for 
information 
i 
under data information interaction. 
+
s
i
y
 
represents the probability that the data in the labeled 
sample set is positive for a given data feature. To 
accurately express the structural relationships, path 
lengths and other factors in the intelligent network, the 
study introduces heterogeneous graphs and meta-paths 
shown in Fig. 3. 

Innovative Application of Intelligent Mechanical Manufacturing Based… Informatica 49 (2025) 11–24 15 
u1
u2
u3
i1
i2
i3
i4
i5
s1
s2
(a) Heterogeneous graph (b) Network structure
User
Project
Category
(c) Example of meta path
i1 s1 u1
u1 i3 u3
u3
i3
Affiliation
Interactive
Interactive Affiliation Interactive
Interactive
 
Figure 3: Schematic diagram of heterogeneous graph 
and meta path 
 
In Fig. 3(a), the heterogeneous graph is a complex data 
structure that can accommodate multiple types of nodes 
and relationships. At the same time, nodes can exhibit 
diverse performance characteristics. To explore the 
higher-order relationships between nodes in depth, the 
introduction of meta-paths can effectively portray these 
complex, non-direct inter-node connections. In Fig. 3(b), 
the mechanical network structure constructs a hidden 
relationship between users, items and categories through 
a meta-path. This path not only shows the possible 
connections between the three, but also provides a new 
perspective for understanding the underlying workings 
of the network in great detail. In Fig. 3(c), after 
completing the detailed portrayal of different types of 
node association patterns, the node associations in the 
whole network become clearly visible. This 
visualization provides a solid foundation for the 
subsequent analysis and mining work of intelligent 
mechanical manufacturing. 
3.2 Optimization strategies for intelligent mechanical 
manufacturing incorporating GNN 
A significant amount of unlabeled data can be 
automatically learned by the model to create valuable 
feature representations through SSL, which are essential 
for ensuing prediction, classification, or optimization 
tasks. GNN can create more precise and thorough 
models of mechanical systems by taking use of the 
intricate interactions that exist between mechanical 
components. GNNs are a subset of neural networks that 
are specifically trained on graph-structured data (GSD). 
By capturing the intricate interrelationships between 
nodes, GNNs are able to learn the properties of the data 
more precisely [18]. Therefore, the research will 
incorporate GNN for the optimization of intelligent 
mechanical manufacturing. Equation (6) illustrates the 
formalization process of the graph learning-based 
recommendation system. 
 
arg ( ( ) ) = T Maxf S G
   (6) 
 
In Equation (6), 
T
 represents the optimal 
recommendation result. 
()  S
 represents the 
recommendation model based on graph learning. 
G 
represents the data structure information of the 
interaction graph. 

 represents the parameters in the 
interaction graph. Then the complex relationship 
between users or devices is captured by GNN based on 
the demand degree of related parts. The demand degree 
formalization process is shown in Equation (7). 
 
min_
_
max_ min_
−
=
−
i
i
Pop Pop
nor Pop
Pop Pop
   (7) 
 
In Equation (7), 
_
i
nor Pop
 represents the demand 
degree after normalization of the part data. 
min_ Pop
 
and 
min_ Pop
 
represents the maximum and minimum 
value of the part demand degree. Then the difference 
value of demand degree between two parts can be 
obtained, and the formula is as in (8). 
,
_ _ _ =−
i j i j
Pop Bias nor Pop nor Pop
   (8) 
In Equation (8), 
,
_
ij
Pop Bias
 
represents the difference 
in demand degree between parts 
i 
and j . An example 
of the recommended system data graphic is obtained as 
shown in Fig. 4. 
 
Digital Electronics
Computer office 
category
 
Figure 4: Recommended system data graph example 
 
As illustrated in Fig. 4, the nodes represent users and 
projects, while the edges represent the users' preferences 
or interactions with projects. The figure demonstrates a 
variety of interaction modes, encompassing both direct 
user-project interactions and indirect interactions 
through other user nodes, thereby reflecting social 
influence. After changing the basic data of the 
recommender system to GSD, the complex relationship 
between users and items can be captured more clearly 
and the items can be categorized by attributes. 
Meanwhile, the complex interactions between users and 
items can form different types of edges, reflecting the 
preference intensity or interest patterns of different users 
for mechanical engineering parts. These edges link user 
nodes and item nodes, which form the graph structure of 
the recommender system [19]. This graphical 
representation method makes complex user project 
relationships intuitive and easy to understand. The GNN 
learning method based on graph results has advantages 

