acceptodds
Under review as a conference paper at ICLR 2027

GLEX: Executable Program Vocabularies for Text-Free Graph Foundation Models

Abstract

Few-shot transfer across heterogeneous graphs is challenging in the text-free setting, where source and target domains may differ in feature semantics, structural statistics, and label spaces. Existing transfer interfaces are often built on representations, aligned coordinates, or structural patterns whose usefulness can depend on the graph on which they are formed. We introduce GLEX (Graph Lexicon of Executable Programs), which instead builds a transferable vocabulary of reusable graph computations. The same computation can be instantiated differently on each graph, providing a common interface while naturally adapting to heterogeneous features and structures. On an unseen target, the pretrained encoder remains fixed, and a few labeled examples determine which computations are most useful for the target task and refine the resulting predictions. Extensive experiments on seven heterogeneous target graphs cover both node and graph classification from one-shot to ten-shot settings. GLEX improves average one-shot node-classification accuracy from 46.52% to 52.95% over the strongest text-free multi-domain baseline BRIDGE, while also improving average graph-classification accuracy from 51.97% to 53.54%. Ablation studies identify the executable vocabulary and support-conditioned adaptation as the main contributors, while efficiency, sensitivity, and program-use analyses further characterize how GLEX behaves across target graphs.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph foundation models, few-shot graph learning, cross-domain graph transfer

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_2F16OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes GLEX, a text‑free graph foundation model for few‑shot cross‑graph transfer. Instead of transferring fixed representations, it leverages executable graph programs that adapt to heterogeneous graphs. GLEX induces graph‑specific program vocabularies, uses a frozen pre‑trained encoder, and conducts closed‑form target adaptation without fine‑tuning. Evaluations on seven datasets yield strong one‑shot node‑classification gains but limited graph‑classification improvements, with key limitations including hand‑crafted program primitives and lack of large‑graph validation.

Critical strengths
  1. Novel transfer paradigm of executable graph computations: Instead of transferring representations or structural instances, GLEX transfers computation logic. It generates graph‑adaptive outputs without explicit feature alignment. The pretrained encoder is frozen; closed‑form adaptation avoids fine‑tuning and scales with support‑set size for stable few‑shot performance.
  2. Strong few‑shot node‑level results and solid evaluations: GLEX outperforms the strong baseline BRIDGE by 6.43% on average for one‑shot node classification over seven heterogeneous datasets. It includes thorough ablations, diagnostics and theoretical analysis to back up empirical observations.
Critical weaknesses
  1. Strong inductive bias from hand‑designed program primitives: Candidate programs are restricted to a manually‑defined pool. Greedy selection only picks existing operators and cannot generate new computation primitives, risking failure for tasks requiring out‑of‑distribution graph operations.
  2. Unclear practical utility of greedy per‑graph induction & limited generality: Ablations show a fixed program bank works nearly as well on average, so greedy induction helps only for specific datasets. Gains are marginal for graph‑level classification, and real‑world large‑graph experiments are missing.
Questions for authors
  1. The greedy per‑graph vocabulary induction brings only minor average improvement compared with a fixed program bank. Could you identify what graph properties determine whether this greedy induction yields large empirical gains?
  2. Why does GLEX obtain substantial improvements for node classification but only marginal gains for graph‑level classification? Is this caused by the program vocabulary itself or by the pooling design for ego‑graphs?
  3. All program primitives are manually predefined. What failure modes can we expect when target tasks require graph computations outside this hand‑built candidate pool?
  4. Vocabulary induction involves heavy computation. How does it scale to real‑world large graphs, and can sampling fully mitigate the overhead?
3 comments

Thank you for recognizing the transfer interface and the node-level results. Your concerns are addressed below.

W1 & Q3:

Thank you for raising this important question about the scope and limitations of the predefined grammar. We agree that greedy selection can only choose and compose programs from the predefined grammar; it cannot invent a new primitive at transfer time. This is an intentional inductive bias that keeps every selected program explicit and executable on an unseen graph. The current grammar nevertheless covers a broad family of computations commonly used in graph representation learning, including multi-hop propagation, contrasts between propagation depths, nonlinear scalar transformations, and pointwise feature interactions. Their bounded compositions express multiscale smoothing, high-pass responses, and feature-dependent interactions. The consistent results across citation, co-purchase, social, and whole-graph benchmarks suggest that this scope covers diverse, though not arbitrary, target computations.

