PaperlayerView exact revision on arXiv

Source-linked AI summary

Graph Lifelong Learning: A Survey

Falih Gozi Febrinanto, Feng Xia, Kristen Moore, Chandra Thapa, Charu Aggarwal

arXiv:2202.10688v2cs.LGcs.AI

TL;DR

Graph-learning methods often assume a complete, static graph, but real graph structure and tasks can evolve, and general lifelong-learning methods do not directly fit graphs. This survey reviews graph lifelong learning, organizes its methods, and discusses applications and challenges. It presents a four-category method taxonomy and comparisons across scenarios and applications.

  • Problem

    Graph-learning methods commonly assume a complete graph, while evolving graph structure and emerging tasks require continual learning that retains earlier knowledge.

    Source: §I · p. 1, A. Motivation of Graph Lifelong Learning · p. 2

  • Method

    The survey synthesizes graph lifelong-learning methods and discusses their applications and open challenges.

    Source: §X · p. 17

  • Results

    The survey categorizes methods as architectural, rehearsal, regularization, and hybrid approaches and compares them across scenarios and applications.

    Source: §X · p. 17

  • Takeaways & Limitations

    Graph lifelong learning targets acquiring knowledge from evolving graphs while refining prior knowledge to address new tasks.

    Source: §I · pp. 1–2

Abstract

from arXiv · show

Graph learning is a popular approach for performing machine learning on graph-structured data. It has revolutionized the machine learning ability to model graph data to address downstream tasks. Its application is wide due to the availability of graph data ranging from all types of networks to information systems. Most graph learning methods assume that the graph is static and its complete structure is known during training. This limits their applicability since they cannot be applied to problems where the underlying graph grows over time and/or new tasks emerge incrementally. Such applications require a lifelong learning approach that can learn the graph continuously and accommodate new information whilst retaining previously learned knowledge. Lifelong learning methods that enable continuous learning in regular domains like images and text cannot be directly applied to continuously evolving graph data, due to its irregular structure. As a result, graph lifelong learning is gaining attention from the research community. This survey paper provides a comprehensive overview of recent advancements in graph lifelong learning, including the categorization of existing methods, and the discussions of potential applications and open research problems.

I. INTRODUCTION

Graph learning often assumes a complete, static graph, while real graphs and their tasks can change over time. This survey introduces graph lifelong learning as a way to accommodate sequentially arriving knowledge while retaining performance on earlier tasks.

I. INTRODUCTION

  • Sequential learning can cause catastrophic forgetting when new information interferes with prior knowledge, reducing performance on earlier tasks.This motivates lifelong learning’s goal of transferring, acquiring, and refining knowledge as tasks arrive.Source: §I · p. 1
  • General lifelong-learning methods for ordered inputs do not directly transfer to graphs, whose lack of canonical node order complicates representation and distance measurement.Graph representation learning uses low-dimensional vectors, while conventional GCN approaches may require the entire graph, creating efficiency problems for incremental learning.Source: §I · p. 2
  • Graph lifelong learning must detect new graph instances or tasks and accommodate knowledge from added data instances or classes.The survey also lists efficient representation learning without requiring the entire graph structure as a need.Source: §I · p. 2

A. Motivation of Graph Lifelong Learning

  • New nodes, relations, features, and downstream tasks can change a graph, whereas common graph-learning methods assume its representation is complete before training.Inductive GNNs can process unseen nodes or edges, but do not accommodate new classes without a strategy to retrain and add knowledge.Source: A. Motivation of Graph Lifelong Learning · p. 2
  • Graph lifelong learning addresses evolving graphs and new tasks by acquiring new information while refining previous knowledge and retaining performance on earlier tasks.The survey identifies efficient retraining, when to retrain, and how much past data to preserve as strategies to address.Source: Figure 2 · p. 2, A. Motivation of Graph Lifelong Learning · p. 3

B. Applications

The survey discusses graph lifelong learning applications, distinguishes this topic from prior surveys of lifelong learning, graph learning, and dynamic graphs, and reviews methods, comparisons, and open challenges.