16   Informatica 49 (2025) 11–24 H. Wang 
in mining the relationship between nodes and improving data utilization. The GNN schematic is shown in Fig. 5. 
Input
Hidden layer
. . .
Activation 
function
Hidden layer
. . .
Activation 
function
. . .
Output
 
Figure 5: Diagram of neural network 
 
In Fig. 5, each node in the GNN is forced by the learnt 
function mapping to aggregate data from its nearby 
nodes in order to gather the neighboring nodes' feature 
information. The information transfer path in this case 
points from the target node to the neighbor node. The 
representation of the node is then updated based on the 
information obtained from the aggregation and the 
node's own features. During the iteration process, the 
node representations will gradually stabilize, thus 
capturing the structural information of the graph and the 
complex relationships between the nodes. The GNN can 
deeply explore and utilize the complex and high-order 
connectivity information in the user-item interaction 
graph through its unique structural design, and it can 
also mine and disseminate more diversified items for 
users [20]. The propagation rule is shown in Equation 
(9). 
 
1
( ( ( )) ) 
+
=    +
l
ik
h W AGG h k N i b
   (9) 
 
In Equation (9), 
,
_
ij
Pop Bias represents the 
propagation of GNN. 
 
represents nonlinear 
activation function. 
W
 
and 
b
 are weights and bias 
vectors. 
() AGG
 
represents the aggregation function. 
k
h
 
is the node association information within the 
region. 
() Ni
 represents the set of nodes within the 
range of node 
i
. In order to mine the data features 
more fully for extraction, the study uses graph encoder 
to extract the data features. Each node is represented as 
a feature vector, and edges represent the relationships 
between nodes. GNN aggregates information from 
neighboring nodes through a message passing 
mechanism. At each layer, nodes collect features from 
their neighbors and integrate this information through 
max-pooling. The aggregated features are transformed 
by nonlinear activation functions to learn advanced 
representations of nodes. Finally, by stacking multiple 
GNN layers, the model can capture the multi-hop 
neighbor information of nodes and learn complex graph 
structural features. The general structural formula of 
graph encoder is Equation (10). 
( ) ( 1)
( , )
−
=
ll
E H E G
   (10) 
 
In Equation (10), 
,
_
ij
Pop Bias represents the node 
l 
layer embedding representation. 
H 
represents the 
encoder of IA. 
G
 represents the view after IA. The 
l
th layer representation obtained through neighbor node 
layer embedding is Equation (11). 
 
( ) ( )
1

=

c
i
c l c l
ij
cc
jN
ij
ee
NN
   (11) 
 
In Equation (11), 
() cl
i
e
 and 
() cl
j
e
 represent the 
embedding representations obtained by the neighbor 
nodes. 
c
i
N
 represents the set of all neighbors of part 
i
 
in the part similarity graph with de-demand degree. 
c
j
N
 
represents the set of all neighbors of part j in the part 
similarity graph with de-demanded degree. Then the 
final vectors of the parts are obtained using the 
multilayer information propagation aggregation layer as 
in Equation (12). 
 
()
0

=
=

L
c c l
i l i
l
ee
   (12) 
 
In Equation (12), 
c
i
e
 represents the final embedding 
vector. 
L
 represents all layers of GNN IA. 
() cl
i
e
 
is the 
superposition result of the initial embedding. 

l
 
represents the embedding representation weight 
parameter of 
l
 layer. The SSL is then combined with 
GNN to further enhance the intelligence of intelligent 
mechanical manufacturing, especially in anomaly 
detection and fault prediction. The derived optimization 
model architecture based on SSL and GNN intelligent 
manufacturing is shown in Fig. 6. 
In Fig. 6, the integration of GNN and SSL began with 
the construction of a multi-behavioral interaction feature 
extraction mechanism. Among them, SSL is used to 
automatically learn feature representations from a large 

Innovative Application of Intelligent Mechanical Manufacturing Based… Informatica 49 (2025) 11–24 17 
amount of unlabeled data from mechanical operations in 
order to identify abnormal patterns. Then, these feature 
representations and SsSs are used as inputs to GNN, 
which updates node representations by aggregating 
neighbor node information, captures high-order 
connectivity relationships between devices, and predicts 
potential failures. Among them, Bayesian personalized 
ranking BPR is a method for learning user preferences, 
while p, q, and n are embedded neural network data. 
Through the information propagation and aggregation 
mechanism of GNN, the model is able to capture 
higher-order connectivity between devices and thus 
identify abnormal behaviors that significantly deviate 
from normal patterns. The model greatly raises the 
degree of equipment intelligence and assists in 
anticipating maintenance needs in order to lessen the 
impact of malfunctions on output. 
 