Across the diverse benchmarks evaluated in this work, the current grammar provides sufficiently informative programs, and we have not identified a concrete failure case for the complete GLEX pipeline. In principle, a limitation would arise only if two nodes or graphs produced identical outputs under every candidate program while the task required different predictions for them. Selecting among the existing programs could not separate such inputs. This describes a general theoretical boundary of any fixed candidate pool. The framework is also extensible: if a future task reveals a missing computation, the corresponding primitive can be added without changing the vocabulary-selection or adaptation procedure.

W2 & Q1:

Thank you for valuable comment. To examine the practical value of induction, we compare GLEX with four predefined ten-program banks using a fixed SVD rank of 32 and the same one-shot support/query splits. GLEX selects ten programs from unlabeled target-graph executions. Each predefined bank also contains ten programs and shares standardized signals at depths zero, one, and two, while emphasizing a different computation family: low-order multiscale transforms, higher-order transforms, contrasts between propagation depths, or transported nonlinear signals. The original paper reports a compact four-program reference bank; this additional control complements that ablation by comparing vocabulary choices under the same ten-program budget.

TargetGLEXLow-order multiscaleHigher orderContrastTransport
Cora48.5051.2548.9447.6951.59
CiteSeer35.3340.1338.7136.9835.42
PubMed58.6056.0054.1151.4953.38
Photo73.3171.4968.2063.9471.53
Computers56.9360.9658.7556.2061.02
Facebook47.3541.9141.2139.6642.17
LastFM41.9741.0343.9248.9840.58
Average51.7151.8250.5549.2850.81

No predefined bank is uniformly best across targets. For example, the contrast bank outperforms the transport bank by 8.40 points on LastFM but trails it by 7.59 points on Photo. GLEX achieves 51.71% on average, within 0.11 points of the strongest fixed bank, without target-specific hand selection. Its practical value lies in providing a single label-free rule for choosing a competitive vocabulary on an unseen graph. Manually defining such a bank is difficult: selecting ten programs from the candidate pool creates a large combinatorial space, and the best combination cannot be known before seeing the target. The few labeled examples available in the one-shot setting also provide too little evidence for a reliable target-specific search. We do not observe a simple graph property that predicts the gain. Even within the citation domain, GLEX leads all four fixed banks on PubMed but not on Cora or CiteSeer, showing that graph family alone is insufficient. A reliable predictor would require a substantially larger collection of target graphs; GLEX instead makes the choice directly from unlabeled program executions. We have added the corresponding discussion to the revised appendix.

W2 & Q2:

Thank you for highlighting the difference between node- and graph-level gains. The smaller graph margin is specific to the comparison with the strong BRIDGE result, rather than a general limitation of GLEX on graph-level tasks. In the main results, GLEX improves over BRIDGE, SAMGPT, and GCOPE by 6.43, 7.60, and 11.90 points for node classification, and by 1.57, 4.65, and 10.64 points for graph classification, respectively, showing that the graph-level gains remain substantial against the other strong baselines.

To further evaluate transfer when each entire graph has its own ground-truth label, we add experiments on the whole-graph datasets PROTEINS, ENZYMES, and NCI1, using the original seven graphs as source domains. All methods use the same settings. GLEX exceeds the strongest baseline BRIDGE by 20.61, 35.32, and 13.66 points on the three datasets, respectively. These results show that the premise of consistently marginal gains on graph-level tasks does not hold: GLEX’s small advantage over BRIDGE is specific to the original ego-graph comparison, rather than a general limitation caused by the program vocabulary or pooling design. We have added this experiment to the revised paper.

MethodPROTEINSENZYMESNCI1
GCN50.4519.0749.98
GAT49.8619.2451.09
BRIDGE52.7322.2950.98
GLEX73.3457.6164.75

W2 & Q4:

Thank you for raising the question of computational scalability and the effect of sampling.

(1) Effectiveness on a large graph. To evaluate whether GLEX remains effective at a substantially larger scale, we further evaluate it on ogbn-arxiv, which contains 169,343 nodes, 1,166,243 directed edges, and 40 classes. We apply SVD to its official 128-dimensional node attributes and use the resulting 16-dimensional inputs for all methods. GLEX achieves 22.43% accuracy, compared with 19.03% for BRIDGE and 15.53% for GraphPrompt. These results confirm that GLEX remains effective on a large attributed graph.

(2) Effectiveness of sampling. Under the default grammar configuration, vocabulary induction first executes 150 candidate programs and then uses their execution values to select the final ten. Each candidate produces a node-by-feature matrix with 169,343 rows (one per node) and 16 columns (one per SVD feature), amounting to 169,343 × 16 = 2,709,488 scalar values per candidate. Without sampling, the selection stage uses all of these values from every candidate. With sampling, it instead retains 512 evenly spaced values per candidate, for a total of 76,800 values across the 150 candidates, and uses them to select the final ten programs. This reduces selection time from 408.76 s to 2.42 s and total induction time from 415.41 s to 9.13 s, while lowering peak process RAM from 9.13 to 3.68 GiB. Candidate generation still takes 6.70 s because every candidate must be executed on the full graph. Sampling therefore removes most of the selection overhead.

