INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 521 www.rsisinternational.org

Enhancing Reinforcement Learning through Graph Neural
Networks: A Novel Approach

1Priya Singh, 2Dr. Md. Abdul Aziz Al Aman, 3Dr. Saroj Kumar

1Department of Computer Science Techno India University Kolkata, West Bengal

2Department of CSE-AI Techno India University Kolkata, West Bengal

3AiLabs-Artificial Intelligence, DGC DataCore Systems (India) pvt. Ltd. Kolkata, West Bengal

DOI: https://doi.org/10.51244/IJRSI.2025.1210000046

Received: 10 October 2025; Accepted: 14 October 2025; Published: 01 November 2025

ABSTRACT

Reinforcement Learning (RL) has showcased remarkable success in various domains. However, its performance
often degrades in the environment with complex structures and distributed rewards. Graph-Based Reinforcement
Learning (GBRL) is an approach that combines the strengths of Graph Theory with Reinforcement Learning to
optimize complex decision making problems in any networked system. This paper proposes an approach of
integrating Reinforcement Learning approaches with Graph Neural Networks (GNNs)to enhance the learning
pipeline and model structured data by utilising their capacity. We present an approach that uses GNNs
represented as graphs that enables RL agents to get dependencies between entities and access information
through them. This paper exhibits GBRL techniques and their application in different domains. A framework of
GBRL methods and its advantages over RL methods in working on graph-based data. This work highlights the
synergy between graph-based learning and decision-making, offering a promising direction for solving high-
dimensional and structured RL tasks more effectively. We also summarize the key challenges and the open
research directions in this field.

Keywords– Graph-Based Reinforcement Learning (GBRL), Reinforcement Learning (RL), Graph Neural
Networks (GNN), Convolution Neural Networks (CNNs), Deep Reinforcement Learning (DRL), Temporal
Difference (TD), Asynchronous Advantage Actor-Critic (A3C), Advantage Actor-Critic (A2C), Deterministic
Policy Gradient (DPG), Deep Deterministic Policy Gradient (DDPG), Multi-Relational Graph (MR-Graph),
Multi-Relational GNN, Markov’s Decision Process (MDP), Monte Carlo (MC) methods, Deep Q-Network
(DQN)

INTRODUCTION

Recently the development of networked systems has boosted the interest in the methods that can efficiently
handle complex and structured data which these systems generate. GBRL emerges as an encouraging method
that utilises the power of reinforcement learning with graph theory to address decision making problems. The
relationships within entities represented as graphs, GBRL allows refine understanding of problem space,
enabling an effective learning and decision making. Conventional RL approaches struggle with interconnection
and complexity present in graph structured data. GBRL deals with these challenges by using graph methods with
RL algorithms, giving better exploitation and exploration of the network structure. This has tremendously
implicated various domains, like biological systems, social networks, communication networks, and
transportation systems, where optimization and understanding between interactions among entities is crucial.
This paper aims to give a broad overview of GBRL, exploring its approaches, potential future directions and
applications. The paper begins with the fundamental concepts of RL and graph theory, followed by the

INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 522 www.rsisinternational.org

Figure 1: Graph Based Reinforcement Learning (GBRL) : Goals with problems that GBRL solves details of
GBRL techniques. With this paper, we intend to highlight advantages of GBRL and identify key challenges with
the proposed strategies for advancing this field.

A. Graph Based Reinforcement Learning

The GBRL techniques pertinence are broad, as they exhibit versatility of the graph frameworks formed of
interconnected entities with their relationships. Beyond the description of the problems, it is normal to address
questions on optimization, where the goal is intervening the system to improve properties. The requirements for
GBRL can be summarized as:

