acceptodds
Under review as a conference paper at ICLR 2027

Selectivity Matters: Source Node Influence Pruning for Unsupervised Graph Domain Adaptation

Abstract

Unsupervised Graph Domain Adaptation (UGDA) aims to facilitate knowledge transfer from a labeled source graph to an unlabeled target graph by mitigating cross-domain distribution shifts. Existing methods primarily focus on node-level feature alignment in latent spaces, relying on the implicit assumption that all source nodes contribute positively to the alignment. However, this assumption often fails because a node's semantic information is intrinsically coupled with its topological graph structure. Due to structural shifts, source nodes with severe structural deviations (e.g., structural outliers) lack semantic counterparts in the target graph, and forcing alignment on them introduces severe noise and causes negative transfer. To bridge this gap, we argue that selective source node utilization is superior to full-graph training, thereby shifting the research paradigm from feature-level alignment to data-level refinement. To this end, we propose Source Node Influence Pruning (SNIP), a novel model-agnostic, data-centric refinement framework. Specifically, SNIP quantifies the structural discrepancy between individual source nodes and the target domain by integrating multiple centrality measures, assigning each source node an influence score. A rank-based normalization mechanism is further employed to eliminate scale variations across different measures, allowing SNIP to effectively identify and filter out structurally incompatible nodes with low influence scores. As a plug-and-play method, SNIP constructs a refined “sub-source” graph that is inherently more beneficial for subsequent alignment. Comprehensive experiments across eight transfer scenarios on five real-world datasets demonstrate that SNIP consistently outperforms competitive baselines and significantly enhances adaptation performance, validating the superiority of selective node utilization over full-graph training.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Unsupervised Graph Domain Adaptation, Data-Centric Refinement, Source Node Pruning, Negative Transfer

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_B9gdOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

In this paper, the authors study GDA from a data-centric perspective, where they argue that existing methods implicitly assume all source nodes contribute positively to alignment, while structurally deviated source nodes may cause negative transfer. To address this issue, they propose SNIP, a model-agnostic source node pruning framework. SNIP measures the structural discrepancy between each source node and the target graph using multiple centrality measures, combines them through rank-based normalization, and prunes a predefined proportion of source nodes with the lowest scores before training existing GDA backbones. Experiments on different benchmarks verify the performance of the proposed SNIP.

Critical strengths
  • The paper considers GDA from a data-centric perspective, and examines whether all source nodes contribute positively to cross-domain alignment.
  • Experiments on citation and social network benchmarks verify the performance of the proposed SNIP.
Critical weaknesses
  • This paper is incremental. Selecting or reweighting source samples to reduce domain discrepancy is a classical idea in domain adaptation [1, 2], and data-level refinement of the source graph has also been explored in GDA [3, 4]. SNIP mainly prunes source nodes with several centrality-based heuristics, so I think it does not constitute a new research paradigm from "how to align" to "what to align" as claimed in the main manuscript.
  • Some key claims of the main manuscript are not verified by experiments. For example, the claim that structurally deviated source nodes lack semantic counterparts in the target domain is not clearly examined.
  • I also have concerns about the reliability of the experimental results. For instance, in Table 1, the results of AdaGCN on C→A and D→A are exactly the same, and the results of SGC on C→A and D→C are also exactly the same. Please double check these same results.
  • Typos and presentation issues:
    • "We vary r from 0. to 0.6" in Section 5.4, while Figure 4 only shows the range up to 0.5;
    • "we remove one component at a time" in Section 5.3 without a capital letter;

[1] Huang J, Gretton A, Borgwardt K, et al. Correcting sample selection bias by unlabeled data[J]. Advances in neural information processing systems, 2006, 19.

[2] Cao Z, Ma L, Long M, et al. Partial adversarial domain adaptation[C]//European conference on computer vision. Cham: Springer International Publishing, 2018: 139-155.

[3] Liu S, Li T, Feng Y, et al. Structural re-weighting improves graph domain adaptation[C]//International conference on machine learning. PMLR, 2023: 21778-21793.

[4] Huang R, Xu J, Jiang X, et al. Can modifying data address graph domain adaptation?[C]//Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining. 2024: 1131-1142.

Questions for authors

Please see weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer B9gd

Q1. The proposed method is incremental because source selection and reweighting have already been studied in domain adaptation.