Recommendation supervision task based on 
behavioral characteristics
Behavioral influence weight 
embedding propagation
Embedded propagation of 
implicit behavioral features
The final 
embedding of 
user nodes
Final 
embedding of 
project nodes
BPR
Interactive heterogeneity
. . .
. . .
Initialize 
embedding
p
q
n
Multi task joint 
optimization
Self supervised assisted learning tasks
GCN encoding
GCN encoding
Embedded 1
Embedded 2
Loss
 
Figure 6: Intelligent manufacturing optimization model 
architecture 
4  Effectiveness analysis of SSL and 
GNN in intelligent mechanical 
manufacturing 
4.1 Comparative experimental analysis of graph SSL 
algorithms based on intelligent manufacturing 
To analyze the effectiveness of the application of 
intelligent mechanical manufacturing based on SSL and 
GNN designed by the study, the superiority and 
reliability of the research methodology is mainly 
verified. The study uses Case Western Reserve 
University (CWRU) and Phm ieee 2012 dataset of 
publicly available datasets as experimental data. The 
CWRU dataset is a well-known dataset used for bearing 
fault diagnosis that contains bearing operating data 
under various operating conditions and can represent the 
reliability of the entire system. However, some types of 
faults have fewer samples, which can lead to data 
imbalance problems. The Phm IEEE 2012 dataset is a 
maintenance prediction and health monitoring dataset 
that contains operational data of various mechanical 
equipment and can reflect various conditions that may 
occur during actual operation of the equipment. 
However, the data dimension is relatively large. Both 
types of data have good representativeness and public 
availability. For the CWRU dataset, the study removed 
all samples containing missing values and standardized 
the features. The PHM IEEE 2012 dataset has been 
normalized to ensure that the eigenvalues are located on 
the same scale. The cross-validation method is used to 
train and test the model effect. There are three graph 
encoder layers and a 40-entity embedding dimension. 
The selection of three GNN layers is based on 
experimental results and provides a good balance 
between model performance and computational 
efficiency. A reduced number of layers might not 
entirely capture the intricacies of the graph structure, 
while an increased number of layers could result in 
overfitting. The 40-dimensional embedding space is 
large enough to capture rich feature representations 
while avoiding the curse of dimensionality. The 
experimental environment is shown in Table 2. 
 
Table 2: Intelligent mechanical manufacturing 
experimental environment 
Environment entry Environmental parameters 
Operating system Linux CentOS7 
Development language Python3.7 
Developing deep learning 
frameworks 
Tensorflow1.14 
Hard disk 1T 
Memory 16G 
GPU 
GeForce RTX 2080 
SUPER 
CPU 
Intel(R) Xeon(R) Silver 
4215R CPU @ 3.20GHz 
 
The GNN layers was set to 4, and the number of 
feature-based part-similar neighbors sampled was set to 
50. The initial data and user embedding are set to 80, 
and the regularization parameter is set to 0.005. The 
initial weight is set to 1/4, and the learning rate is 
adjusted to 0.001. The Top-N recommendation hit rate 
(HR) is used as the evaluation index, and the specific 
comparison of each recommender system is made 
between the cumulative gain (CG), discounted 
cumulative gain (DCG), normalized discounted 
cumulative gain (NDCG) and training time (TT) and 
parallel processing capability (PPC). The 
single-behavior neural graph collaborative filtering 
(NGCF), light graph convolutional network (LGCN), 
multi-behavior model efficient heterogeneous 
collaborative filtering (EHCF) and the proposed model 
of the study are compared for experimental validation. 
Among them, CG can intuitively display the 
performance of recommendation systems at different 

18   Informatica 49 (2025) 11–24 H. Wang 
ranking depths. DCG is capable of simulating the user 
experience while browsing recommendation lists in a 
realistic manner. NDCG possesses the capability to 
directly compare the performance of disparate 
recommendation systems or users. TT and PPC are 
capable of evaluating the computational efficiency and 
scalability of the model. The model evaluation 
performance is shown in Fig. 7. 
(b) Phm ieee 2012 dataset (a) CWRU dataset
0.30
0.25
0.20
0.15
0.10
0.05
0.00
CG
HR@10 HR@50 DCG@50 NDCG@50
RM
NGCF
LGCN
0.6
0.5
0.4
0.3
0.2
0.1
0.0
CG
HR@10 HR@50 DCG@50 NDCG@50
RM
NGCF
LGCN
(c) CWRU dataset+Phm ieee 2012 dataset
10
8
6
4
2
TT/s
TT PPC
RM
NGCF
LGCN
Indicator type
70
60
50
40
30
PPC/%
 