The measurements use one fresh CPU process per condition on a Xeon Platinum 8375C with two numerical-library threads. Generation and selection times exclude graph loading and SVD, while peak host RAM includes graph loading and runtime libraries. The two runs use the same rank-16 attribute inputs and differ only in the sampling cap. We have added these results and the corresponding analysis to the revised paper.

SettingScalar values used per candidateGeneration (s)Selection (s)Total (s)RAM (GiB)
Sampling5126.702.429.133.68
No sampling2,709,4886.65408.76415.419.13

I appreciate the authors' additional experiments, which address several of my concerns, particularly regarding graph-level performance and scalability. However, the new comparisons with fixed program banks still leave the practical benefit of graph-specific vocabulary induction insufficiently established. Moreover, the reliance on hand-crafted program primitives remains a limitation. I therefore maintain my original score.

Reviewer_5RuvOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a framework, called GLEX, for transferable graph representation learning. Unlike many graph foundation models that utilize textual embeddings, structural patterns, etc., as their vocabularies, GLEX builds vocabularies from executable programs. For downstream tasks, the pretrained encoder is frozen, however, the adaptation to specific graph is handled naturally through vocabulary construction and selection. The authors also provide both theoretical analysis and conduct one-shot and few-shot experiments to further support the effectiveness of GLEX.

Critical strengths
  • S1: The intuition of using reusable computations as vocabulary is reasonable since it naturally supports adaptation to graphs with various feature dimensions and from different domains without relying on the semantics.
  • S2: The research questions are well selected to cover several essential aspects of the proposed method and their presentations are clear.
Critical weaknesses
  • W1: Missing references. More discussions on prior works such as GVT [1] and ProGFM [2] can be useful. Similar ideas like learning from propagations and their transformations are shown in some of these prior works. They also should be added to the few-shot experiments.
  • W2: Graph classification experiments are designed only on local ego-graphs rather than the whole graph, which is less convincing when claiming graph-level transferability.

[1] View Space: Learning Representation across Arbitrary Graphs, 2026

[2] Beyond Feature and Structure Alignment: Learning Transferable Propagation Knowledge for Graph Foundation Models, 2026

Questions for authors
  • Q1: Many datasets in this paper have textual features, for example, bag of words or even rich raw texts that can be embedded via LLMs, can you compare GLEX with LM-based GFMs? You may use an embedding model to generate node features from raw texts then feed them to GLEX.
  • Q2: Or alternatively, can you add an additional experiment on large-scale graphs with only numerical features or even no features? Such a task is out of the scope of text attributed graph-based methods and can further establish the usefulness of GLEX.
3 comments

Thank you for the positive assessment of reusable computations and the research questions. Your concerns are addressed below.

W1:

Thank you for mentioning these closely related methods. We have added the following comparison to Related Work in the revised paper: “GVT aligns propagated feature views, while ProGFM transfers propagation prototypes (Lee et al., 2026; Wang et al., 2026). In contrast, GLEX selects graph-specific executable programs from a shared vocabulary and adapts them with support labels.”

We have also added Recurrent GVT to the main one-shot node-classification table. We trained Recurrent GVT using the same experimental protocol as the other methods. The results are listed below. GLEX achieves an average accuracy of 52.95%, compared with 39.74% for Recurrent GVT, an improvement of 13.21 percentage points. GVT is formulated and officially implemented for node classification and does not provide a graph-classification protocol. We therefore evaluate it only in its supported setting rather than introduce an unvalidated graph-level extension. We could not locate a public implementation of ProGFM and therefore do not report results from an unverified reimplementation.

Target datasetGLEX (%)Recurrent GVT (%)
Cora48.7834.22
CiteSeer38.9332.73
PubMed59.2045.80
Amazon Photo73.3859.93
Amazon Computers57.6751.80
Facebook45.3832.40
LastFM47.2921.32
Average52.9539.74

W2:

We use local ego-graphs in the original experiments because the seven benchmarks provide labels only for nodes, not for entire graphs. Each ego-graph therefore inherits the label of its center node, allowing us to evaluate graph-level prediction using the supervision available in these datasets.