1. Graphs are appropriate for problems under consideration, with edges and nodes having clear semantics.

2. Decision-making is possible to intervene in systems beyond mere observation.

3. To solve a problem approximately is acceptable.

LITERATURE REVIEW

The progress of GNNs enables effective representation learning in various fields [1] including social networks,
NLP [2, 3], recommender system [4], social events [5, 6, 7], computer vision and physics [8]. GNN models can
show optimal performance over huge datasets on different tasks like node clustering [9, 10], link prediction [11,
12, 13], graph classification [14, 15, 16, 17] etc. As per the difference in modeling for actual graph data, we can
divide GNN methods into homogeneous GNN, heterogeneous GNN, and multiple graph learning models.
Homogeneous GNN are those methods that overlook data type of nodes or attributes of edges on graphs.
Classical methods include GCN [18], Graph-SAGE [19] and GAT [20]. Heterogeneous GNN approaches
consider the heterogeneity of node or edge types. Considering diverse edges in actual data, more relational GNN
including R-GCN [21], FdGars [22], SemiGNN [23] and GraphConsis [24] were developed. Apart from
homogeneous and heterogeneous GNNs, multi-graph neural network models [25, 26, 27, 28, 29] fuse the
multiple characterizations to learn the embedding of graph data. Recently, GNNs have been used to compute
and model graph-based data to predict relationships in graphs and improve the ability to reason. These models
establish rules for graph-based data, from biological network analysis to social media, optimization and network
modeling [30]. GNN emerges as a tool to solve graph mining problems. With the advancement of technology,
RL algorithms have generated many directions of development. The value based basic algorithms are Q-
Learning algorithms [31] and DQN [32] algorithms, that use value functions to gauge and reduce occurrence of
the optimal situations. whereas policy-based algorithms directly perform iterative calculations on the policy such
as PPO [33]. The Actor-Critic RL methods combine advantages of both value-based and policy-based. Formerly
there were a few attempts to integrate GNNs and RL. A model, DGN+GNN [34] is used to generalise network
topologies, where GNNs allow the RL agents to operate on different networks. A RL based graph model,
G2S+BERT+RL [35] is for natural question generation, GNN is made to process the directed graph passage.
Other works [36, 37, 38] investigate how GNNs improve the generalisation ability of RL. Also there are
numerous studies that show RL can be optimized on graphs. DeepPath [39], GraphNAS [40], Policy-GNN [41],
RL-HGNN [42] are a few examples.

INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 523 www.rsisinternational.org

METHODOLOGIES AND APPROACHES

For this subsection, the goal is the introduction of the RL algorithms and to draw their connections to graphs.
Graph RL methods presented rely on a huge variety of algorithms, each of which is characterized by distinct
principles and assumptions. It is rather not possible to cover all algorithms, it is necessary to give some of them
for solving MDPs.

A. Dimensions of RL Algorithms

There are different RL methodologies used in different applications.

Model-based algorithm and Model-free algorithm: To be specific, let's assume state space (S) and action space
(A) known as a model M = (P, R) refers to having an estimate of transition function (P) and reward functions
(R). Model-based algorithms can add knowledge to greatly optimize learning speed up. It can be used in the
form of mathematical set descriptions that define P and R fully. It is beneficial where authenticity is expensive
to generate and where there is a negative impact on executing poor policies (e.g., robotics). When equipped with
the perspective of the world, policy is planned by the agent, either decision-time planning, or background
planning. In Model-free algorithms, P is corresponding to the density estimation problem, while R is the
Supervised Learning problem. It has simple learning architectures. This has higher model complexity, i.e. they
need more interactions with the environment to train.

On-policy and off-policy: The differentiation relies on the occurrence of these two policies: behavior policy
which is used to interact with the environment and target policy is the policy that is being learned. The behavior
and target policy are same in on-policy methods while they are different in off-policy algorithms. On-policy
methods are used as special cases while off-policy methods are more flexible.

Sample-based and Temporal Difference: The sample-based methods rely upon environmental interaction
samples, rather than complete knowledge of MDP. Temporal Difference (TD) methods are based on samples of
experience. They are biased with less variance.

RL Methods

Recently, RL methods have achieved a huge success in varied applications that automatically handles sequential
decision based problems in environments via decision making and goal directed learning, these types of methods
achieve huge success in many games. RL is developed as an MDP, a sequential-decision making mathematical
model whose actions affect current rewards, the subsequent states with future rewards. In MDP, prediction
problems and control problems are implemented using dynamic programming.

