PaperlayerView exact revision on arXiv

Source-linked AI summary

Graph Self-Supervised Learning: A Survey

Yixin Liu, Ming Jin, Shirui Pan, Chuan Zhou, Yu Zheng, Feng Xia, Philip S. Yu

arXiv:2103.00111v5cs.LG

TL;DR

Graph SSL’s diverse pretext-task designs motivate a unified framework and systematic taxonomy. This survey organizes methods into four categories and reports that representative contrast-based methods outperform generalization-based and auxiliary-property methods, while some hybrid methods surpass supervised baselines.

  • Problem

    The growing diversity of graph-SSL pretext tasks creates a need for a unified framework and systematic taxonomy to summarize methods and applications.

    Source: §1 · p. 2

  • Method

    The survey mathematically formalizes graph SSL, classifies methods as generation-based, auxiliary property-based, contrast-based, or hybrid, and compiles applications and empirical resources.

    Source: §11 · p. 18

  • Results

    Representative contrast-based methods outperform generalization-based and auxiliary-property methods, while hybrid methods are competitive and some surpass supervised baselines.

    Source: §8 · p. 16

  • Takeaways & Limitations

    The survey collects datasets, evaluation methods, performance comparisons, and open-source implementations as resources for empirical graph-SSL research.

    Source: §8 · p. 16, §11 · p. 18

    Graph SSL still lacks a theoretical foundation, with most methods designed intuitively and evaluated through empirical experiments.

    Source: §10 · p. 17

Abstract

from arXiv · show

Deep learning on graphs has attracted significant interests recently. However, most of the works have focused on (semi-) supervised learning, resulting in shortcomings including heavy label reliance, poor generalization, and weak robustness. To address these issues, self-supervised learning (SSL), which extracts informative knowledge through well-designed pretext tasks without relying on manual labels, has become a promising and trending learning paradigm for graph data. Different from SSL on other domains like computer vision and natural language processing, SSL on graphs has an exclusive background, design ideas, and taxonomies. Under the umbrella of graph self-supervised learning, we present a timely and comprehensive review of the existing approaches which employ SSL techniques for graph data. We construct a unified framework that mathematically formalizes the paradigm of graph SSL. According to the objectives of pretext tasks, we divide these approaches into four categories: generation-based, auxiliary property-based, contrast-based, and hybrid approaches. We further describe the applications of graph SSL across various research fields and summarize the commonly used datasets, evaluation benchmark, performance comparison and open-source codes of graph SSL. Finally, we discuss the remaining challenges and potential future directions in this research field.

1 INTRODUCTION

Graph SSL addresses the label dependence of supervised graph learning through pretext tasks that learn from data itself, but graph topology and node dependencies make its task design distinct from SSL for images and text. This survey formalizes graph SSL, organizes its methods, and reviews its resources, applications, limitations, and future directions.

Introduction

  • Manual labels make supervised graph learning costly and can undermine generalization and robustness; SSL instead learns representations from automatically acquired pretext-task signals.These signals come from the data itself, without manual annotation.Source: §1 · p. 1
  • Graph SSL task design must account for irregular non-Euclidean structure and dependencies between nodes, unlike the grid-based or independent examples common in CV and NLP.For graph generation, both topology and node features matter; for contrastive tasks, node dependencies cannot be ignored.Source: §1 · pp. 1–2, Figure 1 · p. 2
  • From early random-walk graph embeddings, and graph autoencoders to newer pretext-task designs, the field's diversity motivates a unified framework and taxonomy.The survey mathematically formalizes graph SSL and groups methods into generation-based, auxiliary property-based, contrast-based, and hybrid approaches.Source: §1 · p. 2
  • The survey reviews classical and recent methods with fine-grained classifications and comparisons, and compiles datasets, benchmarks, performance comparisons, implementations, applications, limitations, and six future directions.It also builds taxonomies of downstream tasks and SSL learning schemes.Source: §1 · p. 2
  • Compared with prior graph SSL surveys [36], [37], this survey distinguishes itself through a mathematical encoder-decoder framework, finer taxonomy, newer coverage, more resources and applications, and forward-looking discussion.It focuses specifically on graph-structured data rather than general SSL, CV, or contrastive learning surveys.Source: §1 · p. 2

2 DEFINITION AND NOTATION

The section defines graph SSL terminology and establishes graph and encoder notation for describing the methods reviewed in the survey.