B. Applications

  • Graph lifelong learning is presented as relevant to social networks, traffic prediction, recommender systems, and anomaly detection as their data or underlying networks change.Examples include evolving social ties, expanding road networks and shifting traffic patterns, changing user-item interactions, and time-varying anomaly definitions.Source: B. Applications · p. 3

C. Related Surveys

  • Prior surveys cover lifelong learning, graph learning, or dynamic graph embedding, but the survey identifies their combined graph-lifelong-learning challenges as unaddressed.Dynamic graph embedding surveys focus on timestamp-specific representations rather than refining prior knowledge or transferring it to new tasks.Source: D. Contributions · pp. 3–4, C. Related Surveys · p. 3

D. Contributions

  • The survey reviews and categorizes graph lifelong-learning algorithms, discusses potential applications, and identifies open issues and challenges.It presents these topics as contributions to the emerging field of graph lifelong learning.Source: D. Contributions · p. 4
  • The survey compares existing graph lifelong-learning models across applications and scenarios.It also discusses comparisons of methods in terms of computational aspects within each category.Source: D. Contributions · p. 4

E. Paper Organization

The survey reviews lifelong learning concepts and graph-learning properties, then organizes graph lifelong learning methods and distinguishes them from related learning paradigms. It also outlines the paper’s sections, from domain overview through methods, benchmarks, open challenges, and conclusion.

E. Paper Organization

  • The paper proceeds from a domain overview and taxonomy to architectural, rehearsal, regularization, and hybrid methods, then compares methods, reviews benchmarks, and discusses open challenges.The overview covers lifelong learning, related learning paradigms, graph-learning properties, scenarios, notation, and the taxonomy.Source: E. Paper Organization · p. 4, §II · p. 4

II. OVERVIEW

  • The overview introduces the graph lifelong learning scenarios, domain notation, and taxonomy that frame the methods reviewed later.The taxonomy is further explained in Subsection II-F.Source: §II · p. 4

A. Lifelong Learning

  • Lifelong learning seeks to accommodate new knowledge while refining prior knowledge, and its established categories are architectural, regularization, and rehearsal.The survey focuses on adapting these approaches to graph representation learning and graph data architecture.Source: A. Lifelong Learning · p. 4

B. Relevant ML concepts

  • Transfer learning, multitask learning, and online learning differ from lifelong learning in how they transfer, share, or sequentially learn knowledge and whether they retain prior experience.Lifelong learning accommodates new distributions while trying to avoid forgetting; OOD generalization instead targets robust prediction across changed test distributions.Source: B. Relevant ML concepts · pp. 4–5

C. Relevant Graph Learning Characteristics

Graph lifelong learning combines graph representation learning with incremental learning, retention, and transfer. It differs from temporal graph learning that updates embeddings without addressing new downstream tasks or catastrophic forgetting, and the survey distinguishes three update scenarios.

C. Relevant Graph Learning Characteristics

  • Dynamic graph learning transfers knowledge across changing snapshots or streams, but does not address new downstream tasks or backward transfer to refine prior embeddings.The survey contrasts this with lifelong learning’s focus on catastrophic forgetting.Source: C. Relevant Graph Learning… · pp. 5–6
  • Graph lifelong learning combines graph representation learning with incremental learning, knowledge retention, and transfer of prior knowledge to new tasks.Graph representation learning converts irregular graph data into feature vectors; the model should learn future instances or classes without training from scratch and avoid catastrophic forgetting.Source: C. Relevant Graph Learning… · p. 6

D. Graph Lifelong Learning Scenarios

  • The survey distinguishes learning new instances, new classes, and both together; new instances can add nodes, attributes, and relations without adding classes.New classes require retaining prior experience, while new instances and classes can provide data from existing or unseen classes.Source: D. Graph Lifelong Learning Scenarios · p. 6

E. Notation of Domain

The survey formalizes graph data as evolving graphs and sequential tasks, then categorizes graph lifelong learning methods by how they accommodate graph structure. Its architectural coverage includes Feature Graph Networks, which reframe node classification and report results in data- and class-incremental settings.