The definition of MDP says the tuple {S, A, T , R, p(s0), γ}, where S is set of all possible states, which are
generalization of environment, A means set of actions which are adopted in states, which are agent’s all possible
actions, R : S×A×S→R shows reward function, which are rewards returned by environment to agent after an
action’s execution, T : S×A→ p(S) is denoted by state transition function, γ ∈ [0, 1] indicates the discount factor,
that is treated as hyperparameter of agent and uses to promote faster availability of rewards to agent. Figure 2
shows the process of agent’s interaction with the environment.

Here, to find a policy π is our goal that maximises Q(s, a) the expected action-value function, (1) defines the
target policy.

��∗ = ��
��

��(��, ��)

= ��
��

��,�� [∑��
��=0 ��+��|�� = ��, �� = ��] (1)

where ��∗ is equal or better than all other policies. Further we provide fundamental concepts of the different
popular RL methods.

INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 524 www.rsisinternational.org

Figure 2. Process of interaction between the agent and the environment.

1) Q-Learning: Q-learning and Temporal Difference (TD) learning methods are off-policy learners, which are
an important result to early research on RL. The values on the Q-table are directly updated by target policy in
Q-learning to achieve selection of optimal policy, whereas behavior policy employs ε-greedy policy for the semi-
random exploration of environment. The action-value function Q has a learning goal to be learned in Q-learning.
By direct approximation the optimal action-value function q∗ are learned, so that Q-learning algorithm can
formulate as (2).

��(��, ��) ← ��(��, ��) + ��[��+1 + ��
��

��(��+1, ��) − ��(��, ��)] (2)

where the equation (2) denotes state st which explores an environment with behavior policy that is based on
values in Q-table at timestamp t. It performs an action a, the reward as R and a new state as ��+1 is obtained
based on environmental feedback. The given equation has been executed to update the latest Q-table. It will
continue updation of new state ��+1, after completion of the operation until termination time t.

2) Reinforce: REINFORCE doesn’t optimize directly based on policy space, rather learns parameterized policy
without being known about the intermediate value estimation function, that uses Monte Carlo method to learn
policy parameters with estimated returns and full trace. This method makes a neural-network based policy which
takes inputs as states and gives output as probability distribution in operation space. This policy π parameterize
θ a set of weights so that π(s; θ) ≡ π(s), that is action probability distribution on state, and REINFORCE is
updated (3).

��,�� = ��,�� (�� − ��,��)
��

��,��
�� (3)

where a non-negative learning factor is ��,��, discounted reward value is denoted by r, and representation function
of state ��,�� is for reducing variance of gradient estimate.

3) Actor-Critic: Actor-Critic algorithm uses a value function and a parameterized policy, where a better Â(��, ��)
estimate of computation of the policy gradient is provided by the value function. This algorithm elaborates both
state and policy value functions by combining advantages of policy gradient and value function based
algorithms. The Actor in this method, denotes policy function for learning policy that can be obtained as any
number of rewards as possible. Whereas the Critic represents an estimated value function that is used for
evaluating estimated value of current policy. Figure 3 here showcases the framework of the Actor-Critic
algorithm. For the RL algorithms, the basic framework like A3C, A2C, DPG, DDPG is the Actor-Critic
Architecture.

4) Deep Q-Network: Q-learning algorithm along with Q-table in large scale graph-structured data suffers from
the problem of large number of intermediate state values, causing dimensional disaster. To address this
challenge, a method is proposed which combines neural networks and value-function approximation through
action or state space reduction and function approximation. DQN is misused to learn the policies by the
leveraging of deep neural networks. It is proposed to use Q-learning algorithms for representing states with graph

INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 525 www.rsisinternational.org

embeddings, according to properties of graph-structured data, to minimize impact of non-Euclidean structure on
Q-table scale. In general, DQN combines Q-learning with deep learning and approximates action-value function
with deep neural-networks,

Figure 3. Framework for Actor-Critic algorithm. Here from the given environment, the Actor receives a state
and selects the action to be performed. In the meantime, Critic receives its current state and state that is generated
by previous interactions and calculates TD error to update Actor and Critic.

and obtains {st, at, rt, st+1} the trace from interaction of agent with environment. It can implement successful
policies from high dimensional sensory inputs by using the end-to-end RL. The loss function L(θt) of this method
is given by (4).