2.1 Term Definitions

  • Manual labels are human-annotated, while pseudo labels are generated automatically from data and can reduce labeling costs when labels are scarce or data is vast.Pretext tasks, such as graph reconstruction, are learned from unlabeled data using pseudo labels; downstream tasks such as node and graph classification evaluate learned representations and generally use manual labels.Source: §2.1 · pp. 2–3
  • Supervised learning uses manual labels, unsupervised learning uses none, and self-supervised learning generates its supervision signals from data through pretext tasks.These pretext tasks aim to improve performance and generalization on downstream tasks.Source: §2.1 · p. 3

2.2 Notations

  • A plain graph is G = (V, E), with nodes V, edges E, neighborhoods N(vi), and adjacency matrix A; attributed graphs additionally associate features with nodes and/or edges.When only nodes have features, their matrix is X and the attributed graph is G = (V, E, X).Source: §2.2 · p. 3
  • Most reviewed methods use GNN encoders to transform node features X and connectivity A into node representations H, then apply readout functions to obtain graph representations hG.The survey also introduces GNN, readout-function, and commonly used loss-function formulations in appendices.Source: §2.2 · p. 3

3 FRAMEWORK AND CATEGORIZATION

The section formalizes graph SSL as an encoder–pretext-decoder process trained on unlabeled graph data, then categorizes it by pretext objective, training scheme, and downstream prediction scale.

3.1 Unified Framework and Mathematical Formulation of Graph Self-Supervised Learning

  • Graph SSL uses an encoder to embed nodes and a pretext decoder to optimize an SSL loss, then applies the trained encoder to supervised downstream tasks.The framework defines the SSL objective over unlabeled graph data and introduces a downstream decoder for labeled tasks.Source: §3.1 · p. 3

3.2 Taxonomy of Graph Self-supervised Learning

  • Graph SSL comprises generation-based reconstruction, auxiliary-property prediction, contrastive learning, and hybrid methods that combine multiple pretext tasks.Generation methods reconstruct features or graph structure; auxiliary-property methods use regression or pseudo-label classification; contrastive methods maximize agreement for positive augmented pairs and minimize it for negative samples. Hybrid methods combine objectives, as in edge reconstruction with node-level contrast.Source: §3.2 · pp. 4–5
  • The categories trade simplicity and flexibility against costs or design challenges, including reconstruction memory, domain knowledge, empirical tuning, and balancing hybrid objectives.Generation can be memory-consuming on large graphs; auxiliary-property selection may require domain knowledge; contrastive design often requires time-consuming experiments; hybrid methods must balance components.Source: §3.2 · p. 5

3.3 Taxonomy of Self-Supervised Training Schemes

  • The three training schemes are pre-training and fine-tuning (PF), joint learning (JL), and unsupervised representation learning (URL).PF pre-trains an encoder before supervised fine-tuning, with pretraining and fine-tuning datasets allowed to differ. JL combines self-supervised and downstream losses using a trade-off parameter; URL freezes the encoder for the downstream stage and uses the same dataset in both stages, without encoder-training supervision.Source: §3.3 · p. 5

3.4 Taxonomy of Downstream Tasks

  • Downstream tasks are grouped by prediction scale into node-level, link-level, and graph-level tasks.They predict node properties, edge or node-pair properties, or properties of individual graphs, respectively.Source: §3.4 · pp. 5–6

4 GENERATION-BASED METHODS

Generation-based graph SSL uses input graphs as supervision, reconstructing either graph features or topology. Its two branches use distinct targets and therefore capture different levels of information.

4.1 Feature Generation

  • Feature generation recovers graph features from masked, noisy, or clean inputs, targeting node, edge, or transformed feature matrices.Examples include masked-feature recovery in Graph Completion [17], noisy-feature recovery in MGAE, and reconstruction from clean inputs in GALA.Source: §4.1 · p. 6

4.2 Structure Generation

  • Structure generation usually reconstructs the full or partial adjacency matrix, using encoders and decoders to recover edges or masked links.GAE uses a sigmoid inner-product decoder and binary cross-entropy; reweighting positive edges or subsampling unconnected pairs can address adjacency sparsity.Source: §4.2 · pp. 6–7
  • Structure generation tends to capture node-pair information through edge reconstruction, whereas feature generation often captures node-level knowledge.The approaches consequently use distinct decoder and loss designs.Source: §4.2 · p. 7

5 AUXILIARY PROPERTY-BASED METHODS

Auxiliary property-based graph SSL creates supervision signals from graph properties without manual labels, using discrete classification targets or continuous regression targets.