 
Figure 7: Model evaluation performance 
 
Fig 7 compares the performance of different 
recommendation models on two datasets, including HR, 
discount cumulative gain (DCG), normalized discount 
cumulative gain (NDCG), as well as TT and PPC. In Fig. 
7, the research model performs significantly better on 
both CWRU and Phm ieee 2012 datasets. In Fig. 7(a), 
on the CWRU dataset, the research model outperforms 
the single-behavior NGCF model by 5.6%, 6.4%, and 
4.8% for HR, DCG, and NDCG, respectively. It 
demonstrates that the study model is more accurate at 
forecasting objects that users might come into contact 
with later on. It also improves 10.8%, 8.5%, and 6.3% 
over the HR, DCG, and NDCG of the multi-behavioral 
model EHCF, respectively. In Fig. 7(b), the performance 
comparison of the research model on the Phm ieee 2012 
dataset is similar to that on the CWRU dataset. In Fig. 
7(c), the TT of the model decreased by 2.4% and 1.8%, 
respectively, compared to the other two. PPC increased 
by 16.4% and 10.2%, respectively. As the complexity of 
the dataset increases, the performance improvement of 
the research model becomes more significant. It 
indicates its strong robustness in handling complex user 
behavior data. Overall, by combining SSL and GNN, the 
research models can more deeply explore complex 
features and patterns in data, including subtle 
differences in user behavior and complex relationships 
between items, to better understand and predict user 
preferences. Moreover, it performs well in terms of 
computational efficiency and scalability. To further 
evaluate the effect of the model in the practical 
application of intelligent mechanical manufacturing, the 
study selects NGCF and EHCF models and the research 
model to conduct noise robustness anti-interference 
experiments. Adding Gaussian noise with a mean of 0 
and an adjustable standard deviation to the dataset can 
simulate the random error of the sensor readings. Then 
gradually control the noise level and observe the 
performance changes of the system at different noise 
levels. The noise robustness results of intelligent 
mechanical manufacturing practical application are 
shown in Fig. 8. 

Innovative Application of Intelligent Mechanical Manufacturing Based… Informatica 49 (2025) 11–24 19 
 
(b) Phm ieee 2012 experimental results of dataset (a) CWRU experimental results of dataset
0.17
0.16
0.15
0.14
0.13
0.12
0.11
NDCG@50
0% 5% 10% 15%
RM
NGCF
LGCN
0.075
0.070
0.065
0.060
0.055
0.050
0.045
RM
NGCF
LGCN
Noise ratio Noise ratio
20%
NDCG@50
0% 5% 10% 15% 20%
 
 
Figure 8: Comparison experiment results of noise robustness 
 
In Fig. 8, the practical application of the model under 
both datasets displays an obvious decreasing trend with 
the increase of the noise intensity during mechanical 
fabrication. In Fig. 8(a), on the CWRU dataset, the 
research model shows a gain of 5.36% and 6.27% in the 
NDCG metrics over the NGCF and EHCF models, 
respectively. In Fig. 8(b), on the Phm ieee 2012 dataset, 
the study model has a gain of 4.26% and 5.85% in 
NDCG metrics over the NGCF and EHCF models, 
respectively. The experimental result data show that the 
research model is more effective in maintaining the 
accuracy and relevance ranking of the recommendation 
results in a complex and variable noise environment. 
This reflects its superior data processing capability and 
feature extraction strategy. It is proved that the research 
model has wide adaptability and stability in dealing with 
different sources and types of noise, which can provide 
more reliable recommendation services in the practical 
application of intelligent mechanical manufacturing. To 
demonstrate the contribution of individual components 
of SSL and GNN to the overall performance, the study 
conducted an ablation experiment analysis. The 
proportion of relevant items in the top 10 items of the 
recommendation list is denoted as Precision@10. The 
proportion of items successfully recommended by the 
recommender system among all relevant items is 
denoted as Recall@10. The reconciled average of 
Precision and Recall is compared with the F1-Score@10 
which measures the overall performance of the model. 
The results are shown in Table 3. 
 
Table 3: Results of ablation experiment table 
Performance 
indicators 
SSL model 
only 
GNN 
model only 
SSL+GNN 
model 
HR@10 0.60 0.65 0.75 
NDCG@10 0.45 0.50 0.60 
Precision@10 0.55 0.60 0.70 
Recall@10 0.50 0.55 0.65 
F1-Score@10 0.52 0.57 0.67 
 
According to Table 3, the complete model outperforms 
models using SSL or GNN alone in all performance 
metrics. The integration of SSL to learn useful feature 
representations of data and GNN to capture and utilize 
complex relationships in graph structured data furnishes 
a potent instrument for intelligent mechanical 
manufacturing. To further validate the superiority of the 
model, the experimental section is expanded to 
benchmark new models beyond NGCF and EHCF. The 
model includes sparse aggregation of graph 
convolutions for inductive learning of node 
representations (GraphSAGE), graph convolutional 
network (GCN), and graph isomorphism network (GIN) 
used for inductive learning of node representations. The 
benchmark test results are shown in Table 4. 
 
