Source-linked AI summary
Graph Transfer Learning via Adversarial Domain Adaptation with Graph Convolution
Quanyu Dai, Xiao-Ming Wu, Jiaren Xiao, Xiao Shen, Dan Wang
TL;DR
Scarce labels and differences between network domains make cross-network node classification challenging. AdaGCN combines graph convolution with adversarial domain adaptation, and experiments report effectiveness across unlabeled and scarcely labeled target settings, with label efficiency and robustness to distribution discrepancy.
Problem
Method
AdaGCN combines graph convolution for class-discriminative node representations with adversarial domain adaptation for domain-invariant representations.
Source: §1 · p. 2Results
Experiments on real-world information networks report superior performance to state-of-the-art baselines, label efficiency, and robustness against distribution discrepancy.
Source: §1 · p. 2Takeaways & Limitations
AdaGCN is reported to work with completely unlabeled or scarcely labeled target networks and to depend little on attributes shared between networks.
Source: §1 · p. 2The paper identifies transfer from multiple source networks and conditional adversarial adaptation for multimodal distributions as future directions.
Source: §6 · p. 13
Abstract
from arXiv · showhide
This paper studies the problem of cross-network node classification to overcome the insufficiency of labeled data in a single network. It aims to leverage the label information in a partially labeled source network to assist node classification in a completely unlabeled or partially labeled target network. Existing methods for single network learning cannot solve this problem due to the domain shift across networks. Some multi-network learning methods heavily rely on the existence of cross-network connections, thus are inapplicable for this problem. To tackle this problem, we propose a novel \textcolor{black}{graph} transfer learning framework AdaGCN by leveraging the techniques of adversarial domain adaptation and graph convolution. It consists of two components: a semi-supervised learning component and an adversarial domain adaptation component. The former aims to learn class discriminative node representations with given label information of the source and target networks, while the latter contributes to mitigating the distribution divergence between the source and target domains to facilitate knowledge transfer. Extensive empirical evaluations on real-world datasets show that AdaGCN can successfully transfer class information with a low label rate on the source network and a substantial divergence between the source and target domains. The source code for reproducing the experimental results is available at https://github.com/daiquanyu/AdaGCN.
1 INTRODUCTION
The paper frames cross-network node classification as a way to address scarce labels, domain divergence, and missing cross-network edges. AdaGCN combines graph convolution with adversarial domain adaptation, and experiments report strong performance, label efficiency, and robustness to distribution discrepancy.
1 INTRODUCTION
- Cross-network classification must handle domain divergence, few labeled source nodes, and absent cross-network edges, while target networks may be unlabeled or partially labeled.Single-network methods face domain-shift difficulties, while some multi-network methods depend on cross-network connections.Source: §1 · pp. 1–2
- Prior graph domain adaptation work is described as limited: CDNE does not jointly model network structure and node attributes and relies on preprocessing that densifies sparse adjacency matrices.The denser matrices increase computational complexity for its autoencoder-based architecture.Source: §1 · p. 2
- AdaGCN combines semi-supervised graph convolution for class-discriminative representations with adversarial domain adaptation for domain-invariant representations.The semi-supervised component uses available source and target labels; adversarial learning pits a representation learner against a domain critic.Source: §1 · p. 2
- AdaGCN demonstrates superior performance to state-of-the-art baselines, label efficiency, and robustness against distribution discrepancy on real-world information networks.The experiments also cover completely unlabeled and scarcely labeled target networks and report low dependence on attributes shared across networks.Source: §1 · p. 2
2 RELATED WORK
The related work reviews single-network learning, multi-network learning, and domain adaptation. It identifies a gap in methods that can transfer node-classification knowledge across graph domains without relying on cross-network links.