Response: SNIP addresses a distinct problem in UGDA: identifying structurally incompatible source nodes using target-domain topological information and refining the source graph before adaptation. Unlike conventional sample selection, node pruning in graphs affects not only the training samples but also graph connectivity and message passing. Furthermore, SNIP differs fundamentally from StruRW, which reweights structural relationships, and GraphAlign, which synthesizes graph data through feature and topology optimization. SNIP instead directly constructs a structurally compatible induced subgraph from the original source graph based on target-aware structural compatibility. Its novelty lies in explicitly incorporating cross-domain structural compatibility into source-node refinement, providing a lightweight, model-agnostic approach complementary to existing UGDA methods.

Q2. The claim of shifting the research paradigm from "how to align" to "what to align" is overstated.

Response: The distinction between "how to align" and "what to align" reflects a fundamental difference in how cross-domain distribution shifts are addressed. Existing UGDA methods primarily focus on improving alignment mechanisms while largely treating the entire source graph as transferable. Although reweighting methods account for heterogeneous contributions by adjusting the importance of samples or structural relationships during training, they generally retain the original source graph. In contrast, SNIP explicitly identifies and removes structurally incompatible source nodes before adaptation, thereby refining both the source-node population and its underlying topology.

This distinction is particularly important under structural shifts, where forcing incompatible source structures into alignment may introduce negative transfer. Simply adjusting their contributions during training does not directly address whether these structures should be included in the adaptation process. SNIP addresses this issue at the data level, shifting the focus from how to align existing source knowledge to what source knowledge should participate in alignment. Importantly, SNIP is complementary to existing approaches, improving adaptation without modifying their architectures or optimization objectives.

Q3. The claim that structurally deviated nodes lack semantic counterparts in the target domain is not directly verified.

Response: Our central hypothesis is that structurally incompatible source nodes may contribute negatively to cross-domain alignment, rather than that every structurally deviated node necessarily lacks a semantic counterpart. Fig. 2 provides direct empirical motivation for this hypothesis. Under matched pruning ratios, structural pruning consistently outperforms random pruning, indicating that the structural compatibility criterion helps distinguish source nodes with different contributions to adaptation. Furthermore, the improvements achieved across different backbones and transfer scenarios reinforce the practical relevance of selective source-node utilization. These results support the proposed mechanism without requiring a strict one-to-one correspondence between source and target nodes.

Q4. Identical baseline results appear across different transfer scenarios in Table 1.

Response: As explicitly stated in Section 5.2, the baseline results are adopted from previously published studies [1-2] rather than independently reproduced in this work. Therefore, identical entries should first be examined in relation to their original sources and cannot, by themselves, be taken as evidence of unreliable SNIP experiments. Importantly, the evaluation of SNIP focuses on comparisons with its corresponding backbones, AdaGCN and A2GNN, under the same transfer settings. We will verify the identified entries against the original references and correct any transcription errors to ensure accurate reporting.

[1] Rethinking propagation for unsupervised graph domain adaptation. In Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI, pp. 13963–13971, 2024.

[2] Smoothness really matters: A simple yet effective approach for unsupervised graph domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, AAAI, pp. 15875–15883, 2025.

Q5. Typographical and presentation inconsistencies.

Response: We thank the reviewer for identifying these presentation issues. We will correct the inconsistency between the pruning-ratio range described in Section 5.4 and that displayed in Fig. 4, as well as the capitalization error in Section 5.3. These corrections do not affect the methodology or experimental conclusions.

Thanks for the response, but my main concerns remain. AdaGCN is the backbone of SNIP, and the improvements in Table 1 are computed against its numbers, so the duplicated results on C→A and D→A directly affect whether the claimed gains hold. I think the backbone should be rerun rather than copied from earlier papers. The response also lacks any experimental comparison with StruRW or GraphAlign. Since no new evidence is provided, I will keep my original rating.

Reviewer_L1btOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies source-node selection for unsupervised graph domain adaptation and proposes Source Node Influence Pruning (SNIP), a model-agnostic preprocessing framework. SNIP computes degree, betweenness, closeness, and eigenvector centrality, and measures each source node’s structural discrepancy using its distance from the corresponding target-graph mean. These discrepancies are transformed into percentile ranks and combined through a weighted sum. A prescribed fraction of low-scoring source nodes is then removed, and the resulting induced subgraph is used to train an existing UGDA model. Experiments cover eight transfer directions constructed from three citation networks and two Twitch networks, using AdaGCN and A2GNN as backbones. The paper reports improvements over both backbones and provides component ablations, sensitivity analyses, and a conditional comparison of target-risk upper bounds before and after pruning.