Table 4: Model benchmark test table 
Performance 
indicators 
GraphSAG
E 
GC
N 
GI
N 
SSL+GN
N model 
HR@10 0.58 0.62 
0.7
0 
0.75 
NDCG@10 0.48 0.52 
0.5
8 
0.60 
Precision@1
0 
0.52 0.57 
0.6
3 
0.70 
Recall@10 0.48 0.52 
0.6
0 
0.65 
F1-Score@1
0 
0.50 0.54 
0.6
1 
0.67 
 
In Table 4, the SSL+GNN model outperforms other 
baseline models in all key performance indicators. This 
indicates that the SSL+GNN model can more effectively 
support decision-making, optimize resource allocation, 
and improve production efficiency in practical 
applications. 
 
4.2 Experiments on parameter analysis of 
graph SSL algorithms based on intelligent 
manufacturing 
The research model's performance can be better 
understood, its configuration can be optimized, and its 
capacity for generalization can be enhanced by 

20   Informatica 49 (2025) 11–24 H. Wang 
conducting a thorough analysis of its key parameters 
following the conclusion of the model's numerous 
performance comparison trials. Among these are the 
number of GNN coding layers L, the number of SsSs K, 
and the SSL scale weight coefficient β. The L of layers 
2-4 can achieve a good balance, while K needs to be 
determined through experimentation to find the optimal 
value. A smaller beta value emphasizes the importance 
of downstream tasks, while a larger beta value gives 
more weight to self-supervised tasks, usually set at 
0.001. The experiments first set different values of β on 
Movielens-M, Last-FM and Book-Crossing datasets. It 
also keeps K and L constant, both set to 40. The key 
performance metrics such as NDCG, recall, etc. are 
recorded for each β-value. The impact of the obtained 
SSL scale weight coefficient β is shown in Fig. 9. 
 
(a) Movielens-1M (b) Last-FM (c) Book-Crossing
Recall
NDCG
Recall
NDCG
Recall
NDCG
0.26
0.24
0.22
0.20
0.18
0.30
0.28
0.26
Numerical value
Numerical value
Numerical value
0.24
0.22
0.20
0.18
0.16
0.10
0.09
0.08
0.07
0.06
0.05
0.00 0.00 0.00
β Value β Value β Value
 
Figure 9: The influence of scale weight coefficient β on SSL 
 
In Fig. 9, the larger the value of β is taken, the NDCG 
and recall performance of the model decreases. In Fig. 
9(a), on the Movielens-M dataset, the model 
performance is best in comparison when the value of β 
is taken at 0.001. Its model performance decreases 
rapidly after taking a value of 0.004, while the recall 
value is always greater than the NDCG. In Fig. 9(b), on 
the Last-FM dataset, the model performance is best 
when β is taken at an initial value, and the model 
performance decreases after the value is increased to 
0.002. In Fig. 9(c), on the Book-Crossing dataset, the 
model performance appears to decrease and then 
rebound as the value of β is increased. However, the 
rebound is small and far from the initial value. The 
results show that the research model is extremely 
sensitive to β, and the performance of the research 
model is close to the ideal state when β takes smaller 
values. In studying the effect of different SsS numbers K 
on the model, the value of L is set constant and β is set 
to 0.001. The obtained effect of SsS number K is shown 
in Fig. 10. 
In Fig. 10, as the value of K grows, the research model 
performance shows a general trend of increasing and 
then declining, and the recall value is always bigger than 
NDCG. In Fig. 10(a), the model performance shows the 
most obvious trend of increasing and then decreasing on 
the Movielens-M dataset, where the optimal K value is 
30. Meanwhile, the model performance is relatively 
stable on the other two datasets. The results show that 
the model performance varies most significantly on the 
Movielens-M dataset, and the characteristics of the 
dataset make the SsSs more sensitive to the model 
performance. Among them, the optimal K value enables 
the model to fully utilize the SsSs to enhance the 
learning effect of the main task. This reflects the 
important role of the model for intelligent mechanical 
manufacturing SSL. In studying the effect of different 
GNN coding layers L on the model performance, β is set 
to take a constant value and K is 30. The effect of GNN 
coding layers L obtained is shown in Fig. 11. 
In Fig. 11, the recall and NDCG performance of the 
research model on all three datasets increases and 
subsequently drops as the value of L increases. In Fig. 
11(a), on the Movielens-M dataset, the model 
performance increases to a maximum at L=2, but 
continues to decrease when L>3 and reaches a steady 
state at L=4. On the other two datasets, the model 
performance also maximizes at L=2 and continues to 
decrease afterward. The results show that the number of 
SsSs has a significant effect on model performance, and 
the right amount of SsSs can help the model learn more 
useful features. However, too many signals may 
introduce noise and lead to a decrease in model 
performance. In summary, the performance of the 
research model in the field of intelligent mechanical 
manufacturing is optimized when the parameters β=0.01, 
K=30, and L=2 is selected. Choosing β=0.01 means that 
the weight of SsSs in the overall objective function is 
low. It helps the model to focus on self-supervised tasks 
without overly relying on these signals, thus maintaining 
sensitivity to the final task. Choosing K=30 means that 
an appropriate amount of SsSs is used in model training. 
It helps the model capture complex structures and 
patterns in the data while avoiding overfitting. Choosing 
L=2 means that the model can capture key information 
about the graph structure while maintaining 
computational efficiency. This finding not only validates 
the effectiveness of SSL and GNN in complex industrial 