��(��) = ��[(�� + ��
��+1

��−1
(��+1, ��+1) − ��

(��, ��))2](4)

To stabilize the training process DQN gives two techniques: (1) A buffer replay to reuse past experiences. (2) A
separate target-network that can be periodically updated.

B. Graph Embedding Methods

Graph Neural Networks term means a deep embedding method. Let's recall the graph G = (V, E) in which vi
nodes have feature vectors ��

and, optionally, with ��.��
as edge features. The goal here is to get ℎ��

an embedding
vector for each of the nodes that captures features along with structure of interactions on a given graph. The
calculation or generation of the embedding vectors takes place in l ∈ 1, 2, ..., L layers, where L represents the
final layer. ℎ��

(��)
can be used to denote the embedding of vi nodes in the l layer. Notation W(l) , indexed by the

subscript, represents a weight matrix which depicts the block of learnable parameters in the layer l of the Graph
Neural Network model. Unless the embeddings were initialized with node features, ℎ��

(0)
= ��

, ∀ vi ∈ V.

1) Message Passing Neural Network: Message Passing Neural Network (MPNN) (Gilmer et al., 2017) is the
framework which abstracts many of the graph learning architectures, along with serving as one of the useful
conceptual models present for deep embedding methods. It has layers that apply M(l) a message function and U
(l) a vertex-update function, to compute embeddings.

��
(��+1)

= ∑
��Ɲ(��)

��(��)(ℎ��
(��)

, ℎ��
(��)

, ��,��
)

ℎ��
(��+1)

= ��(��)(ℎ��
(��)

, ��
(��+1)

) (5)

INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 526 www.rsisinternational.org

where Ɲ(vi) represents the open neighborhood of node vi . Afterwards, I the readout function is applied to
calculate

embedding for entire graph from the final node embeddings set: ��({ℎ��
(��)

|})I({h (L)vi|vi ∈ V }). The message
function and vertex-update functions are learned. The readout function can either be fixed a priori or learned.

2)Graph Convolutional Network: Graph Convolutional Network method (GCN) (Kipf & Welling, 2017) is
simpler to merely rely on multiplication of features of nodes with weight matrix together with degree-based
normalization. It is a first-order approximation of local spectral filters on the graphs. As it can be created as the
series of matrix multiplication, it scales well to the large graphs with millions of edges while giving superior
performance to the other methods of embedding at the time. It can be formulated as:

��
(��+1)

= ��(��1
(��) ∑��Ɲ[��]

ℎ��
(��)

√(1+��(��)(1+��(��))
) (6)

where degree of node vi is indicated by deg(vi), and closed neighborhood of node vi by Ɲ[��], including all its
neighbors as well as vi itself.

3) Graph Attention Network: Graph Attention Network model (GAT) (Veličković et al., 2018) proposed use of
the attention mechanisms (Bahdanau et al., 2016), it is considered as a way to do flexible aggregation on the
neighbor features. Coefficients of learnable aggregation enables an increase in the expressibility of the model,
which translates gains in the predictive performances over GCN for the classification of nodes. Let ��,��

(��)
be

denoted as the attention coefficient that shows importance of features of node vj to the features of node vi in the
layer l. It is computed as:

��,��
(��)

=
��( ��(��[��1

(��)
ℎ��

(��)
||��1

(��)
ℎ��

(��)
||��2

(��)
��,��

]))

∑
��∈Ɲ[��]

��(��(��[��1
(��)

ℎ��
(��)

||��1
(��)

ℎ��
(��)

||��2
(��)

��,��
))

(7)

Where the exponential function is represented by exp(x) = ex, weight vector is θ that parametrize attention
mechanism and concatenation is denoted by [·∥·]. The activation function LeakyReLU(x), which gives the non-
zero values for negative inputs, according to αLR the small slope, is equal to αLRx if x < 0, and x otherwise. For
the given attention coefficient, node embeddings can be calculated according to the rule given below.

