IEEE Computational Intelligence Magazine - February 2023 - 36

ofuniting the two. This paper provides a survey ofrecent developments
in the emerging field ofgraph lifelong learning, which
enable continuous learning mechanisms in graph learning models.
Moreover, this survey paper reviews new perspectives of
incremental learning in graph data different from other relevant
learning concepts, such as spatio-temporal neural networks and
dynamic graph learning. We summarize the benefit of graph
lifelong learning that can overcome the catastrophic forgetting
problem, refine previous experiences, and accommodate new
knowledge to the graph learning model. Based on our knowledge,
we are the first to review the emerging topic ofgraph lifelong
learning. Overall, the main contributions of this survey
paper are:
❏ A discussion of the potential applications that benefit from
implementing graph lifelong learning.
❏ A comprehensive review and categorization of the current
progress in graph lifelong learning algorithms.
❏ A comparison of existing models ofgraph lifelong learning
based on its different applications and scenarios.
❏ A discussion ofopen issues and challenges ofgraph lifelong
learning as insights for future research directions.
E. Paper Organization
The remainder of this survey is organized into the following
sections. Section II provides an overview of the domain of
graph lifelong learning, including lifelong learning in general,
relevant graph learning properties, graph lifelong learning scenarios,
notation of domain, and taxonomy of the current
works. The categorization of graph lifelong learning techniques
is further explained in Subsection II-F. That categorization
includes architectural approaches in Section III, rehearsal
approaches in Section IV, regularization approaches in
Section V, and hybrid approaches in Section VI. Section VII
compares each categorization to consider its implementation
and summarizes each method's characteristics based on scenarios
and their applications. Section VIII reviews benchmark settings
for graph lifelong learning, including datasets and
evaluation metrics. Section IX describes the open issues and
challenges for future research direction. Lastly, the conclusion
ofthis survey paper is given in Section X.
II. OVERVIEW
This section begins with a discussion of lifelong learning for
regular deep learning or machine learning methods in Section
II-A. Then Section II-B discusses some related learning paradigms
and learning objectives in typical machine learning
models and highlights how they differ from lifelong learning.
Section II-C reviews relevant graph learning properties that
have similarities with graph lifelong learning problems. Section
II-D then explains several graph lifelong learning scenarios.
Section II-E then explains the notation and problem definition
ofgraph lifelong learning. Finally, Section II-F presents a categorization
of graph lifelong learning approaches that we will
review in the later sections.
36 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | FEBRUARY 2023
A. Lifelong Learning
Lifelong learning aims to enable continuous learning to
accommodate newly appeared knowledge and refine previous
knowledge. Prior studies on lifelong learning models can be
divided into three categories: architectural, regularization, and
rehearsal [24], [48].
The architectural approach modifies network architecture
by adding more units and expanding and compressing the
architecture [15], [22], [49]. Rusu et al. [21] introduced a progressive
network (ProgNN) that was able to adjust new capacity
alongside pre-trained models so that it gave the flexibility
to accommodate the latest knowledge and reuse the old ones.
Similarly, Yoon et al. [25] proposed a dynamic expandable
network (DEN) that is able to decide the network capacity
with only a necessary number of units when new knowledge
arrives.
The regularization approach uses some additional loss
terms to maintain the stability of prior parameters when learning
new parameters. The most common technique, elastic
weight consolidation (EWC), was proposed by Kirkpatrick
et al. [19], which regularizes model parameters by penalizing
the changes based on task importance, so new tasks will have
optimum parameter region where the model performs without
any catastrophic forgetting ofprevious experiences. Li and
Hoiem [20] developed a model called learning without forgetting
(LWF) that employs distillation loss to maintain the
network's output as close to its previous values.
The rehearsal approach relies on a retraining process from
prior task samples that can keep the performance of previous
experiences [50]. Rebuffi et al. [51] proposed incremental classifier
and representation learning (ICARL) that can store samples
ofeach task that is constantly replayed while learning new
tasks.
In order to apply lifelong learning models like those mentioned
above to problems in graph-based domains, they must
first be adapted to incorporate graph learning techniques that
address graph representation learning and graph data architecture.
This survey paper focuses on reviewing methods that
have the specific aim of deploying lifelong learning settings in
graph-based domains.
B. Relevant ML Concepts
This section discusses some learning paradigms in typical
machine learning models, such as transfer learning, multitask
learning, and online learning [14] and highlights how they differ
from lifelong learning. Transfer learning seeks to take the
knowledge gained from learning a task in a particular source
domain and transfer it over to a different, but related, target
domain [58]. The transfer learning paradigm has no specific
mechanism designed to retain past knowledge and use it to
learn new knowledge, which is the main goal oflifelong learning.
Multitask learning aims to share knowledge across various
tasks to help the learning process [59]. Instead of optimizing a
single task, multitask learning aims to maximize the performance
of the tasks simultaneously. Data is assumed to be static
IEEE Computational Intelligence Magazine - February 2023

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