Innovative Application of Intelligent Mechanical Manufacturing Based… Informatica 49 (2025) 11–24 21 
scenarios, but also demonstrates the importance of 
model performance that can be significantly improved 
by fine-tuning the model parameters. 
 
(b) Last-FM (c) Book-Crossing
Recall
NDCG
Recall
NDCG
Recall
NDCG
0.26
0.25
0.24
0.23
0.22
0.30
0.28
0.26
Numerical value
Numerical value
Numerical value
0.24
0.22
0.20
0.18
0.16
0.10
0.09
0.08
0.07
0.06
0.05
1
0.00 0.00
5 10 1520 30 4050 100 1 5 10 1520 30 4050 100 1 5 10 1520 30 4050 100
(a) Movielens-1M
K Value K Value K Value
 
Fig. 10 The influence of the number of self-supervised signals K 
 
Recall
NDCG
Recall
NDCG
0.26
0.25
0.24
0.23
0.22
0.30
0.28
0.26
Numerical value
Numerical value
Numerical value
0.24
0.22
0.20
0.18
0.16
0.10
0.09
0.08
0.07
0.06
0.05
1
0.00 0.00
2 3 4 5 1 2 3 4 5 1 2 3 4 5
(a) Movielens-1M (b) Last-FM (c) Book-Crossing
Recall
NDCG
L Value L Value L Value
 
 
Figure 11: The influence of L encoding layers in neural networks 
 
5  Discussion 
The study used a fusion method of SSL and GNN to 
construct an intelligent manufacturing optimization 
model. Compared with the intelligent scheduling model 
obtained by Yang S. [10], this model could capture more 
complex data structures. The reason was that the model 
proposed in this article used a reasonable amount of 
SsSs. Compared with the detection method of Rani V et 
al. [12], the model proposed in this study had better 
robustness. The reason was that the feature extraction 
strategy of the model in this article could maintain the 
accuracy of the results. SSL automatically mined 
patterns from a large amount of unlabeled data through 
contrastive learning, and the test results on the CWRU 
dataset showed that the model's HR, DCG, and NDCG 
have improved by 5.6%, 6.4%, and 4.8%, respectively. 
This indicated that SSL could effectively identify 
potential faults in bearings under different operating 
conditions without relying on external labels for fault 
detection. Meanwhile, GNN enhanced the predictive 
ability of the model by capturing the complex 
interactions between devices. On the Phm IEEE 2012 
dataset, the NDCG metrics of the model improved by 
5.36% and 6.27% compared to the baseline model. It 
reflected the superior performance of GNN in 
processing high-dimensional data and predicted 
maintenance scenarios in real-world scenarios, 
accurately predicting possible equipment failures. 
Through experimental analysis, it was found that the 
model performance reaches its optimum when the 
parameters β=0.01, K=30, and L=2. This discovery 
not only validated the effectiveness of SSL and GNN in 
complex industrial scenarios, but also demonstrated that 
fine-tuning model parameters could significantly 
improve model performance. In the noise robustness 
experiment, the model was able to maintain a high 
NDCG index even after the addition of Gaussian noise, 
indicating that the model could maintain the accuracy 
and relevance ranking of recommendation results in 
noisy environments. In summary, the studied intelligent 
manufacturing optimization model had demonstrated 
significant performance improvement and advantages in 
practical applications by combining the technical 
features of SSL and GNN. The model had the capacity 
to automatically learn data features and capture complex 
relationships between devices. Furthermore, it could 
improve performance through parameter optimization, 
thereby providing an effective technical solution for the 
field of intelligent manufacturing. 

