IEEE Computational Intelligence Magazine - February 2023 - 34

FIGURE 2 Graph Evolution. Graph lifelong learning aims to accommodate a new task Tt based on the graph Gt ¼ Gt1 þ DGt while maintaining the
performance in solving previous tasks T1; T2; T3; ... ; Tt1 .
A. Motivation of Graph Lifelong Learning
Graph learning is an advanced method that aims to complete
tasks and learn representations on graph-based data. Graph
learning can address several problems, such as node prediction
[28], edge prediction [29], and graph prediction [30].
Deep learning in graphs, based on graph neural networks
(GNNs), [31] refines and extends existing neural network algorithms
such as convolutional neural networks (CNNs) and recurrent
neural networks (RNNs). Various GNNsmodels have been
proposed, such as graph recurrent neural networks (Graph
RNNs) [32], GCNs [27], graph autoencoders (GAEs) [33], and
graph spatial-temporal networks [34].
Current popular graph learning methods assume that the
graph representation is perfect or complete before the learning
process begins. However, in many real-world applications, the
graph structure can flexibly and dynamically transform when
there is a shift in information from neighborhoods over time, as
illustrated in Fig. 2. When human social networks grow and
evolve, their graph representations accordingly become more
complex. That process is based on several factors, such as adding
or deleting entities in the graph, changing relationship types,
updating entity features, and adding downstream tasks. As with
many popular machine learning algorithms, graph learning algorithms
are limited by the isolated learning paradigm concept [14].
With that paradigm, the models cannot perform continuous
learning to accommodate new tasks and retain past information.
GNNs are able to perform inductive learning [35],
which can accommodate changes to the graph structure. In
particular, an inductively trained GNNs can forward-pass
newly added nodes/edges, enabling the model to be applied
to a different graph without retraining. However, being
able to apply a previously trained model to unseen data is
34 IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | FEBRUARY 2023
not sufficient for incremental learning settings, as it cannot
accommodate cases where new nodes/edges introduce new
classes to the classification task. In such a case, the model
does not have a knowledge base to classify inputs to the
new class [36], which suggests that another strategy is
needed to efficiently retrain the model and accommodate
new data and classes for incremental learning.
Addressing the dynamic graph [37], [38], [39] is also not
enough for incremental learning. The focus of dynamic graph
representation learning is to learn the embedding of the
dynamic graphs in every timestamp [40], [41], [42]. However,
since it was not developed with the intention to incrementally
solve new tasks, it does not address refining the performance
ofprevious experiences when learning new tasks. In particular,
there is no mechanism to minimize catastrophic forgetting and
preserve performance on the old tasks. The difference between
relevant graph learning properties is discussed in more detail in
Section II-C.
The above limitations of graph learning and dynamic
graph learning have motivated the development ofapproaches
to extend lifelong learning to graphs in recent years. 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. To achieve this, graph
lifelong learning addresses strategies to efficiently retrain the
model, decide the right time to retrain the model, and count
how much past data should be preserved. In response to the
increasing research and development attention on graph lifelong
learning, this survey paper comprehensively discusses the
current research progress, potential applications, and challenges
ofgraph lifelong learning.

IEEE Computational Intelligence Magazine - February 2023

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