2.1 Single Network Learning
- Single-network embedding methods often learn representations without tailoring them to node classification, whereas semi-supervised graph methods integrate topology, attributes, and observed labels.GCN combines local node attributes and graph topology; GraphSAGE varies aggregation, and GAT uses attention to aggregate neighbor features.Source: §2.1 · p. 3
2.2 Multi-Network Learning
- EOE and DMNE model cross-network relationships but rely on cross-network connections, while network alignment instead assumes common nodes and seeks node correspondences.Other transfer work addresses tasks including social-tie inference, link prediction, and node classification; the paper focuses on single-source node classification with graph convolution and adversarial adaptation.Source: §2.2 · p. 3
2.3 Domain Adaptation
- Adversarial domain adaptation learns representations that reduce source–target divergence, but prior approaches chiefly address vector data with IID samples rather than correlated graph entities.The reviewed approaches include DANN’s generator–discriminator game and methods using Wasserstein distance or discriminative information for alignment.Source: §2.3 · pp. 3–4
3 PROBLEM DEFINITION
The problem is to classify nodes in a target network using knowledge from a partially labeled source network despite domain divergence. The networks may have different attribute sets, which the formulation combines into a shared feature space.
3 PROBLEM DEFINITION
- The formulation represents each network with nodes, a weighted adjacency matrix, and node attributes, and allows source and target networks to have different attributes.It forms a combined attribute set from both networks and reformulates their feature matrices to include that set; an absent node-attribute association is represented by zero.Source: §3 · p. 4
- The task is to learn a classifier for the target network using a partially labeled source network, although the domains differ and share the same label space.The target may be completely unlabeled or partially labeled; the formal setup assumes an unlabeled target for simplicity and says the method extends to partial labels.Source: §3 · p. 4
- A network domain is defined as an attributed network together with its node-classification function.The source and target domains are distinct, but use the same label space.Source: §3 · p. 4
4 PROPOSED METHOD
AdaGCN combines graph convolution and adversarial domain adaptation to learn class-discriminative, domain-invariant representations for cross-network node classification. Its training couples source-label classification with domain alignment, and its graph convolution can be tuned to control smoothing strength.
4.1 An Overview of Model Architecture
- AdaGCN pairs graph convolution with adversarial domain adaptation to address domain divergence, absent cross-network edges, and few common attributes.A shared representation learner feeds a label classifier and a domain critic, supporting target classification from source labels and extending to partially labeled targets.Source: §4.1 · pp. 4–5
4.2 Network Representation Learning
- IGCN raises the graph filter to a tunable exponent nI, controlling smoothing strength while addressing overfitting risks from stacking GCN layers.The paper also describes standard GCN filtering as Laplacian smoothing, averaging each node's features with its neighbors'.Source: §4.2 · p. 5
4.3 Semi-Supervised Learning
- The label classifier trains on labeled source nodes with cross-entropy, and available target labels can be added for semi-supervised learning.The classifier supports multi-class softmax or multi-label sigmoid outputs.Source: §4.3 · p. 6
4.4 Adversarial Domain Adaptation
- The adversarial component aligns source and target representation distributions using a Wasserstein-1 objective and a domain critic.The critic is constrained to be 1-Lipschitz; gradient penalty supports that constraint, and the generator minimizes the Wasserstein distance after critic training.Source: §4.4 · p. 6
4.5 Overall Loss and Model Training
- Training alternates critic updates with joint representation-learner and classifier updates that minimize classification and domain-adaptation losses.The described procedure uses full-batch gradient descent, while mini-batch training is also possible with existing methods.Source: §4.5 · p. 7
4.6 Time Complexity Analysis
- The model's overall time complexity is linear in network size, with GCN computation linear in the number of edges and classifier and critic computation linear in nodes.The IGCN layer adds a constant factor nI to the GCN-layer complexity.Source: §4.6 · p. 7
5 EXPERIMENTS
The experiments evaluate multi-label transfer across three attributed citation-network domains and six source-to-target tasks. They compare AdaGCN variants with methods from several learning families under shared implementation settings.
5 EXPERIMENTS
- The evaluation asks how methods compare, how source and target label rates and domain discrepancy affect transfer, and how graph convolution and hyperparameters matter.The section also visualizes learned node embeddings.Source: §5 · p. 7
5.1.1 Datasets
- Experiments examine six transfer directions among DBLPv7, Citationv1, and ACMv9, using papers' topic categories as labels and title keywords as attributes.The networks cover different publication periods and originate from DBLP, Microsoft Academic Graph, and ACM.Source: §5.1.1 · p. 7
5.1.2 Baselines
- Baselines span network embedding, graph-based semi-supervised learning, attribute-only domain adaptation, and network transfer methods.Comparisons include GCN and GraphSAGE, WDGRL, and NetTr and CDNE.Source: §5.1.2 · pp. 7–8
5.1.3 Implementation Details
- AdaGCN’s source and target GCNs each have three hidden layers sized 1000-100-16; AdaIGCN differs only in its representation learner, using one IGCN layer and two fully connected layers.All transfer learning tasks use the same parameter configurations unless otherwise specified; the model is trained for 1000 epochs.Source: §5.1.3 · p. 8
5.2 Performance Comparison (RQ1)
AdaGCN transfers labels from a partially labeled source to an unlabeled target, outperforming most baselines in the unsupervised setting and GCN in the semi-supervised setting. AdaIGCN improves further, while adding even a small number of artificial cross-network edges degrades combined-network GCN performance.