5 AUXILIARY PROPERTY-BASED METHODS

  • Auxiliary property-based SSL generates pseudo-labels automatically from node-, link-, or graph-level properties, avoiding the manual annotation costs of supervised learning.It retains a sample-label training paradigm, but the labels are self-generated from graph data.Source: §5 · p. 7

5.1 Auxiliary Property Classification

  • Auxiliary property classification uses discrete pseudo-labels, a classifier decoder, and cross-entropy loss, with clustering-based and pair relation-based variants.The variants differ in how they define the finite pseudo-label set.Source: §5.1 · p. 7

5.1.1 Clustering-based Methods

  • Clustering-based methods label nodes by feature or structural clusters; M3S clusters embeddings, aligns clusters to class labels, and trains on high-confidence unlabeled nodes.Other approaches include feature-based Node Clustering, structural Graph Partitioning, Cluster Preserving [50] with soft pseudo-labels, and CAGNN [56], which refines feature clusters using inter-cluster edges.Source: §5.1.1 · p. 8

5.1.2 Pair Relation-based Methods

  • Pair relation-based methods assign pseudo-labels to node pairs, including shortest-path distance in S2GRL [57], bounded distance in PairwiseDistance, and relative centrality in Centrality Score Ranking [50].Centrality Score Ranking compares eigencentrality, betweenness, closeness, and subgraph centrality scores.Source: §5.1.2 · pp. 8–9

5.2 Auxiliary Property Regression

  • Auxiliary property regression predicts continuous graph properties with mean squared error, including node properties and pairwise feature similarity.NodeProperty predicts degree, local importance, or local clustering coefficient; Distance2Cluster predicts distances to cluster centers, while PairwiseAttrSim and SimP-GCN use cosine similarity of raw features. The survey notes that classification methods are more diverse and calls for more continuous properties in regression.Source: §5.2 · p. 9

6 CONTRAST-BASED METHODS

This section surveys contrast-based graph SSL, which uses graph augmentations and contrastive tasks across multiple granularities to maximize agreement between semantically similar instances and minimize it for unrelated ones. It covers graph augmentations, contrastive pretext tasks, mutual-information estimation, and triplet-margin, BYOL-inspired, and Barlow Twins objectives, including triplet loss’s limits as an MI estimator and differences in loss preferences across graph scales.

6.1 Graph Augmentations

  • Graph augmentations address the difficulty of transferring Euclidean-domain transformations to graphs and are classified as attributive, topological, or hybrid.Figure 7 illustrates node feature masking, node feature shuffle, edge modification, graph diffusion, and subgraph sampling.Source: §6.1 · pp. 9–10, Figure 7 · p. 10

6.1.1 Attributive augmentations

  • Attributive augmentations mask node features row-wise or column-wise, adapt masking by node centrality, or shuffle selected node features between positions.GCA leaves important features unmasked and assigns higher masking probability to less important nodes.Source: §6.1.1 · p. 10

6.1.2 Topological augmentations

  • Topological augmentations modify edges by dropping or inserting them, while graph diffusion connects indirectly linked nodes with calculated weights.Edge modifications can be adaptive or adversarial,; diffusion includes heat-kernel and Personalized PageRank forms.Source: §6.1.2 · p. 10

6.1.3 Hybrid augmentations

  • Hybrid augmentation changes both node features and adjacency; subgraph sampling selects nodes and links using uniform, random-walk, or top-k importance-based strategies.Common hybrid views also combine other augmentation strategies, as in GRACE and MVGRL.Source: §6.1.3 · pp. 10–11

6.2.1 Same-Scale Contrast

  • Same-scale methods contrast nodes across views or graph-level representations, using negative-sample objectives or negative-free knowledge distillation.GRACE contrasts same-node views against other nodes; BGRL uses BYOL-style momentum distillation, while GraphCL(G) applies SimCLR at graph level.Source: §6.2.1 · pp. 11–12