ℎ��
(��+1)

= ∑��∈Ɲ[��] ��,��
(��)

��1
(��)

ℎ��
(��)

(8)

Connections Between GNN and RL

This section will discuss the relationship between GNN and RL in context to GBRL framework. Learning
techniques that are designed for the operations on graphs are used commonly as the function approximators
which are considered as part of RL algorithms. GBRL methods are goal driven and constructive, which allows
for the flexibility in finding embeddings relevant to the objective function that is to be optimized without any
fine-grained supervision signal. Alternatively, primitive graph learning benchmark to rely on the supervised
learning and availability of granular examples. The work on molecular optimization is an exception (You et al.,
2018a), which has used a GCN as a differentiator trained on the example molecules which provide a part of the
reward signal. It is noteworthy as another line of recent work that connects GNNs and RL. In this work, a
representation of a graph is constructed where nodes are states, and edges represent transitions to be determined
by actions.

CHALLENGES AND GAPS

Irrespective of the potential of GBRL, several gaps and challenges in this research are needed to be addressed.
These challenges highlight the opportunities for further improvement of GBRL methods:

INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 527 www.rsisinternational.org

1) Scalability: Scalability to large-scaled graph-structured environments is a challenge of GBRL algorithms.
Processing huge amounts of data efficiently and managing its computational complexity remains its significant
hurdles.

2) Learning in Dynamic Environments: Many graphs are dynamic in the real world, with edges and nodes
evolving over time. Recent GBRL methods struggle in adapting these changes, demanding algorithms capable
of handling dynamic graphs for the development.

3) Generalization and Transfer Learning: Essuring GBRL models to generalize within different domains and
tasks is crucial. Difficulties faced by current approaches in transferring knowledge to a new and unseen
environment, with spotlighted need to improve transfer learning techniques.

4) Explainability and Interpretability: With increasingly complex GBRL models ensuring their decisions are
interpretable and explainable for practical applications. Remains a significant challenge is enhancing
transparency.

5) Quality and Data Sparsity: For effective learning high-grade with broad scope of graph data is essential.
However, noise, data sparsity or missing information can inhibit performance of GBRL models. It is critical to
develop methods for handling these data issues.

6) Multi-Agent Coordination: In multiple agents involved scenarios, ensuring an effective collaboration in a
graph-structured environment and coordinating the actions of these agents can be challenging.

7) Evaluation and Benchmarking: For GBRL, the deficiency of standard criteria and evaluation metrics makes
it difficult comparing different approaches and measuring progress. Making comprehensive criteria frameworks
is necessary for the development of the field.

8) Societal and Ethical Implications: As implementation of GBRL techniques to real-world problems, it is
crucial to understand and address their societal and ethical implications. Ensuring accountability, fairness and
reducing biases in GBRL models are the most important considerations.

With these challenges and gaps addressed, the GBRL techniques can progress in developing more practical,
scalable and robust solutions for complex decision-making problems.

FUTURE RESEARCH SCOPE

GBRL has great capabilities for addressing different kinds of complex real-world problems. Whereas, there are
several open questions and challenges that need further investigation. The below given areas represent various
directions for future research in the field of GBRL:

1) Efficiency and Scalability: It is crucial to develop algorithms which can further scale to large and complex
graph-structured environments. Optimization of computational resources and improvement of scalability of
GBRL methods should be future research focus.

2) Generalization and Transfer Learning: An important area to explore is enhancement of capabilities of GBRL
algorithms in transferring learned knowledge within different domains along with the generalisation of new and
unseen environments. Investigation techniques for new domain adaptations and transfer learning will also help
in achieving this goal.

3) Multiple Agent Systems: GBRL extension to multiple agent systems, where agents learn and interact
simultaneously within these graph-structured environments, presents a unique challenge. Further research in
these directions can lead to advancement in coordination among agents and collaborative decision-making.