E. Notation of Domain

  • Graph lifelong learning represents evolving graph data as G = (G1, G2, ..., Gt), with Gt = Gt−1 + ∆Gt, and learns sequential tasks while maintaining prior-task performance and avoiding catastrophic forgetting.The graph at each time point is represented by an adjacency matrix and node-feature matrix.Source: E. Notation of Domain · p. 6

F. Taxonomy

  • The survey categorizes graph lifelong learning methods as architectural, regularization, rehearsal, and hybrid, adding hybrid for methods that combine approaches.The categorization adapts general lifelong-learning categories to graph structure considerations.Source: F. Taxonomy · p. 6, Figure 3 · p. 7

III. ARCHITECTURAL APPROACH

  • Architectural approaches adapt graph lifelong learning by changing graph structure, expanding units, or compressing models; examples include FGN and HPN.FGN converts graph data architecture into a representation usable by regular learning architectures, while HPNs use hierarchical abstractions of prototypes.Source: §III · pp. 6–7

A. Feature Graph Networks

  • FGN transforms node classification into feature-graph classification, enabling independent node processing in CNN-like mini-batches without preprocessing the entire graph.It represents original nodes as independent graphs and original features as nodes; cross-correlation builds feature interactions, with broadcast, transform, and attention layers available to construct feature adjacency.Source: A. Feature Graph Networks · pp. 7–8
  • On Pubmed, FGN reports 0.872 ± 0.009 overall performance in data-incremental learning and 0.857 ± 0.003 performance in class-incremental learning, with a 3.49 forgetting rate.The survey reports the 3.49 forgetting rate as the lowest among compared models.Source: A. Feature Graph Networks · p. 8

B. Hierarchical Prototype Networks

Hierarchical Prototype Networks encode graph knowledge at atomic, node, and class levels, selecting and updating prototypes as new embeddings arrive. The survey also describes experience replay, which stores selected examples for training on subsequent tasks.

B. Hierarchical Prototype Networks

  • HPNs encode node attributes and multi-hop structure in atomic embeddings, then form atomic-, node-, and class-level prototypes for node classification.The hierarchy uses convolutional-network feature maps at different layers.Source: B. Hierarchical Prototype Networks · p. 8
  • HPNs compare incoming embeddings with existing prototypes using cosine similarity, refining close matches and creating prototypes for distinct new tasks.A threshold filters redundant incoming embeddings before updating atomic prototypes.Source: B. Hierarchical Prototype Networks · p. 8
  • HPNs report the best average accuracy of 93.7 ± 1.5 on Cora and the lowest forgetting rate, -0.9 ± 0.9, on Actor among compared methods.The survey also reports good memory efficiency for HPNs.Source: B. Hierarchical Prototype Networks · p. 8

A. Experience Replay GNN Framework

  • ER-GNN replays examples from a stored experience buffer while learning each new task, using selection strategies including mean of feature, coverage maximization, and influence maximization.The selected examples are added to the buffer after parameter updates.Source: A. Experience Replay GNN Framework · p. 9
  • ER-GNN achieves 95.66% mean accuracy on Cora and 17.08% mean forgetting, reported as the lowest forgetting compared with graph-learning finetuning.The evaluation learns node-classification tasks as new nodes from new classes are added.Source: A. Experience Replay GNN Framework · p. 9

B. Lifelong Open-world Node Classification

Lifelong open-world node classification incrementally accommodates new graph tasks and classes while using knowledge from previous tasks. Its experiments find that limited-history training performs almost like full-history training.

B. Lifelong Open-world Node Classification

  • The model stores prior knowledge explicitly as historic data or implicitly in parameters, then uses it to refine the network across tasks.It also introduces a measurement for regulating graph evolution and can integrate isotropic and anisotropic GNNs.Source: B. Lifelong Open-world Node… · p. 9
  • When new classes arrive, the model adds corresponding parameters to the output layer for the next task.The incremental setup assumes the graph structure, including nodes and edges, changes over time.Source: B. Lifelong Open-world Node… · p. 9
  • Rehearsal can use either a cold restart from scratch or a warm restart from the previous task’s final parameters, with history size limiting past data.Source: B. Lifelong Open-world Node… · p. 9
  • At least 95% accuracy is achieved with half-dataset coverage, and limited-history incremental learning is almost similar to full-history training.The experiments use GAT, GCN, and GraphSage on PharmaBio and DBLP scenarios with newly added classes.Source: B. Lifelong Open-world Node… · pp. 9–10

V. REGULARIZATION APPROACH

The regularization approach consolidates knowledge in a single fixed-capacity model while learning new tasks. The surveyed methods apply this idea to graph embeddings, few-shot class-incremental learning, and topology-aware parameter preservation.