6.2.2 Cross-Scale Contrast

  • Cross-scale methods contrast representations across granularities: patch-global approaches pair nodes with graph summaries, while context-global approaches pair subgraphs with global representations.DGI establishes node-to-graph contrast, MVGRL contrasts across augmented views, and SUGAR selects informative subgraphs for contrast with a sketched graph.Source: §6.2.2 · pp. 12–14

6.3 Mutual Information Estimation

  • Contrastive objectives distinguish positive instance pairs from negatives by pulling their representations together and pushing negative pairs apart.Common estimators include Jensen–Shannon divergence and InfoNCE.Source: §6.3 · p. 14
  • Jensen–Shannon divergence is a common, efficient lower-bound MI estimator, while InfoNCE is a lower-bound estimator with one positive and N negative pairs.InfoNCE can use a temperature-scaled dot-product discriminator, as in GRACE and GCC.Source: §6.3.1 · p. 14, §6.3.2 · p. 14

6.3.3 Triplet Loss

  • Triplet-margin loss does not guarantee MI maximization because it cannot represent an MI lower bound.Jiao et al. define this loss for estimating MI between data instances.Source: §6.3.3 · p. 14

6.3.4 BYOL Loss

  • BYOL-inspired methods such as BGRL avoid negative samples and can be interpreted as non-bound MI estimators.An online predictor and a mean-square-error pretext decoder help prevent collapse, alongside mechanisms such as momentum encoders and stop-gradient.Source: §6.3.4 · pp. 14–15

6.3.5 Barlow Twins Loss

  • Barlow Twins uses a redundancy-reduction loss to avoid negatives; same-scale methods commonly use InfoNCE or BYOL, while DGI-derived cross-scale methods often prefer JSD.The objective is defined for representations of two views of a batch; the survey contrasts typical loss choices by scale.Source: §6.3.5 · p. 15

7 HYBRID METHODS

Hybrid graph SSL combines multiple pretext tasks in a weighted multi-task objective, bringing together generation, contrastive learning, and auxiliary-property supervision.

7 HYBRID METHODS

  • Hybrid methods jointly train on two or more self-supervised objectives, combining their losses with trade-off weights.The formulation assigns each task a loss, pretext decoder, and data distribution.Source: §7 · p. 15
  • Generation-based hybrids combine feature and structure tasks, including GPT-GNN, Graph-Bert [39], PT-DGNN [105], and Manessi et al..GPT-GNN masks nodes and edges, then reconstructs features with MSE loss and predicts masked edges with contrastive loss; Graph-Bert reconstructs raw attributes and graph diffusion values. PT-DGNN applies this combination to dynamic graphs, while Manessi et al. use three feature-generation tasks.Source: §7 · p. 15
  • Generative-contrastive hybrids pair reconstruction with contrast objectives in GMI, CG3, and MVMI-FT [107].GMI maximizes agreement between node embeddings and neighbor features while reconstructing adjacency; CG3 uses local and global views for contrast and topology reconstruction, and MVMI-FT combines cross-scale contrast with graph reconstruction.Source: §7 · p. 15
  • Multiple contrastive objectives capture different views or scales in GraphLoG, HDMI, G-Zoom [110], and LnL-GNN [111].GraphLoG contrasts subgraph-subgraph, graph-graph, and graph-context pairs; HDMI combines same- and cross-scale objectives, G-Zoom uses three same-scale contrast levels, and LnL-GNN maximizes mutual information across local and non-local neighborhoods.Source: §7 · pp. 15–16
  • Auxiliary-property hybrids combine property tasks with one another or with generation, including Hu et al. [50], GROVER, and Kou et al..Hu et al. combine denoising link reconstruction, centrality ranking, and cluster preservation; GROVER predicts contextual node properties and graph-level motifs, while Kou et al. mix generation and property classification in a clustering model.Source: §7 · p. 16

8 EMPIRICAL STUDY

The section compares representative graph SSL methods on node and graph classification and collects resources for empirical evaluation. Node-classification results show that performance varies with pretext-task design and label access.