22   Informatica 49 (2025) 11–24 H. Wang 
6  Conclusion 
Aiming at the problem of innovative application of 
intelligent mechanical manufacturing, the research 
conducted a fusion processing based on SSL and GNN 
methods. Among them, considering the comparative 
damage determination of intelligent machinery, a 
comparative SSL model was constructed to utilize the 
data to automatically identify and predict the 
abnormalities of the machinery during operation. 
Meanwhile, heterogeneous graph and meta-path were 
introduced to accurately express the structural 
relationship, path length and other factors in the 
intelligent network. Then on the basis of SSL, it was 
fused with GNN for optimization, and the propagation 
rules of GNN were described to improve the level of 
intelligence to form a recommendation system. Finally, 
the optimization model architecture based on SSL and 
GNN intelligent manufacturing was derived. The results 
indicated that the research model improved HR, DCG 
and NDCG by 5.6%, 6.4%, and 4.8% over the NGCF 
model on both datasets, respectively. It also improved 
each performance by 10.8%, 8.5%, and 6.3% over the 
EHCF model, respectively. In the practical application 
of intelligent mechanical manufacturing, the research 
model exhibited a gain of 5.36% and 6.27% in NDCG 
metrics over the NGCF and EHCF models, respectively. 
There was a gain of 4.26% and 5.85% over the NDCG 
metrics of the NGCF and EHCF models, respectively. 
The enhancement of NDCG exerts a direct influence on 
the efficacy of decision-making processes within the 
production context, facilitating the optimization of 
resource allocation and the mitigation of unanticipated 
downtime. DCG has been shown to enhance the 
efficiency and output of the production line. Improving 
HR performance helps reduce production costs, improve 
user experience, and make products more competitive in 
the marketplace. The results of the study show that the 
research model is superior compared to other models in 
terms of dataset performance and practical application, 
and has good intelligent mechanical manufacturing 
application value. Moreover, it has contributed to the 
performance of user classification and recommendation 
systems. Although the model performed well in 
experiments, its scalability may be limited by data 
quality, computational resources, and model 
generalization ability when implemented in actual 
industrial environments. The incompleteness of 
industrial data, noise, and environmental differences can 
lead to a decrease in model performance. In addition, the 
high computational requirements of GNN and SSL 
models can be a challenge in resource-constrained 
environments. Therefore, the model needs to be adapted 
and optimized in new environments to ensure its 
efficiency and accuracy in different industrial scenarios. 
 
Funding 
2023 Henan Province Key Research and Development 
and Promotion Project (Science and Technology 
Research Project): Research and Application of a Virtual 
Reality-Based Visualization Platform for Maintenance 
Safety Knowledge of Electromechanical Equipment (No. 
232102321061) 
Acknowledgement 
2023 Henan Province Key Research and Development 
and Promotion Project (Science and Technology 
Research Project): Experimental Research on 
low-temperature Waste Heat Power Generation System 
based on radial inflow turbine (No. 232102320223) 
References 
[1] Jose J, Mathew V. (2024). Internet of Things-A 
Model for Data Analytics of KPI Platform in 
Continuous Process Industry. Informatica, 48(1). 
https://doi.org/10.1007/s10570-024-00699-x 
[2] Lv, Z., Guo, J., & Lv, H. (2023). Safety Poka Yoke 
in Zero-Defect Manufacturing Based on Digital 
Twins. IEEE Transactions on Industrial 
Informatics, 19, 1176–1184. 
https://doi.org/10.1109/TII.2021.3139897  
[3] Kumar, S., Gopi, T., Harikeerthana, N., Gupta, M., 
Gaur, V., Królczyk, G., & Wu, C. (2022). Machine 
learning techniques in additive manufacturing: a 
state-of-the-art review on design, processes and 
production control. Journal of Intelligent 
Manufacturing, 34(1), 21–55. 
https://doi.org/10.1007/s10845-022-02029-5 
[4] Dave, R., & Purohit, J. N. (n.d.). Leveraging Deep 
Learning Techniques to Obtain Efficacious 
Segmentation Results. Archives of Advanced 
Engineering Science, 1(1), 11-26. 
https://doi.org/10.47852/bonviewaaes32021220  
[5] Ravni čan J, Marinko A, Noveski G, et al. (2022). A 
Prestudy of Machine Learning in Industrial Quality 
Control Pipelines. Informatica, 46(2). 
https://doi.org/10.1007/s10570-022-00665-2 
[6] He H. (2024). Automatic Network Traffic 
Scheduling Algorithm Based on Deep 
Reinforcement Learning. Informatica, 48(22). 
https://doi.org/10.1007/s10570-024-00799-x 
[7] Sharma T, Kaur K. (2023). A Deep Learning-fuzzy 
Based Hybrid Ensemble Approach for Aspect 
Level Sentiment Classification. Informatica, 47(6). 
https://doi.org/10.1007/s10570-023-00694-5 
[8] Yao, X., Ma, N., Zhang, J., Wang, K., Yang, E., & 
Faccio, M. (2022). Enhancing wisdom 
manufacturing as industrial metaverse for industry 
and society 5.0. Journal of Intelligent 
Manufacturing. 35, 235–255. 
https://doi.org/10.1007/s10845-022-02027-7 
[9] Chihi, I., & Sidhom, L. (2023). Fault diagnosis and 
self-healing for smart manufacturing: a review. 
Journal of Intelligent Manufacturing. 35, 
2441–2473. 
https://doi.org/10.1007/s10845-023-02165-6 
[10] Yang, S., & Xu, Z. (2021). Intelligent scheduling 
and reconfiguration via deep reinforcement 
learning in smart manufacturing. International 