V. REGULARIZATION APPROACH

  • Regularization methods use a fixed-capacity model and a loss term to consolidate new-task learning while retaining prior knowledge.The survey identifies DiCGRL [68], GPIL, and TWP among graph methods using this approach.Source: §V · p. 10

A. Disentangle-based Continual Graph Representation learning

  • DiCGRL [68] disentangles graph relation triplets into semantic components, selects related components through attention, and updates activated neighbors to retain prior-task performance.It uses regularization constraints when training incoming triplets and supports knowledge-graph and network embeddings.Source: A. Disentangle-based Continual Graph… · p. 10
  • DiCGRL [68] reaches 47.7% maximum average H@10 on FB15K-237 and 85% average accuracy on Pubmed, with results close to upper-bound retraining.Source: A. Disentangle-based Continual Graph… · p. 10

B. Graph Few-shot Class-incremental Learning

C. Topology-aware Weight Preserving

  • TWP estimates parameter importance from taskwise and topology-preserving gradients, then penalizes changes to important parameters while learning new tasks.Lower-importance parameters remain freer to adjust, while the method also promotes minimizing importance scores to maintain plasticity.Source: C. Topology-aware Weight Preserving · p. 11
  • TWP achieves 0.976 average performance and 0.001 ± 0.062 average forgetting on Reddit, the highest performance and lowest forgetting scores reported for that dataset.Source: C. Topology-aware Weight Preserving · p. 11

D. Translation-based Knowledge Graph Embedding

Translation-based knowledge graph embedding uses regularization to preserve earlier representations as new triples arrive. The following hybrid section describes ContinualGNN’s combination of rehearsal and regularization for streaming graph changes.

D. Translation-based Knowledge Graph Embedding

  • Song and Park penalize rapid changes between old and new embedding parameters to reduce risk on new triples while preserving earlier representations.The approach assumes entity relations remain fixed after training and uses regularization to accommodate new embeddings.Source: D. Translation-based Knowledge Graph… · pp. 11–12
  • The translation-based embedding method achieves 94.27 H@10 on WN18 triple classification, outperforming EWC and finetuning and nearly matching joint learning.Source: D. Translation-based Knowledge Graph… · p. 12

A. ContinualGNN

  • ContinualGNN combines replay with regularization: propagation identifies affected nodes around new patterns, and rehearsal preserves existing knowledge as patterns are learned.The hybrid approach combines multiple lifelong-learning strategies to use their respective strengths.Source: A. ContinualGNN · p. 12, §VI · p. 12
  • ContinualGNN [10] reaches 0.9212 maximum average accuracy on Elliptic, outperforming incremental finetuning and nearly matching joint-training upper bounds.Source: A. ContinualGNN · p. 12

B. Lifelong Dynamic Attributed Network Embedding

LDANE learns node embeddings for growing attributed networks using structural and attribute objectives alongside architectural expansion and selective retraining. On SYN-2, it reports the highest multilabel-classification performance among the compared embedding models.

B. Lifelong Dynamic Attributed Network Embedding

  • LDANE combines global structure reconstruction, local neighbor proximity, and attribute-similarity constraints in its embedding objective.The global term reconstructs adjacency structure; the attribute term encourages similar nodes to embed more similarly than dissimilar nodes.Source: B. Lifelong Dynamic Attributed Network… · p. 13
  • LDANE [72] expands autoencoder input and output neurons as node counts grow, selectively retrains affected weights, and can split the autoencoder when features change significantly.Rehearsal and regularization support retraining and limit parameter drift.Source: B. Lifelong Dynamic Attributed Network… · p. 13
  • On SYN-2, LDANE achieves the highest reported multilabel-classification micro-F1 of 0.93 and macro-F1 of 0.92 among compared embedding models.The comparisons include models that do not employ lifelong learning techniques.Source: B. Lifelong Dynamic Attributed Network… · p. 13

C. TrafficStream

TrafficStream targets evolving traffic graphs with efficient pattern updates and knowledge retention. The survey also compares lifelong-learning methods and outlines benchmarking considerations.