Critical strengths

A practically relevant problem and a reusable design. Consistent reported improvements over the selected backbones.

Critical weaknesses

Insufficient justification of the selection criterion. Distance to target mean centrality does not necessarily indicate transferability, particularly for multimodal or heavy-tailed distributions. Moreover, pruning changes retained nodes’ structural properties, but selection scores are computed only on the original graph.

Table 1 also contains reporting inconsistencies: A2GNN’s displayed Macro-F1 values average to 77.005 rather than 76.72, and the claim of top-two performance on every metric is incorrect.

Limited algorithm-specific theoretical support. Proposition 1 is an algebraic comparison of two risk bounds. It does not establish that SNIP’s ranking reduces the Wasserstein discrepancy or that the resulting alignment gain exceeds the pruning cost.

Unclear model-selection protocol. The paper does not specify how hyperparameters and early stopping were selected without target labels. Given the search over pruning ratios and centrality weights, this omission prevents verification of the claimed unsupervised evaluation setting.

Comparisons against class-stratified pruning, source-only trimming, and degree-only selection are missing. Three-run means without variability or matched-run details also leave the reliability of smaller improvements unclear.

Computing global centralities offline can still incur substantial preprocessing costs. All evaluated graphs contain fewer than 10,000 nodes, and no preprocessing time, memory usage, or end-to-end runtime is reported to substantiate the claimed scalability.

Questions for authors

See the weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer L1bt - Part 1.

Q1. Mean centrality may not reliably indicate transferability, particularly for multimodal or heavy-tailed distributions.

Response: SNIP employs the distance to the target mean centrality as a practical measure of structural compatibility, rather than assuming it perfectly reflects node transferability. By integrating four complementary centrality measures through rank-based normalization, SNIP captures diverse structural characteristics instead of relying on a single metric. Although the target mean may not fully characterize highly multimodal or heavy-tailed distributions, it provides a simple and effective structural reference. We also evaluate more sophisticated target-profile estimators, which yield only marginal performance differences. Therefore, we adopt the mean-based criterion for its simplicity and effectiveness. The consistent improvements over random pruning and the component ablations further support the validity of this design choice.

We investigate four alternative strategies for estimating the target structural profile:

(1) Mean: Uses the average centrality value of target nodes as the structural reference.

(2) Median: Uses the median target centrality to reduce sensitivity to extreme values.

(3) KNN: Measures structural discrepancy using the average distance to the k nearest target nodes in the corresponding centrality dimension.

(4) KDE: Estimates the target centrality distribution using kernel density estimation and evaluates source-node compatibility based on the estimated density.

These alternatives characterize target structural patterns from different perspectives. Mean provides the simplest global reference, while Median, KNN, and KDE offer greater flexibility in capturing heterogeneous structural distributions.

ProfileA→D (Macro-F1)A→D (Micro-F1)D→A (Macro-F1)D→A (Micro-F1)A→C (Macro-F1)A→C (Micro-F1)C→A (Macro-F1)C→A (Micro-F1)C→D (Macro-F1)C→D (Micro-F1)D→C (Macro-F1)D→C (Micro-F1)
Mean78.3279.1076.7175.1182.0983.1378.3577.0578.0579.2780.1882.00
Median78.0779.0376.2774.9582.1583.3178.0376.5878.2379.3480.0682.06
KNN77.9279.2676.4175.1381.9583.2278.0176.8977.6478.8580.3382.19
KDE77.8078.9476.3375.0582.1383.3678.0677.1678.0179.2579.9781.82

Q2. Pruning changes the structural properties of retained nodes, while the scores are computed only once.

Response: Computing centralities on the original source graph is an intentional design choice. SNIP aims to assess each node's structural role in the original graph and determine its compatibility with the target domain before pruning. Recomputing centralities after node removal would characterize a modified graph rather than the original source structure and introduce additional computational overhead. Therefore, the one-shot scoring mechanism is consistent with SNIP's objective of lightweight source-domain refinement.

Q3. Table 1 contains numerical inconsistencies and an incorrect top-two performance claim.

Response: We appreciate the reviewer for identifying these reporting issues. The displayed A2GNN Macro-F1 values yield an average of 77.01 rather than 76.72, and the corresponding average improvement should be approximately 1.94 percentage points. We will correct these numerical inconsistencies and revise the top-two performance statement accordingly. Importantly, these corrections do not change the observed improvements of SNIP over its corresponding backbones across the evaluated transfer scenarios.

