DIB-OD: Preserving the Invariant Core for Robust Heterogeneous Graph Adaptation via Decoupled Information Bottleneck and Online Distillation

Yang Yan; Yunxuan Li; Qiuyan Wang; Tianjin Huang; Qiudong Yu

cs.LGApr 13, 2026

DIB-OD: Preserving the Invariant Core for Robust Heterogeneous Graph Adaptation via Decoupled Information Bottleneck and Online Distillation

Authors: Yang Yan, Yunxuan Li, Qiuyan Wang, Tianjin Huang, Qiudong Yu

Organizations: School of Information Technology and Engineering, Tianjin University of Technology and Education, Tianjin, China · School of Computer Science and Technology, Tiangong University, Tianjin, China · Department of Computer Science at University of Exeter

Abstract

Graph pre-training can facilitate knowledge transfer across graph datasets, but severe structural and feature shifts may cause negative transfer and adaptation-induced overwriting of reusable knowledge. We propose DIB-OD, a heterogeneous graph adaptation framework that combines a Decoupled Information Bottleneck with Online Distillation. A multiview teacher first learns a compressed, task-relevant representation, which is distilled into a transferable core branch and a complementary residual branch. Mutual-information objectives and HSIC-based dependence regularization discourage information overlap between the branches, while a confidence-aware semantic regularizer and a frozen teacher anchor reliable pretrained information during target-domain adaptation. Experiments on seven graph-classification datasets spanning chemical, biological, and social-network domains show consistent improvements over representative baselines, particularly under challenging cross-type transfer.

Figures & tables

Datasets	$N\rightarrow F$	$M\rightarrow N$	$F\rightarrow M$	$M\rightarrow F$	$P\rightarrow D$	$D\rightarrow P$	$I\rightarrow d$	$d\rightarrow I$
GAT	44.67 ± 0.12	49.71 ± 0.57	55.46 ± 0.13	53.63± 0.55	50.73 ± 1.34	60.58 ± 0.71	43.39 ± 0.05	48.94 ± 1.16
DANN	45.40 ± 1.36	49.45 ± 0.34	55.61 ± 0.46	48.73 ± 3.21	43.41 ± 4.74	50.37 ± 1.09	46.18 ± 1.50	50.92 ± 1.84
UDAGCN	56.57 ± 2.80	48.62 ± 6.98	50.61 ± 3.53	48.35 ± 5.45	60.68 ± 0.15	49.47 ± 1.14	49.47 ± 1.14	50.76 ± 0.11
DANE	51.24 ± 2.70	57.41 ± 0.99	52.71 ± 5.73	56.61 ± 2.69	48.72 ± 2.64	61.84 ± 2.10	48.72 ± 2.64	46.30 ± 1.29
ASN	53.22 ± 4.29	49.62 ± 0.06	44.64 ± 0.01	44.64 ± 0.00	56.64 ± 0.00	58.66 ± 0.00	43.19 ± 0.00	50.01 ± 0.02
GRADE	53.22 ± 4.29	50.80 ± 1.76	51.07 ± 5.25	46.80 ± 4.30	44.33 ± 3.11	42.21 ± 2.38	47.50 ± 0.63	50.86 ± 0.00

Table 1: Intra-Domain Generalization Performance

Datasets	$N\rightarrow P$	$P\rightarrow N$	$M\rightarrow P$	$P\rightarrow M$	$M\rightarrow D$	$D\rightarrow M$	$I\rightarrow P$	$D\rightarrow N$
GAT	59.57 ± 0.01	50.38 ± 0.01	52.56 ± 1.18	55.35 ± 0.23	49.19 ± 1.20	44.78 ± 0.18	40.42 ± 0.03	49.38 ± 0.38
DANN	55.58 ± 8.72	50.38 ± 0.00	61.40 ± 2.23	44.84 ± 0.41	44.64 ± 0.00	44.60 ± 0.07	40.38 ± 0.07	50.87 ± 0.90
UDAGCN	44.87 ± 0.25	45.56 ± 4.59	54.16 ± 0.83	73.29 ± 0.00	51.15 ± 3.11	52.20 ± 2.88	49.98 ± 0.23	43.30 ± 0.95
DANE	34.91 ± 3.44	41.89 ± 2.29	52.59 ± 2.08	59.75 ± 1.13	47.26 ± 7.32	46.11 ± 5.16	60.62 ± 3.88	47.63 ± 3.45
ASN	40.43 ± 0.00	49.62 ± 0.01	59.57 ± 0.00	55.34 ± 0.02	58.66 ± 0.00	55.36 ± 0.00	59.75 ± 0.01	49.62 ± 0.03
GRADE	59.48 ± 0.08	50.51 ± 0.69	49.31 ± 8.60	48.94 ± 5.27	47.79 ± 8.91	58.53 ± 7.28	40.49 ± 0.49	50.08 ± 0.37