Performance Comparison of Node Classification

  • Node classification is evaluated on citation networks in transductive learning and PPI in inductive learning, with accuracy as the metric; graph-classification results and dataset and code resources are in Appendices D–F.The transductive datasets are Cora, Citeseer, and Pubmed; the inductive evaluation uses PPI. Appendices E and F describe benchmark datasets and open-source implementations, respectively.Source: §8 · p. 16
  • BGRL’s BYOL loss and G-BT’s Barlow Twins loss do not yield superior performance, indicating that transferred objectives do not always improve graph SSL.The comparison concerns methods adopting advanced contrastive objectives from visual learning.Source: §8 · p. 16, Table 5 · p. 16
  • Hybrid methods are competitive, with some outperforming supervised baselines; G-Zoom [110] combines contrastive tasks at three levels.The authors attribute gains to supervision signals from diverse perspectives.Source: Table 5 · p. 16, §8 · p. 16
  • MVGRL, MERIT, and SubG-Con outperform generalization-based and auxiliary property-based methods, while early random-walk and autoencoder methods underperform most graph SSL methods.The authors suggest that simple unsupervised targets fail to fully exploit the data for supervision signals.Source: §8 · p. 16, Table 5 · p. 16
  • Methods trained with accessible labels in the PF/JL group generally outperform URL methods, whose encoders are trained only with SSL pretext tasks.URL representations feed directly into classification decoders; PF/JL encoders can access training labels.Source: §8 · p. 16

9 PRACTICAL APPLICATIONS

Graph SSL has been applied in recommender systems, anomaly detection, and chemistry. Across these fields, researchers use generation, contrastive, hybrid, and auxiliary-property tasks to address domain-specific problems.

Recommender Systems

  • Graph SSL supports recommendation tasks including cold-start handling, model pre-training, and selection-bias reduction.Hao et al. use reconstruction for cold-start users and items; S2-MHCN and DHCN use contrastive hypergraph learning, Liu et al. [117] use a debiased contrastive loss, and PMGT uses two generation tasks for multimodal side information.Source: §9 · pp. 16–17

Anomaly Detection

  • Graph SSL suits unsupervised anomaly detection, using hybrid generation, contrastive learning, and hop-count prediction.DOMINANT, SpecAE, AEGIS, and SL-GAD combine structure and feature generation; CoLA and ANEMONE use contrastive learning, while HCM [125] predicts node-pair hop counts.Source: §9 · p. 17

Chemistry

  • In chemistry, graph SSL learns molecular representations and supports drug–target and drug-drug interaction prediction.MolCLR [126] and CKGNN [127] use graph-level contrastive pretext tasks; GraphMS and MIRACLE address drug–target and drug-drug interaction prediction. GROVER and Hu et al. are noted as previously reviewed molecule-focused SSL work.Source: §9 · p. 17

10 FUTURE DIRECTIONS

The section identifies unresolved theoretical, methodological, and application challenges in graph SSL and outlines directions for addressing them. It calls for stronger foundations and methods that reflect graph data characteristics and support broader use.

Theoretical Foundation

  • Graph SSL needs a theoretical foundation connecting empirical methods to graph signal processing and spectral graph theory; MI estimator choices remain empirical,.Most methods are designed intuitively and assessed experimentally, despite the support MI estimation theory provides for some contrastive-learning work.Source: §10 · p. 17

Interpretability and Robustness

  • Graph SSL should become more explainable and robust, since black-box methods often overlook interpretability and real-world graph data can be noisy or adversarial,.This matters for risk-sensitive, privacy-related applications such as fraud detection.Source: §10 · p. 17

Pretext Tasks for Complex Types of Graphs

  • Graph SSL should develop pretext tasks for heterogeneous, spatial-temporal, dynamic, and hypergraph data that capture their distinctive characteristics.MI maximization has been used for complex graphs, but can miss rich signals such as temporal dynamics.Source: §10 · p. 17

Augmentation for Graph Contrastive Learning

  • Graph contrastive learning needs more diverse augmentations with reliable invariance, beyond masking, edge modification, subgraph sampling, or graph diffusion.Directions include adaptive augmentation, automatic selection, and stronger augmented samples [67] informed by graph structure and attributes.Source: §10 · p. 17

Learning with Multiple Pretext Tasks

  • Graph SSL should explore advanced hybrid methods that adaptively combine multiple pretext tasks to provide diverse supervision and more informative representations.Most approaches use one pretext task, while only a few hybrid methods combine tasks.Source: §10 · p. 17

Broader Scope of Applications

  • Graph SSL has untapped potential beyond recommender systems, anomaly detection, and chemistry, including financial networks, cybersecurity, community detection, and federated learning.The paper points to broad application prospects where manual labeling is costly, especially when annotation depends heavily on domain knowledge.Source: §10 · p. 17

11 CONCLUSION … B.4 Commonly Used Loss Functions

The paper surveys graph self-supervised learning (SSL) through a unified framework and four-part taxonomy, and assembles resources, applications, and future research directions. Its appendices provide a timeline and reference definitions for graph types, GNNs, and common losses.