AuthorsOpenReview ↗

Response to Reviewer L1bt - Part 2.

Q4. Proposition 1 does not establish that SNIP's scoring rule necessarily reduces the Wasserstein discrepancy.

Response: Proposition 1 is intended to characterize the fundamental condition under which source-node pruning tightens the target-risk upper bound, rather than establish an unconditional guarantee for a particular scoring function. Specifically, the alignment gain must exceed the pruning cost. SNIP operationalizes this principle by preferentially removing source nodes with large structural discrepancies from the target domain. The theoretical analysis explains the trade-off underlying selective pruning, while the experimental results demonstrate the effectiveness of the proposed scoring strategy.

Q5. The unsupervised hyperparameter-selection and early-stopping protocols are unclear.

Response: SNIP does not require target labels for hyperparameter selection or early stopping. Specifically, the pruning ratio and centrality weights are selected based on the overall training objective (i.e., source-domain classification loss and the cross-domain alignment loss (MMD)), or alternatively set to fixed values. Similarly, early stopping is determined by the overall training objectives (i.e., loss) without accessing target labels. Since both the source classification loss and MMD loss can be computed without target labels, the entire hyperparameter-selection and training process remains consistent with the unsupervised graph domain adaptation setting.

Q6. Comparisons with class-stratified pruning, source-only trimming, and degree-only selection are missing.

Response: These setups examine different aspects of source-node selection. Class-stratified pruning primarily preserves source-class proportions, while source-only trimming removes structural outliers without considering target-domain characteristics. In contrast, SNIP explicitly evaluates source-node compatibility with the target structure, which is central to our motivation. Therefore, class-stratified pruning and source-only trimming address different objectives and are not directly aligned with our primary motivation of mitigating cross-domain structural mismatch through target-aware source-node selection.

Moreover, degree-only selection has already been investigated in Fig. 2, where degree-based structural pruning consistently outperforms random pruning. Table 3 further demonstrates the complementary contributions of multiple centrality measures. Together, these experiments provide evidence for the effectiveness of target-aware structural selection and multi-centrality combination.

Q7. Three-run averages without variability estimates make smaller improvements difficult to assess.

Response: All reported results are averaged over three runs with different random seeds, following a consistent evaluation setting. Due to the large size of the comparison tables, we report only the mean values and omit the standard deviations for clarity and readability. More importantly, SNIP improves both AdaGCN and A2GNN across all eight transfer scenarios, rather than showing isolated gains on individual datasets. This consistency across different backbones, transfer directions, and evaluation metrics provides evidence that the benefits are not restricted to a particular experimental configuration.

AuthorsOpenReview ↗

Response to Reviewer L1bt – Part 3.

Q8. Global centrality computation may be expensive, and scalability is not sufficiently demonstrated.

Response: SNIP computes structural influence scores only once during offline preprocessing and introduces no additional trainable parameters or changes to the backbone optimization procedure. Its computational cost is therefore separated from iterative model training. Moreover, SNIP is not restricted to the four selected centrality measures. Appendix E (i.e., Subsection Impact of Structural Measures) demonstrates that alternative structural measures also achieve competitive performance, allowing computationally cheaper measures to be adopted when efficiency is a priority. Thus, the framework provides flexibility in balancing structural score estimation and preprocessing cost.

Appendix E: To justify our selection rationale, we systematically compare two distinct configurations using A2GNN as the backbone. The default configuration combines degree centrality (D), betweenness centrality (B), closeness centrality (C), and eigenvector centrality (E). Conversely, the alternative configuration includes PageRank (PR), Clustering Coefficient (CC), K-Core Number (KC), and Average Neighbor Degree (AD), which primarily emphasize core-periphery structure, localized clustering or specific random-walk dynamics. Specifically, PageRank measures node importance through localized random-walk based propagation; the Clustering Coefficient reflects how tightly neighbor nodes are connected; the K-core number characterizes a node’s position in the graph’s core–periphery hierarchy; and the Average Neighbor Degree measures the mean connectivity density of a node’s immediate neighborhood.

Table 5: Comparison of different structural measure configurations across six transfer scenarios.