Table 2: Cross-Type Data Performance Comparison

Datasets	$N\rightarrow F$	$F\rightarrow N$	$M\rightarrow N$	$P\rightarrow D$	$N\rightarrow P$	$P\rightarrow N$	$M\rightarrow D$	$I\rightarrow P$
w/o IB	56.75 ± 1.70	69.64 ± 2.56	51.22 ± 2.31	70.37 ± 8.19	59.57 ± 0.01	68.04 ± 2.66	73.59 ± 5.20	69.73 ± 5.36
w/o HSIC	56.56 ± 2.72	69.30 ± 3.03	68.21 ± 2.64	72.82 ± 6.16	67.30 ± 5.31	67.70 ± 2.98	73.69 ± 5.38	67.66 ± 3.43
w/o SSR	56.79 ± 2.60	52.10 ± 3.70	51.44 ± 5.72	73.51 ± 4.84	53.42 ± 2.25	68.06 ± 2.52	72.58 ± 4.53	68.46 ± 5.87
Full Fine-tuning	56.42 ± 3.37	52.56 ± 2.61	52.29 ± 3.64	73.34 ± 4.05	69.63 ± 3.65	67.75 ± 3.12	67.84 ± 2.26	67.94 ± 3.16
Ours	57.30 ± 3.04	69.83 ± 2.91	68.50 ± 2.67	73.77 ± 2.90	70.17 ± 2.41	69.28 ± 2.71	74.53 ± 3.97	70.80 ± 3.33

Table 3: Ablation Results (%) on Within-type and Cross-type Graph Transfer

Panel of Figure 2 — Figure 2: Changes in mutual information during cross-domain adaptation.

Online Knowledge Distillation

See all 10 papers

Graph Domain Adaptation Does Not End with Representation Learning
Sep 22, 2026Ziqian Liu, Yongxue Xu, Enze Zhang +3Unsupervised Domain Adaptation Graph Representation Learning
Sun Yat-sen University, China · University of Wisconsin–Madison, USA · Jiangsu University, China +2
MDGMIX: Boundary-Aware Subgraph Mixing for Multi-Domain Graph Pre-Training
May 25, 2026Ziyu Zheng, Yaming Yang, Ziyu Guan +2Cross-Domain Generalization Cross-Domain Transfer Learning
School of Computer Science and Technology, Xidian University, Xi’an, China · School of Artificial Intelligence, Xidian University, Xi’an, China
DSBD: Dual-Aligned Structural Basis Distillation for Graph Domain Adaptation
Apr 3, 2026Yingxu Wang, Kunyu Zhang, Jiaxin Huang +4Graph Domain Adaptation Domain Adaptation
Zhengzhou University · The Hebei University of Technology · California State Polytechnic University +1

Sep 22, 2026J/K move · Enter open · S save

Graph Domain Adaptation Does Not End with Representation Learning

Ziqian Liu, Yongxue Xu, Enze Zhang, Jiaqi Zhang+2

Sun Yat-sen University, China · University of Wisconsin–Madison, USA · Jiangsu University, China+2

Unsupervised Domain Adaptation Graph Representation Learning Domain Adaptation Graph Domain Adaptation Node Classification

Open paper AskPDF arXiv

Graph domain adaptation (GDA) transfers knowledge from a labeled source graph to an unlabeled target graph under shifts in both node attributes and graph structure. Existing methods primarily adapt graph representations through propagation redesign, distribution alignment, or source-to-target transition modeling, but still rely on a single graph-propagating path for target prediction. This leaves open whether an adapted graph representation exhausts the predictive evidence available in the target domain, since the graph-aware expert and graph-free local expert may exhibit different failure modes under topological shifts. To address this limitation, we propose EviGDA, an Evidence-Augmented Graph Domain Adaptation framework that complements graph representation adaptation with a graph-free local expert. The graph-aware expert performs message passing and entropy-aware marginal alignment, while the graph-free local expert learns solely from source node features and labels without graph propagation or target alignment. The two experts are optimized independently and combined only at inference through a task-level constant probability mixture, preserving complementary evidence without joint training, learned routing, or target pseudo-labels. Extensive experiments on ten datasets and 16 transfer tasks show that EviGDA outperforms state-of-the-art baselines. Complementary prediction paths add value beyond graph representation alignment.

Abstract

Online Knowledge Distillation

Explore similar work

Graph Domain Adaptation Does Not End with Representation Learning

MDGMIX: Boundary-Aware Subgraph Mixing for Multi-Domain Graph Pre-Training

DSBD: Dual-Aligned Structural Basis Distillation for Graph Domain Adaptation