C. TrafficStream

  • TrafficStream limits GNN processing to two-hop-neighbor subgraphs and detects graph evolution to select nodes with significant changes.The approach targets additions such as sensors or stations and shifts in points of interest.Source: C. TrafficStream · p. 14
  • It replays sampled prior data and smooths parameters using EWC, weighting changes by each parameter’s importance to preserve prior knowledge.Less-important parameters can adapt more, while changes to more-important parameters are limited.Source: C. TrafficStream · p. 14
  • TrafficStream achieves high prediction accuracy versus traditional retraining strategies and reduces training complexity through continual learning.Experiments use the real-world PEMS3-Stream dataset and compare against retraining baselines including STGCN, STSGCN, and GRU.Source: C. TrafficStream · p. 14

VII. METHOD COMPARISONS

  • Methods address new instances, new classes, or both across differing applications; architectural expansion, regularization, rehearsal, and hybrid approaches have distinct efficiency and performance trade-offs.Architectural methods must manage memory and complexity; regularization can constrain new-task performance, rehearsal requires careful sample selection and computational efficiency, and hybrid methods add framework complexity.Source: §VII · p. 14

VIII. BENCHMARKS

  • The benchmarks section introduces datasets, experimental details, and performance calculation, while Table III presents method comparisons.The supplied passages do not specify Table III’s individual comparisons.Source: §VIII · p. 14, Table III · p. 15

A. Datasets

The survey describes ways to adapt existing graph datasets for lifelong-learning scenarios and reviews metrics for evaluating performance across sequential tasks.

A. Datasets

  • No dataset is specifically designed for graph lifelong learning; image benchmarks converted through superpixels remain unlike irregular real-world graph structures.Existing graph benchmarks can be modified by distributing classes across sequential tasks with new instances and classes.Source: A. Datasets · p. 15

B. Performance Metrics

  • Average accuracy summarizes task accuracies after incremental learning, while the forgetting measure averages forgetting across tasks.Both metrics are described as measures over tasks after learning incrementally through task T.Source: B. Performance Metrics · pp. 15–16
  • Learning curve area measures average b-shot performance across the learning curve and is high when zero-shot performance is good and learning is quick.Here, b denotes a mini-batch number.Source: B. Performance Metrics · p. 16

IX. OPEN ISSUES

The survey identifies catastrophic forgetting and several graph-specific open issues, including uncertain neighborhoods, extreme evolution, global dependency learning, benchmarking, and distributed training.

IX. OPEN ISSUES

  • Catastrophic forgetting—losing performance on previous tasks while learning new ones—remains a central challenge; finetuning can degrade prior-task performance, while joint training requires access to all data.Joint training may be inefficient with massive data, motivating iterative learning with smaller data amounts.Source: §IX · p. 16

A. Uncertain Neighboorhoods

  • Dynamic graph embedding still needs strategies for uncertain neighborhoods, including when to embed new nodes and relations efficiently as graphs change.The survey notes that some methods address continual embedding, but an optimal strategy remains needed.Source: A. Uncertain Neighboorhoods · p. 16

B. Extreme Evolution

  • Extreme graph evolution raises a retention choice: preserve prior tasks or allow forgetting when those tasks are no longer relevant to the current graph.The survey distinguishes graphs that retain much of their old structure from those whose former structures become extinct.Source: B. Extreme Evolution · p. 16

C. Global Dependencies Learning

  • Incrementally propagating new-instance information across all nodes and edges to learn global graph dependencies remains challenging, especially at large scale.The survey calls for more efficient continual global-dependency learning.Source: C. Global Dependencies Learning · p. 16

D. Comparative Benchmarking

  • Benchmarking remains non-standard across new-instance, new-class, and combined scenarios and across different graph-learning applications.The survey identifies fair comparison as a research need.Source: D. Comparative Benchmarking · p. 16

E. Model Training Parallelisation

  • Graph lifelong learning needs federated approaches that compute on local edge devices and aggregate models, since centralizing data can raise security, performance, and cost concerns.Nodes, relations, and new classes may appear at local points in distributed architectures.Source: E. Model Training Parallelisation · pp. 16–17

X. CONCLUSION

  • The survey synthesizes methods, comparisons, benchmark settings, applications, and open problems to support further work on graph lifelong-learning theory, models, algorithms, and applications.Its future directions include class imbalance and common machine-learning concerns such as fairness, interpretability, privacy, and data quality.Source: §X · p. 17

Related papers

Loading 2202.10688v2…