MethodsA→C MicroA→C MacroA→D MicroA→D MacroC→A MicroC→A MacroC→D MicroC→D MacroD→A MicroD→A MacroD→C MicroD→C MacroAvg. MicroAvg. Macro
A2GNN (Backbone)82.7580.8677.8475.1874.6376.3178.2875.9073.4474.9481.0978.8477.9577.01
Default (D, B, C, E)83.1382.0979.1078.3277.0578.3579.2778.0575.1176.7182.0080.1879.2778.95
Alternative (PR, CC, KC, AD)83.2581.8678.4877.0676.7478.1279.1777.7575.2976.8081.6579.5179.0978.52

Thank you for the clarifications. I will keep my original score.

Reviewer_nYHPOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Source Node Influence Pruning (SNIP), a data-centric refinement framework for Unsupervised Graph Domain Adaptation (UGDA). Unlike existing approaches that primarily focus on feature-level alignment, SNIP emphasizes the importance of selectively utilizing transferable source nodes. Specifically, the proposed method estimates node-level structural influence scores using multiple centrality measures, employs rank-based normalization to integrate these measures, and removes structurally incompatible source nodes before subsequent domain alignment. The framework is model-agnostic and can be integrated into existing UGDA methods. Extensive experiments on five real-world datasets across eight transfer scenarios demonstrate its effectiveness.

Critical strengths
  1. Novel perspective: The paper effectively shifts the focus from "how to align" features to "what to align" by identifying that structurally incompatible source nodes are a primary cause of negative transfer.

  2. Plug-and-Play method: SNIP is model-agnostic and operates purely at the data-processing level, making it easy to integrate with existing UGDA architectures without adding significant training time overhead.

Critical weaknesses
  1. The paper investigates the effects of the pruning ratio and other hyperparameters. A brief discussion of the recommended default settings would be helpful for practitioners interested in applying SNIP to new datasets.

  2. The authors employ four centrality measures to characterize node-level structural compatibility. It would be helpful to briefly explain the motivation for selecting these particular measures in the main text, although additional comparisons are already provided in the appendix.

Questions for authors
  1. The authors need to elaborate on how the pruning ratio and centrality weights are selected in the unsupervised setting, particularly without relying on target-domain labels.
2 comments
AuthorsOpenReview ↗

Response to Reviewer nYHP

Q1. How are the pruning ratio and centrality weights selected without target labels?

Response: SNIP does not require target labels for structural compatibility estimation or source-node pruning. The pruning ratio controls the trade-off between removing structurally incompatible nodes and preserving source information, while the centrality weights balance different structural measures. These hyperparameters can be selected using grid search strategy or fixed configurations without accessing target labels. Moreover, our sensitivity analyses in Fig. 4 and Fig. 6 show that SNIP remains effective across a range of parameter settings, indicating that its effectiveness does not depend on a single carefully tuned configuration.

Q2. What default hyperparameter settings are recommended?

Response: Fig. 4 shows that moderate pruning is generally more effective than aggressive pruning. A pruning ratio of approximately 0.2 is therefore a reasonable initial choice, while uniform centrality weights (0.25 for each measure) provide a simple configuration without introducing metric-specific preferences. These settings provide a practical balance between structural refinement and information preservation. The optimal pruning ratio and centrality weights may depend on the degree of structural shift between domains.

Q3. Why are these four centrality measures selected?

Response: The four measures capture complementary structural properties: degree characterizes local connectivity, betweenness reflects bridging roles, closeness describes global accessibility, and eigenvector centrality measures connectivity to influential neighbors. Their combination provides a broader characterization of node structure than any single metric. Table 3 supports this design: removing degree, betweenness, closeness, or eigenvector centrality decreases average Macro-F1 by 1.01, 0.51, 0.88, and 1.39 percentage points, respectively. Therefore, the four measures contribute complementary information rather than serving as redundant features.

Thank you for your detailed responses to my questions. I will keep my positive score.

Reviewer_gc4qOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Source Node Influence Pruning (SNIP), which removes source nodes based on differences between their centralities and the global mean centrality values of the target graph, using rank-based normalization and weighted node influence scores. The retained induced subgraph is then used for unsupervised graph domain adaptation. The method is applied as a preprocessing step for A2GNN and evaluated on eight transfer scenarios across five datasets.

Critical strengths

The method is simple and can be integrated into existing graph domain adaptation models. It reports improvements in mean performance across two backbones and eight transfer scenarios, suggesting practical value. Comparisons with random pruning, pruning-ratio analysis, and component ablations provide useful additional evidence.