5.2.1 Unsupervised Setting: Partially Labeled Source Network and Completely Unlabeled Target Network
- AdaIGCN improves on AdaGCN by 4.83% Micro-F1 and 4.28% Macro-F1 on average, while AdaGCN gains 13.54% Micro-F1 and 19.03% Macro-F1 over GCN.The authors attribute the further gains to IGCN’s improved node representations and the AdaGCN gains to adversarial adaptation.Source: §5.2.1 · p. 8
- AdaGCN outperforms all baselines in five of six unsupervised tasks and matches the best baseline on the sixth, with average relative gains over CDNE of 4.41% Micro-F1 and 5.81% Macro-F1.The setting uses a 10% labeled source and a completely unlabeled target.Source: §5.2.1 · p. 8, Table 3 · p. 9
5.2.2 Semi-Supervised Setting: Partially Labeled Source and Target Networks
- With 10% source and 5% target training rates, AdaIGCN gains 2.40% Micro-F1 and 2.04% Macro-F1 over AdaGCN, and 3.14% Micro-F1 and 3.55% Macro-F1 over the best baseline CDNE.All models perform better than in the unsupervised setting; AdaGCN consistently outperforms GCN across tasks.Source: Table 4 · p. 9, §5.2.2 · p. 10
5.2.1 Unsupervised Setting: Partially Labeled Source Network and Completely Unlabeled Target Network
- Topology-only DeepWalk and node2vec perform poorly across transfer tasks, while WDGRL also performs poorly and is worse than DNNs on some tasks.The authors relate these outcomes to incomparable independently learned representations and to methods’ inability to use network structure, respectively.Source: §5.2.1 · pp. 8–9
5.2.2 Semi-Supervised Setting: Partially Labeled Source and Target Networks
- Adding artificial cross-network edges at rates of 10^-5% or 10^-3% degrades combined-network GCN performance, while AdaGCN retains a clear advantage.The comparison also considers zero artificial edges.Source: §5.2.2 · p. 10
5.3 Effect of Training Rate (RQ2)
Across varying source and target training rates, AdaGCN and AdaIGCN outperform the baselines in the reported settings. AdaIGCN remains effective with few source labels and improves further over AdaGCN, particularly at low target training rates.
5.3.1 Effect of Source Training Rate
- For almost all baselines except DeepWalk, performance first improves and then stabilizes as the source training rate increases.DeepWalk performance decreases as the source training rate increases.Source: §5.3.1 · p. 10
- AdaIGCN achieves remarkably good performance with only 5% labeled source nodes, while AdaGCN and AdaIGCN consistently outperform all baselines across the four reported tasks and source training rates.The source training rate ranges from 5% to 90%, and the target is unlabeled.Source: §5.3.1 · p. 10, Figure 3 · p. 11
5.3.2 Effect of Target Training Rate
- AdaGCN improves over GCN by 5.08% on average across the two reported tasks at target training rates from 1% to 10%.The source training rate is fixed at 10%; the reported tasks are Citationv1→DBLPv7 and DBLPv7→Citationv1.Source: §5.3.2 · p. 10, Figure 4 · p. 11
- At a 1% target training rate on Citationv1→DBLPv7, AdaIGCN gains 3.90% over AdaGCN.AdaIGCN improves consistently over AdaGCN, with a larger gap at low target training rates.Source: Figure 4 · p. 11, §5.3.2 · p. 11
5.4 Effect of Distribution Discrepancy (RQ3)
AdaGCN and AdaIGCN outperform GCN across the tested common-attribute rates and both transfer tasks, including settings with large distribution shifts. AdaIGCN consistently performs better than AdaGCN.
