IEEE Computational Intelligence Magazine - February 2023 - 35

A challenge emerges in how to address
new tasks that arise sequentially in
graph-based problems, such as the
addition of new nodes as the graph
grows over time and the inclusion of
new class labels in classification tasks.
B. Applications
Graph lifelong learning helps improve learning performance for
problems involving dynamic graph data with additional new tasks
or new instances. This section discusses various research problems
where lifelong learning approaches are needed to mine rich values
and solve current and future tasks in graph-related domains.
1) Social Networks
Graph learning has contributed to solving various prediction
problems on social networks [6]. The work by Han et al. [43]
combines GNNs with a lifelong learning method to detect
fake news on social media; however, it does not address the
dynamic property ofthe graph data. Social networks are timevariant
by nature, evolving when entities establish or cease
relationships. Graph lifelong learning can therefore help adapt
the learning process and update knowledge in social network
problems like that studied in [43].
2) Traffic Prediction
Traffic prediction is one a critical task in the transportation system
that directly affects human daily activities. The challenge
is that the underlying network will expand and evolve over
time, and the traffic patterns will dynamically adjust to the
evolving graph and traffic events that arise. Graph lifelong
learning provides excellent potential for predicting streaming
traffic flows on the dynamic graph network [44].
3) Recommender Systems
Recommendersystems areaclassofMLalgorithmsthatare
designed to recommend items or services to the user based on
what the user is most likely to be interested in. Research has shown
that recommender systems can be developed using graph learning
[45]. In real-world applications, the user interactions and the
collection of items or services to be recommended grow and
change over time, making it a good candidate for graph lifelong
learning.
4) Anomaly Detection
Anomaly detection is vital in several fields, including finance,
security, and environmental monitoring. Graph learning can
be used to mine the relationship between entities to capture
abnormalities from the system [46]. Most recent approaches
ignore temporal information that changes the definitions ofan
anomaly (e.g., data drift) and the representation of anomaly
features in the network over time. This challenge makes graph
lifelong learning an interesting candidate to tackle problems,
accommodate new knowledge, and prevent the forgetting of
previously learned anomaly characteristics.
C. Related Surveys
Several survey papers discuss the concepts of either lifelong
learning or graph learning. The book by Chen and Liu [14]
summarized the definition oflifelong learning as a continuous
learning process and summarized existing techniques for lifelong
learning on different learning approaches such as supervised,
unsupervised, semi-unsupervised, and reinforcement
learning. Delange et al. [47] presented a methodology for
dynamically determining the stability-plasticity trade-off in
lifelong learning by setting hyperparameters. This survey also
discussed the categorization of continual learning methods
using a taxonomy illustration. Biesialska et al. [48] summarized
the role ofcontinuous learning in natural language processing.
Parisi et al. [16] summarized the factors that motivate research
in lifelong learning and its relation to the biological aspect. It
also explained the lifelong learning setting challenge in neural
network models. The above lifelong learning surveys only
concentrate on general machine learning or deep learning
approaches to address tasks in grid or regular domains but do
not cover the concept oflifelong learning in graph data.
For graph learning, Xia et al. [3] comprehensively
reviewed the concept of graph learning with categorizations
such as graph signal processing, matrix factorization, random
walk, and deep learning, and also explained the application of
graph learning in several fields. Wu et al. [1] provided a comprehensive
overview of state-of-the-art GNNs methods
and discussed potential directions in different fields. Zhang
et al. [2] reviewed deep learning methods on graphs, such as
GCNs, GAEs, graph adversarial methods, graph reinforcement
learning (Graph RL), and Graph RNNs.
In addition to those, there are surveys that review other
relevant learning in dynamic graphs that mainly focus on graph
representation learning and graph embedding in dynamic
graphs with different characteristics from lifelong learning.
Kazemi et al. [37] reviewed representation learning in the
dynamic graph, which includes knowledge graphs by categorizing
encoders and decoders techniques. It also reviewed
potential applications and datasets that are widely used. Barros
et al. [38] provided an overview ofdynamic graph embedding,
including the underlying methods and the current works so
far. Those surveys discuss the embedding process in every
timestamp ofdynamic graphs that do not consider refining the
performance of prior knowledge and performing transfer
knowledge to help learn a new task as a lifelong learning characteristic.
In Section II-C, the different traits of other relevant
graph learning properties is explained in more detail.
D. Contributions
Existing surveys in the literature focus only on lifelong learning
or on graph learning, and none ofthem addresses the challenges
FEBRUARY 2023 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 35
IEEE Computational Intelligence Magazine - February 2023

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