IEEE Computational Intelligence Magazine - February 2023 - 41

Graph lifelong learning aims to enable
models to acquire new information based
on changes to the evolving graph's
structure and refinement of previous
knowledge in order to handle new
unseen tasks on graph-based domains.
gradient episodic memory (GEM) [23]. Those methods are
combined directly with three different base methods of graph
representation learning without any knowledge consolidation,
such as GCNs [27], GAT [78], and graph isomorphism network
(GIN) [80]. The scenarios are to learn new tasks based
on adding new data and classes in some datasets such as
Cora [75], Citeseer [75], Actor [81], etc. The result shows that
HPNs get the best average accuracy of93.7 1.5 in the Cora
dataset. The forgetting rate also gets the least rate compared
with other methods, with -0.9 0.9 in the Actor dataset.
HPNs not only perform better in terms of performance but
also have good memory efficiency.
IV. REHEARSAL APPROACH
This approach regulates retraining processes for previous tasks
to strengthen the relationship between memory and performance
on previously learned tasks [50], [51]. This graph learning
approach enables the selection of appropriate samples of graph
representation, such as nodes and edges, for retraining purposes.
The number of samples is carefully considered to minimize the
computational complexity. Recently, graph lifelong learning
methods using rehearsal approaches have been proposed,
such as ER-GNN [9], and lifelong Open-world Node Classification
[36]. Other graph lifelong learning approaches, such as
ContinualGNN [10] and TrafficStream [44], that combine
aspects ofthe rehearsal approach with other different approaches
are discussed in Section VI.
A. Experience Replay GNN Framework
Zhou et al. [9] proposed the concept of lifelong learning on
graph data using graph-based experience replay called ERGNN.
The main objective is to mitigate the catastrophic
forgetting problem in graph data while learning continuously
over time. The model preserves past knowledge in an
experienced buffer obtained from previous learning of tasks
to be replayed while learning a new task. This model implements
an experience node strategy to help to select the
experience nodes to be stored in the experience buffer,
such as the mean of feature (MF) [51] and coverage maximization
(CM) [82]. Apart from those experience node
strategies, ER-GNN proposed a new method called influence
maximization (IM).
The model's experience node replay is influenced by
Complementary Learning System (CLS) that supports the
biological learning process in humans. CLS is a great example
of a complementary contribution to the virtual experience
mechanism of the hippocampus and neocortex in the human
brain system. It helps consolidate knowledge by replaying
memories in the hippocampus before being transferred into
long-term memory using the neocortical system [83]. Referring
to the ER-GNN problem definition, suppose there are a
collection of tasks T ¼fT1; T2; ...; Ti; ...TMg which come
sequentially and each Ti 2 TðjTj¼ MÞ. Every task Ti has
training node-set Dtr
i
i
Dte
and testing node-set Dte
i . When the
model starts to learn task Ti, it uses resources from the training
set Dtr
and maximizes the performance using the testing set
i . It also selects the experience of examples B from the
stored experience buffer. Both the training data Dtr
i and experience
nodes B help to learn a model fu parameterized by u.
ER-GNN uses a cross-entropy loss function for node classification
and implements parameter updates to get optimal
parameters by minimizing the loss using optimization methods
such as Adam optimizer [84]. In the next step, after updating
the parameters, the model performs node selection in the current
training data set Dtr
i to be added as an example into the
experience buffer. MF, CM, and the novel method, IM are
used as a scheme based on the crucial part of the experience
selection strategy that stores experience in the buffer to
increase the performance ofcontinuous graph learning.
To measure the performance of ER-GNN, it uses a case
study to learn new tasks based on additional new nodes from
new classes to perform node classification. Several datasets, such
as Cora [75], Citeseer [75], and Reddit [35], are used in this evaluation.
Some graph learning models, such as Deepwalk [85],
Node2Vec [86], GCN [27], graphSage [35], etc., are used in
comparison with a finetuning strategy which means the model is
trained continually without employing lifelong learning techniques.
The result shows ER-GNN achieves the best mean performance
on the Cora dataset with mean accuracy of95.66%. It also
has the lowest mean forgetting, 17.08%, compared to graph
learning with finetuning strategy. From that experiment, it
shows the performance improvement in reducing forgetting
mean compared to state-of-the-art GNNs that are forced to do
learning a sequence ofnode classification tasks.
B. Lifelong Open-World Node Classification
Galke et al. [36] introduced lifelong open-world classification
to implement incremental training on graph data. This model
could rely on previous knowledge to help the learning process
in a sequence of tasks. The learned knowledge is stored as historic
data explicitly or within model parameters implicitly.
The model then analyses the influence of stored knowledge
that is helpful in refining the network. The model's objective
is to address the challenge in graph learning that nodes in graph
data can not be processed independently because the relation
ofnodes needs to be processed through an embedding mechanism
to share the information across the nodes. The second
objective is to enable the concept of open-world classification
[14] to implement incremental training in graph data and
FEBRUARY 2023 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 41
IEEE Computational Intelligence Magazine - February 2023

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - February 2023