11 CONCLUSION

  • The survey organizes graph SSL into generation-based, auxiliary property-based, contrast-based, and hybrid approaches, and reviews their methods, applications, resources, challenges, and future directions.It also collects datasets, evaluation methods, performance comparisons, and open-source codes.Source: §11 · p. 18

B.1 Notations

  • The paper uses bold uppercase and lowercase letters for matrices and vectors, respectively, and calligraphic fonts for sets; its notations are summarized in Table 6.The table is identified as a summary of commonly used notations and explanations.Source: §B.1 · p. 18, Table 6 · p. 18

B.2 Dynamic and Heterogeneous Graphs

  • Dynamic graphs have node and edge sets that vary over time, while heterogeneous graphs have multiple node and/or edge types; spatial-temporal, bipartite, and multiplex graphs are specified as special cases.Spatial-temporal graphs have fixed adjacency and time-varying features; bipartite graphs have two node types and one edge type, while multiplex graphs have one node type and multiple edge types.Source: §B.2 · p. 18

B.3 Graph Neural Networks and Readout Layers

  • A GNN learns node representations using aggregation and combination functions, while a readout function fuses node representations into a graph representation.Readout can use permutation-invariant summation or pooling algorithms such as DiffPool and HaarPool.Source: §B.3 · pp. 18–19

APPENDIX A TIMELINE OF GRAPH SELF-SUPERVISED LEARNING

  • The timeline traces early random-walk and graph-autoencoder methods, DGI in 2019, the first auxiliary property-based and hybrid methods in 2020, and advanced techniques in 2021.Examples include DeepWalk, LINE, node2vec, GAE/VGAE, adaptive augmentation GCA, JOAO, and AD-GCL.Source: Appendix A · p. 18, Figure 9 · p. 19

B.4 Commonly Used Loss Functions

  • Mean squared error compares predicted and target vectors elementwise and can also be applied to matrices and scalars.The appendix defines MSE as a commonly used loss function.Source: §B.4 · p. 19
  • Cross-entropy uses a softmax prediction and one-hot target, while binary cross-entropy applies to predicted and target scalars in [0, 1] and {0, 1}, respectively.The appendix presents CE and BCE as commonly used loss functions.Source: §B.4 · p. 19

APPENDIX C DOWNSTREAM TASKS … FICATION

Appendix C describes supervised downstream objectives for node-, edge-, and graph-level tasks. The graph-classification evaluation reports that URL methods underperform supervised GIN overall, while cross-scale contrast methods generally outperform same-scale methods.

C.1 Node-level tasks

  • Node classification predicts each node’s discrete label from its embedding, typically using an MLP with Softmax and cross-entropy trained on labeled nodes.The training-node set contains nodes with known labels, whose embeddings are selected from the full embedding matrix.Source: §C.1 · p. 19

C.2 Link-level tasks

  • Edge classification predicts a label for each edge from its endpoint embeddings, using a classifier and cross-entropy loss on training edges.The endpoint embeddings are selected from the full embedding matrix.Source: §C.2 · pp. 19–20

C.3 Graph-level Tasks

  • Graph classification predicts each graph’s label by aggregating node embeddings into a graph representation, then applying a Softmax classifier trained with cross-entropy.The objective is trained on a designated set of training graphs.Source: §C.3 · p. 20

FICATION

  • The graph-classification evaluation uses accuracy on nine benchmark datasets, the split from, and GIN as its supervised baseline.Methods are grouped into URL and PF/JL schemes.Source: FICATION · p. 20
  • URL methods overall underperform supervised GIN, cross-scale contrast methods generally outperform same-scale methods, and graph labels in PF/JL do not always improve encoders over URL.The observations are based on graph-classification results across benchmark datasets.Source: FICATION · p. 20
  • SSL-pretrained encoders can negatively transfer to graph-level downstream tasks, motivating further investigation into how to avoid this effect.The passage cites GCC as an example where involving graph labels does not necessarily improve encoders on every dataset over pure URL.Source: FICATION · p. 20
  • Different methods are evaluated on different graph-dataset sets, so unified benchmarks for graph-level SSL evaluation are needed.Source: FICATION · p. 20

APPENDIX E DATASETS … E.4 Bio-chemical Graphs

The appendix surveys commonly used graph datasets across citation, co-purchase, social, and bio-chemical domains, with benchmark statistics summarized in Table 8.