5.4 Effect of Distribution Discrepancy (RQ3)
- The experiment varies the common-attribute rate by deleting shared attributes; lower rates represent larger distribution discrepancy.The source training rate is 10%, and Citationv1 is the target network in both tasks.Source: §5.4 · p. 11
- Across common-attribute rates of 10%–50%, AdaGCN gains 22.88% Micro-F1 and 24.93% Macro-F1 over GCN on DBLPv7→Citationv1, and 9.02% Micro-F1 and 14.56% Macro-F1 on ACMv9→Citationv1.AdaGCN and AdaIGCN outperform GCN on both tasks throughout the tested rates; AdaIGCN consistently performs better than AdaGCN.Source: Figure 5 · p. 11, §5.4 · p. 11
5.5 Effect of Graph Convolution (RQ4)
Graph convolution on node features substantially improves transfer performance, but excessive smoothing can harm it. AdaIGCN’s smoothing parameter therefore has a useful range rather than an unlimited benefit.
5.5 Effect of Graph Convolution (RQ4)
- Increasing the IGCN smoothing parameter from 0 to 1 yields a 26.62% relative improvement, while larger settings can over-smooth features and harm transfer performance.On DBLPv7→Citationv1, performance first increases and then slightly drops as the parameter varies from 1 to 25; at 0, AdaIGCN reduces to WDGRL.Source: Figure 6 · p. 12, §5.5 · p. 12
5.6 Parameter Sensitivity (RQ5)
AdaGCN’s sensitivity analysis examines the domain adaptation coefficient, gradient penalty, and domain critic training steps. The reported trends favor balancing the two losses, setting the penalty to 10, and training the critic for at least 10 steps.
5.6 Parameter Sensitivity (RQ5)
- The analysis uses the unsupervised setting with a 10% source training rate and reports Micro-F1 for Citationv1→DBLPv7.Other hyperparameters are fixed at their default settings while each parameter is studied; similar tendencies are reported for other tasks.Source: §5.6 · p. 12
- Micro-F1 slightly improves as λ rises from 0.4 to 1.2, then drops quickly, indicating the two losses need balancing.λ balances the semi-supervised and domain-adaptation losses to learn class-discriminative and domain-invariant representations.Source: §5.6 · p. 12
- The best result occurs at γ = 10; smaller or larger settings might degrade performance.The paper recommends γ = 10 for the gradient penalty when training the domain discriminator.Source: §5.6 · p. 12
- Micro-F1 increases as domain critic training steps nd rise from 5 to 10, then stabilizes.The authors link this pattern to their analysis that the critic should be trained sufficiently close to optimality.Source: §5.6 · p. 12
5.7 Visualization of Node Representations
The visualizations indicate that adversarial domain adaptation aligns representations across networks by category, while the IGCN layer improves within-category compactness and alignment. The resulting clusters are described as discriminative.
5.7 Visualization of Node Representations
- With AdaGCN and AdaIGCN, same-category papers from ACMv9 and Citationv1 cluster together, unlike the mixed categories under GCN and IGCN.Figure 7 uses gray and orange for source-network Databases and Computer Vision papers, and red and green for the corresponding target-network papers.Source: §5.7 · p. 12
- The IGCN layer’s adjustable feature smoothing can make same-category representations more compact without increasing model complexity.The authors report this compactness in Figures 7(b) and 7(d).Source: §5.7 · p. 13
- AdaIGCN aligns source and target representations better than AdaGCN, and the cluster boundary is described as clear.The authors interpret the clear boundary as indicating discriminative learned representations.Source: §5.7 · p. 13
6 CONCLUSION
AdaGCN combines semi-supervised graph convolution with adversarial domain adaptation for cross-network node classification. The paper identifies multi-source transfer and conditional adversarial alignment for multimodal data as future directions.
6 CONCLUSION
- AdaGCN combines semi-supervised graph convolution and adversarial domain adaptation to learn class-discriminative, network-invariant node representations for cross-network classification.The semi-supervised component learns a generalized classifier, while the adversarial component supports knowledge transfer from source to target.Source: §6 · p. 13
- The authors identify transfer from multiple source networks and conditional adversarial adaptation for multimodal distributions as future work.They describe networked-data transfer learning as still being in an early stage.Source: §6 · p. 13