4) Dynamic Graphs: As many real-world graphs are dynamic in nature, with edges and nodes changing over
time. Development of GBRL methods that can adapt such dynamic graphs while learning from data is a thrilling
research direction.

INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 528 www.rsisinternational.org

5) Interpretability and Explainability: For practical applications, ensuring the transparency of GBRL models
and their decision-making processes is essential. Future work must focus on enhancing the explainability of the
GBRL algorithms making them much more trustworthy and more accessible to users.

6) Real-World Applications: Further research in the emerging fields such as cybersecurity, healthcare, smart
cities etc is needed to explore the potential of GBRL.

CONCLUSION

GBRL represents a remarkable advance in the machine learning domain, offering distinct capabilities for
handling complicated problems of decision-making in graph environments. By combining the concepts of RL
with graph theory, GBRL enables efficient exploitation and exploration of relationships within a network or
networks, enhancing performances across applications. This paper highlights the principles, methodologies and
various applications of GBRL. It also identifies the key challenges and gaps in current research, emphasising on
the need for adaptive, scalable and interpretable GBRLs. The way to more robust and practical implementation
of GBRL can be addressed through these challenges in real-world scenarios. In conclusion, GBRL provides a
framework to better understand and optimize complex networked systems. As the research progresses in this
area, we predict that GBRL plays a pivotal role in the addressing of some pressing challenges in this
interconnected world.

REFERENCES

1. Christopher JCH Watkins and Peter Dayan. 1992. Q-learning. Machine learning 8, 3-4 (1992), 279–
292.

2. David K Duvenaud, Dougal Maclaurin, Jorge Iparraguirre, Rafael Bom barell, Timothy Hirzel, Al´an
Aspuru-Guzik, and Ryan P Adams. 2015. Convolutional networks on graphs for learning molecular
fingerprints. In Proceedings of the NIPS. 2224–2232.

3. Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare,
Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. 2015. Human-level control
through deep reinforcement learning. nature 518, 7540 (2015), 529–533.

4. William L Hamilton, Rex Ying, and Jure Leskovec. 2017. Inductive representation learning on large
graphs. In Proceedings of the 31st International Conference on Neural Information Processing Systems.
1025–1035.

5. Thomas N Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional
networks. Proceedings of the ICLR.

6. Guixiang Ma, Lifang He, Chun-Ta Lu, Weixiang Shao, Philip S. Yu, Alex D Leow, and Ann B Ragin.
2017. Multi-view clustering with graph embedding for connectome analysis. In Proceedings of the 2017
ACM on Conference on Information and Knowledge Management. 127–136.

7. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy
optimization algorithms. arXiv preprint arXiv:1707.06347 (2017).

8. Wenhan Xiong, Thien Hoang, and William Yang Wang. 2017. DeepPath: A Reinforcement Learning
Method for Knowledge Graph Reasoning. In Proceedings of the EMNLP. ACL, 564–573.

9. Scott Fujimoto, Herke van Hoof, and David Meger. 2018. Addressing Function Approximation Error in
Actor-Critic Methods. In Proceedings of the 35th International Conference on Machine Learning
(Proceedings of Machine Learning Research, Vol. 80). PMLR, Stockholmsm¨assan, Stockholm Sweden,
1587–1596.

10. Seyed Mehran Kazemi and David Poole. 2018. Simple embedding for link prediction in knowledge
graphs. In Proceedings of the NIPS. 4284–4295.

11. Tengfei Ma, Cao Xiao, Jiayu Zhou, and Fei Wang. 2018. Drug similarity integration through attentive
multi-view graph auto-encoders. In Proceedings of the 27th International Joint Conference on Artificial
Intelligence. 3477–3483.

12. Shirui Pan, Ruiqi Hu, Guodong Long, Jing Jiang, Lina Yao, and Chengqi Zhang. 2018. Adversarially
regularized graph autoencoder for graph embedding. In Proceedings of the IJCAI. AAAI Press, 2609–
2615.

INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 529 www.rsisinternational.org

13. Hao Peng, Jianxin Li, Yu He, Yaopeng Liu, Mengjiao Bao, Lihong Wang, Yangqiu Song, and Qiang
Yang. 2018. Large-scale hierarchical text classification with recursively regularized deep graph-cnn. In
Proceedings of the WWW. 1063–1072.

14. Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, Rianne Van Den Berg, Ivan Titov, and Max
Welling. 2018. Modeling relational data with graph convolutional networks. In Proceedings of the
ESWC. Springer, 593–607.

15. Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua
Bengio. 2018. Graph attention networks. Pro ceedings of the ICLR.

16. Zhitao Ying, Jiaxuan You, Christopher Morris, Xiang Ren, Will Hamilton, and Jure Leskovec. 2018.
Hierarchical graph representation learning with differentiable pooling. In Proceedings of the NIPS.
4800–4810.

17. Muhan Zhang and Yixin Chen. 2018. Link prediction based on graph neural networks. In Proceedings of
the NIPS. 5165–5175.

18. Xi Zhang, Lifang He, Kun Chen, Yuan Luo, Jiayu Zhou, and Fei Wang. 2018. Multi-view graph
convolutional network and its applications on neuroimage analysis for parkinson’s disease. In AMIA
Annual Symposium Proceedings, Vol. 2018. 1147.

19. Paul Almasan, Jos´e Su´arez-Varela, Arnau Badia-Sampera, Krzysztof Rusek, Pere Barlet-Ros, and
Albert Cabellos-Aparicio. 2019. Deep reinforcement learning meets graph neural networks: Exploring a
routing optimization use case. arXiv preprint arXiv:1910.07421 (2019).

20. Yu Chen, Lingfei Wu, and Mohammed J Zaki. 2019. Reinforcement Learning Based Graph-to-Sequence
Model for Natural Question Genera tion. In Proceedings of the ICLR.

21. Yang Gao, Hong Yang, Peng Zhang, Chuan Zhou, and Yue Hu. 2019. Graphnas: Graph neural
architecture search with reinforcement learning. arXiv preprint arXiv:1904.09981 (2019).

22. Di Jin, Ziyang Liu, Weihao Li, Dongxiao He, and Weixiong Zhang. 2019. Graph convolutional networks
meet markov random fields: Semi supervised community detection in attribute networks. In Proceedings
of the AAAI. AAAI Press, 152–159.

23. Kai Lei, Meng Qin, Bo Bai, Gong Zhang, and Min Yang. 2019. GCN GAN: A non-linear temporal link
prediction model for weighted dynamic networks. In Proceedings of the IEEE INFOCOM. 388–396.

24. Hao Peng, Jianxin Li, Qiran Gong, Yangqiu Song, Yuanxing Ning, Kunfeng Lai, and Philip S. Yu. 2019.
Fine-grained event categorization with heterogeneous graph convolutional networks. In Proceedings of
the IJCAI. AAAI Press, 3238–3245.

25. Daixin Wang, Jianbin Lin, Peng Cui, Quanhui Jia, Zhen Wang, Yanming Fang, Quan Yu, Jun Zhou,
Shuang Yang, and Yuan Qi. 2019. A Semi supervised Graph Attentive Network for Financial Fraud
Detection. In Proceedings of the IEEE ICDM. 598–607.

26. Jianyu Wang, Rui Wen, Chunming Wu, Yu Huang, and Jian Xion. 2019. Fdgars: Fraudster detection via
graph convolutional networks in online app review system. In Proceedings of the World Wide Web
Conference. 310–316.

27. Victor Bapst, Thomas Keck, A Grabska-Barwi´nska, Craig Donner, Ekin Dogus Cubuk, Samuel S
Schoenholz, Annette Obika, Alexander WR Nelson, Trevor Back, Demis Hassabis, et al. 2020. Unveiling
the predictive power of static structure in glassy systems.

28. Patrick Hart and Alois Knoll. 2020. Graph Neural Networks and Rein forcement Learning for Behavior
Generation in Semantic Environments.

29. Jarom´ır Janisch, Tom´aˇs Pevny, and Viliam Lis ‘ y. 2020. Symbolic Relational Deep Reinforcement
Learning based on Graph Neural Networks. arXiv preprint arXiv:2009.12462 (2020).