APPENDIX E DATASETS

  • The appendix covers four dataset categories—citation, co-purchase, social, and bio-chemical graphs—and summarizes selected benchmark statistics in Table 8.Source: Appendix E · p. 20, Table 8 · p. 21

E.1 Citation Networks

  • Citation-network benchmarks use papers or authors as nodes, relationships such as citations or co-authorship as edges, text-derived features, and field-of-study labels,,,.Source: §E.1 · p. 20

E.2 Co-purchase Networks

  • Amazon co-purchase graphs represent goods as nodes, frequent co-purchases as edges, review-derived features, and product-category labels.Source: §E.2 · p. 20

E.3 Social Networks

  • Reddit data represents forum posts as nodes and user comments as edges, with community labels and features from text embeddings, post scores, and comment counts.Source: §E.3 · p. 20
  • Social graph benchmarks include movie-cast graphs, discussion-thread graphs, and researcher ego-networks labeled by research field.IMDB-B and IMDB-M connect actors sharing a movie; REDDIT-BINARY and REDDIT-MULTI-5K represent discussion threads; COLLAB graphs are researchers’ ego-networks.Source: §E.3 · pp. 20, 22

E.4 Bio-chemical Graphs

  • Bio-chemical datasets include 24 PPI graphs of interacting proteins, MUTAG nitro compounds labeled aromatic or heteroaromatic, and PROTEINS and D&D graphs labeled enzyme or non-enzyme,.Source: §E.4 · p. 22
  • Chemical benchmarks label PTC and PTC-MR compounds for rat carcinogenicity,, NCI-1 and NCI-109 for cancer-cell-line growth inhibition, and eight OGB molecule datasets for specific properties.The OGB datasets are BBBP, Tox21, ToxCast, SIDER, ClinTox, MUV, HIB, and BACE.Source: §E.4 · p. 22

APPENDIX F OPEN-SOURCE IMPLEMENTATIONS · APPENDIX G OTHER GRAPH SSL APPLICATIONS

Appendix F collects available open-source implementations of reviewed graph SSL methods in Table 9, while Appendix G surveys applications spanning expert finding, program repair, scientific domains, and federated learning.

APPENDIX F OPEN-SOURCE IMPLEMENTATIONS

  • Available open-source implementations of reviewed graph SSL methods are linked in Table 9.The survey collects code when it exists for a reviewed method.Source: Appendix F · p. 23

APPENDIX G OTHER GRAPH SSL APPLICATIONS

  • COAD [152] uses same-scale contrastive pretraining for zero-shot expert linking, while SLAPS, SCRL, and SDGE [155] address graph structure learning, few-label learning, and community detection.SLAPS adds denoising node-feature generation to classification; SCRL uses prototype-based pretext tasks; SDGE uses a same-scale contrastive objective.Source: Appendix G · p. 23
  • DrRepair predicts repaired program lines from automatically corrupted programs, while C-SWM uses node-level contrastive learning to train structured world models of multi-object systems.Source: Appendix G · p. 23
  • Sehanobish et al. [158] train a GAT with clustering-based auxiliary-property classification and use learned edge weights as edge features on SARS-CoV-2 and COVID-19 biological datasets.Source: Appendix G · p. 23
  • Sun et al. use hybrid contrastive learning to maximize patch- and graph-level agreement on anatomy-aware graphs extracted from medical images.The application is learning representations for medical images.Source: Appendix G · p. 23
  • FedGL [162] supports federated learning with auxiliary-property classification for a global view of client training, using server-side weighted fusion of clients’ node predictions as pseudo labels.The passage also cites federated learning,.Source: Appendix G · p. 23

Related papers

Loading 2103.00111v5…