Critical weaknesses
  1. The relationship between node influence scores and transferability is not sufficiently validated. A larger deviation from the target graph’s mean centrality does not necessarily imply that a node is harmful to transfer, particularly when the target graph contains different structural patterns or classes with few nodes. The experiments show the overall benefits and costs of pruning, but do not sufficiently establish that the scores distinguish nodes that should be removed from those that should be retained. They also do not examine how removing informative nodes affects individual target classes.

  2. The theory does not establish a connection between the specific scoring rule and improvement in the target-risk upper bound. Proposition 1 correctly accounts for the alignment gain and pruning cost, but its proof is primarily an algebraic decomposition of the upper bounds before and after pruning, without using the centrality scores or ranking rule.

Questions for authors
  1. Could the authors analyze the class distributions and structural properties of the pruned and retained nodes, together with performance changes for individual target classes, to explain how the scores avoid removing nodes that remain useful for transfer?

  2. Under what conditions can scoring based on the target graph’s mean centrality reduce the source-target distribution discrepancy and make the alignment gain exceed the pruning cost?

2 comments
AuthorsOpenReview ↗

Response to Reviewer gc4q

Q1. The relationship between node influence scores and transferability is insufficiently validated.

Response: SNIP does not assume that structural discrepancy perfectly determines node transferability. Instead, it uses target-aware structural compatibility as a proxy for identifying potentially harmful source nodes. Fig. 2(b–c) shows that structural pruning outperforms random pruning, while Tables 1–2 demonstrate consistent improvements across two UGDA backbones. These results support the effectiveness of the selection criterion, although they do not establish a causal relationship at the individual-node level.

Q2. How does pruning affect class distributions and individual target classes?

Response: SNIP is designed to reduce structural mismatch rather than explicitly preserve class distributions. Since the pruning criterion does not use source labels, it may alter class proportions, particularly for structurally heterogeneous classes. Nevertheless, both Micro-F1 and Macro-F1 improve consistently in the reported experiments. Since Macro-F1 assigns equal importance to each class, these improvements suggest that the overall benefits are not limited to majority-class performance but also extend to minority classes.

Q3. The theoretical analysis does not directly justify the proposed scoring rule.

Response: Proposition 1 is intended to characterize the conditions under which source-node pruning improves the target-risk upper bound, rather than to prove that the proposed scoring rule guarantees such improvement. Specifically, pruning is beneficial when the alignment gain exceeds the combined costs of source information loss and increased estimation uncertainty. SNIP provides a practical structural compatibility heuristic motivated by this trade-off, rather than a theoretically optimal selection criterion. We will clarify this distinction to avoid overstating the theoretical guarantee.

Q4. Under what conditions is target-mean-based pruning effective?

Response: Mean-based pruning is most appropriate when the target structural distribution has a representative central tendency and structurally mismatched source nodes lie far from it. The mean-based criterion is a simple implementation of node influence score estimation. We also evaluate more sophisticated alternatives, which yield only marginal performance differences. Therefore, we adopt the mean-based criterion for its simplicity and effectiveness.

We investigate four alternative strategies for estimating the target structural profile:

(1) Mean: Uses the average centrality value of target nodes as the structural reference.

(2) Median: Uses the median target centrality to reduce sensitivity to extreme values.

(3) KNN: Measures structural discrepancy using the average distance to the k nearest target nodes in the corresponding centrality dimension.

(4) KDE: Estimates the target centrality distribution using kernel density estimation and evaluates source-node compatibility based on the estimated density.

These alternatives characterize target structural patterns from different perspectives. Mean provides the simplest global reference, while Median, KNN, and KDE offer greater flexibility in capturing heterogeneous structural distributions.

ProfileA→D (Macro-F1)A→D (Micro-F1)D→A (Macro-F1)D→A (Micro-F1)A→C (Macro-F1)A→C (Micro-F1)C→A (Macro-F1)C→A (Micro-F1)C→D (Macro-F1)C→D (Micro-F1)D→C (Macro-F1)D→C (Micro-F1)
Mean78.3279.1076.7175.1182.0983.1378.3577.0578.0579.2780.1882.00
Median78.0779.0376.2774.9582.1583.3178.0376.5878.2379.3480.0682.06
KNN77.9279.2676.4175.1381.9583.2278.0176.8977.6478.8580.3382.19
KDE77.8078.9476.3375.0582.1383.3678.0677.1678.0179.2579.9781.82

Thank you for the detailed response and the additional experiments. I will raise my rating.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.