Innovative Application of Intelligent Mechanical Manufacturing Based… Informatica 49 (2025) 11–24 23 
Journal of Production Research. 60(16), 
4936–4953. 
https://doi.org/10.1080/00207543.2021.1943037 
[11] Krishnan, R., Rajpurkar, P., & Topol, E. J. (2022). 
Self-supervised learning in medicine and 
healthcare. Nature Biomedical Engineering, 6(12), 
1346–1352. 
https://doi.org/10.1038/s41551-022-00914-1 
[12] Rani, V. L., Nabi, S. T., Kumar, M., Mittal, A., & 
Kumar, K. G. (2023). Self-supervised Learning: A 
Succinct Review. Archives of Computational 
Methods in Engineering, 30(4), 2761–2775. 
https://doi.org/10.1007/s11831-023-09884-2 
[13] Kumar, P., Rawat, P., & Chauhan, S. (2022). 
Contrastive self-supervised learning: review, 
progress, challenges and future research directions. 
International Journal of Multimedia Information 
Retrieval, 11(4), 461–488. 
https://doi.org/10.1007/s13735-022-00245-6 
[14] Wu S, Sun F, Zhang W, Xie X, Cui B. Graph 
Neural Networks in Recommender Systems: A 
Survey. (2022). ACM Computing Surveys, 55(5), 
1–37. https://doi.org/10.1145/3535101 
[15] Zhou Y, Zheng H, Huang X, Hao S, Li D, Zhao J. 
Graph Neural Networks: Taxonomy, Advances, 
and Trends. (2022). ACM Transactions on 
Intelligent Systems and Technology, 13(1), 1–54. 
https://doi.org/10.1145/3495161  
[16] Liu, S., Zheng, P., & Bao, J. (2023). Digital 
Twin-based manufacturing system: a survey based 
on a novel reference model. Journal of Intelligent 
Manufacturing. 35, 2517–2546. 
https://doi.org/10.1007/s10845-023-02172-7 
[17] Wang, Y., Wang, J., Cao, Z., & Farimani, A. B. 
(2022). Molecular contrastive learning of 
representations via graph neural networks. Nature 
Machine Intelligence, 4(3), 279–287. 
https://doi.org/10.1038/s42256-022-00447-x  
[18] Dong, G., Tang, M., Wang, Z., Gao, J., Guo, S., 
Cai, L., Gutierrez, R., Campbell, B., Barnes, L. E., 
& Boukhechba, M. (2022). Graph Neural Networks 
in IoT: A Survey. ACM Transactions on Sensor 
Networks, 19(2), 1–50. 
https://doi.org/10.1145/3565973 
[19] Liu Z, Yang L, Fan Z, Peng H, Yu P S. Federated 
social recommendation with graph neural network. 
ACM Transactions on Intelligent Systems and 
Technology (TIST), 2022, 13(4): 1-24. 
https://doi.org/10.1145/3501815 
[20] Zheng, C., Chen, H., Cheng, Y., Song, Z., Wu, Y., 
Li, C., Cheng, J., Yang, H., & Zhang, S. (n.d.). 
ByteGNN: Efficient Graph Neural Network 
Training At large Scale. Proceedings of The Vldb 
Endowment, 15, 1228–1242. 
https://doi.org/10.14778/3514061.3514069 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

24   Informatica 49 (2025) 11–24 H. Wang