30. Kwei-Herng Lai, Daochen Zha, Kaixiong Zhou, and Xia Hu. 2020. Policy-GNN: Aggregation
Optimization for Graph Neural Networks. In Proceedings of the ACM SIGKDD. New York, NY, USA,
461–471. Reinforced Neighborhood Selection Guided Multi-Relational Graph Neural Networks • 39:43

31. Zhiwei Liu, Yingtong Dou, Philip S. Yu, Yutong Deng, and Hao Peng. 2020. Alleviating the
Inconsistency Problem of Applying Graph Neural Network to Fraud Detection. Proceedings of the
SIGIR, 1569–1572.

32. Hao Peng, Jianxin Li, Qiran Gong, Yuanxin Ning, Senzhang Wang, and Lifang He. 2020. Motif-
Matching Based Subgraph-Level Attentional Convolutional Network for Graph Classification. In
Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 34. 5387–5394.

INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue X October 2025

Page 530 www.rsisinternational.org

33. Ruihong Qiu, Zi Huang, Jingjing Li, and Hongzhi Yin. 2020. Exploiting Cross-session Information for
Session-based Recommendation with Graph Neural Networks. ACM Transactions on Information
Systems (TOIS) 38, 3 (2020), 1–23.

34. Junkai Sun, Junbo Zhang, Qiaofei Li, Xiuwen Yi, Yuxuan Liang, and Yu Zheng. 2020. Predicting
citywide crowd flows in irregular regions using multi-view graph convolutional networks. IEEE
Transactions on Knowledge and Data Engineering (2020).

35. Penghao Sun, Julong Lan, Junfei Li, Zehua Guo, and Yuxiang Hu. 2020. Combining deep reinforcement
learning with graph neural networks for optimal VNF placement. IEEE Communications Letters 25, 1
(2020), 176–180.

36. Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S. Yu Philip. 2020. A
comprehensive survey on graph neural networks. IEEE Transactions on Neural Networks and Learning
Systems (2020).

37. Zhiqiang Zhong, Cheng-Te Li, and Jun Pang. 2020. Reinforcement Learning Enhanced Heterogeneous
Graph Neural Network. arXiv preprint arXiv:2010.13735 (2020).

38. Yuwei Cao, Hao Peng, Jia Wu, Yingtong Dou, Jianxin Li, and Philip S. Yu. 2021. Knowledge-Preserving
Incremental Social Event Detection via Heterogeneous GNNs. In Proceedings of the Web Conference
2021. Association for Computing Machinery, 3383–3395.

39. [39] Hao Peng, Jianxin Li, Yangqiu Song, Renyu Yang, Ranjan Rajiv, Philip S. Yu, and He Lifang. 2021.
Streaming Social Event Detection and Evolution Discovery in Heterogeneous Information Networks.
ACM Transactions on Knowledge Discovery from Data 15, 5 (2021).

40. Hao Peng, Jianxin Li, Senzhang Wang, Lihong Wang, Qiran Gong, Renyu Yang, Bo Li, Philip S. Yu,
and Lifang He. 2021. Hierarchical taxonomy-aware and attentional graph capsule RCNNs for large-scale
multi-label text classification. IEEE Transactions on Knowledge and Data Engineering 33, 6 (2021),
2505–2519.

41. Qingyun Sun, Hao Peng, Jianxin Li, Jia Wu, Yuanxing Ning, Phillip S. Yu, and Lifang He. 2021.
SUGAR: Subgraph Neural Network with Reinforcement Pooling and Self-Supervised Mutual
Information Mechanism. In Proceedings of the Web Conference. 2081–2091.

42. Yang Wang. 2021. Survey on Deep Multi-Modal Data Analytics: Collaboration, Rivalry, and Fusion.
ACM Transactions on Multimedia Computing, Communications, and Applications 17, 1s, Article 10
(2021), 25 pages. In IEEE Intelligent Vehicles Symposium (IV). IEEE, 1589–1594