To directly evaluate transfer to whole graphs, we additionally evaluated GLEX on three datasets with whole-graph labels: PROTEINS, ENZYMES, and NCI1. PROTEINS and ENZYMES are binary and six-class protein-classification datasets, respectively, while NCI1 is a binary molecular anticancer-activity dataset. The original seven datasets served as the source domains, and all methods were evaluated under the same settings. As shown below, GLEX achieved the highest accuracy on all three datasets: 73.34% on PROTEINS, 57.61% on ENZYMES, and 64.75% on NCI1, outperforming the strongest baseline by 20.61, 35.32, and 13.66 percentage points, respectively. These results show that GLEX transfers effectively to whole-graph classification, beyond the local ego-graph setting. We have added this experimental setting and the corresponding results to the revised manuscript.

MethodPROTEINSENZYMESNCI1
GCN50.4519.0749.98
GAT49.8619.2451.09
BRIDGE52.7322.2950.98
GLEX73.3457.6164.75

Q1:

To test whether GLEX can directly use semantic text features, we first replace the original node inputs with frozen text embeddings. We then compare GLEX with GraphText and LLaGA, two language-model-based graph methods, while also reporting representative graph baselines under the same embedding-based protocol.

We use Cora, CiteSeer, PubMed, and ogbn-arxiv because their nodes represent papers with associated text. Each paper is encoded once with frozen Qwen3-Embedding-8B using its title and abstract. Cora, CiteSeer, and PubMed use the original 4,096-dimensional embeddings. Because ogbn-arxiv contains 169,343 nodes, we apply SVD to reduce the inputs of GLEX and the standard graph baselines to 256 dimensions, substantially reducing memory and computation while retaining the dominant variation in the embeddings. All methods are evaluated with one labeled node per class on the same five one-shot splits.

As shown below, GLEX consistently outperforms both language-model-based baselines on all four datasets. For example, it improves over GraphText from 49.73% to 60.41% on Cora and from 11.47% to 27.16% on ogbn-arxiv. LLaGA’s weak performance likely reflects a mismatch between its pretrained graph–language alignment and our cross-domain one-shot setting, where one label per class is insufficient to adapt that alignment. These results show that GLEX can effectively operate on semantic embeddings without requiring a language model to be trained or fine-tuned within the graph-learning pipeline. We have added a dedicated subsection to the revised manuscript describing the experimental setup, comparisons, and results with language-model-derived text features.

DatasetGLEXGCNGCOPEBRIDGEGraphTextLLaGA
Cora60.4155.9253.2758.4849.739.93
CiteSeer43.6042.8042.9344.9334.8010.73
PubMed57.9355.1341.7353.6751.0022.20
ogbn-arxiv27.1610.5317.6625.6811.472.33

Q2:

Following your suggestion, we evaluate GLEX on ogbn-arxiv, a substantially larger graph, without using any node attributes. Ogbn-arxiv contains 169,343 papers, 1,166,243 citation links, and 40 subject classes. We add it to the seven original targets and, for every graph, use the top 16 singular vectors of the undirected adjacency matrix as structure-derived node inputs. GLEX ranks first on all eight graphs, achieving 45.96% mean accuracy, compared with 33.74% for GraphPrompt and 33.68% for BRIDGE. On ogbn-arxiv, GLEX reaches 18.69%, versus 6.43% for GraphPrompt and 5.64% for BRIDGE. These results show that GLEX scales to a large graph and remains effective without text or numerical node attributes. We have added this experiment and the corresponding discussion to RQ1 of the revised paper.

MethodCoraCiteSeerPubMedAmazon PhotoAmazon ComputersFacebookLastFMogbn-arxivAverage
GCN31.1621.3841.5338.7823.0935.7326.844.3127.85
GAT30.2923.7838.5632.6418.5133.4923.363.7025.54
DGI27.4418.8736.9340.3224.3132.6724.263.7326.07
GraphPrompt31.7022.5341.0952.3937.6434.7343.436.4333.74
BRIDGE34.8522.9838.7851.2441.6732.7841.525.6433.68
GLEX44.2928.9657.1160.2555.8940.5861.9118.6945.96

Thank you for the detailed responses. The additional experiments have addressed many of my concerns, especially the whole-graph generalization and scaling up without original node attributes. I will remain my original positive assessment.

Reviewer_JQr8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GLEX uses executable graph transformations as a transfer interface across graphs with heterogeneous features. It selects graph-specific program vocabularies from shared construction rules, pretrains a shared encoder on source graphs, and adapts to a target through support-conditioned predictor weighting and residual correction while keeping the encoder frozen. Evaluation covers few-shot node classification and classification of node-centered ego-graphs on seven target networks.

Critical strengths
  1. Defining transferable units as graph computations is a useful design choice. Propagation, differences, and nonlinear interactions can be instantiated on graphs with different feature spaces, while the target support set selects their relevance. The closed-form predictor weighting and residual fit also make adaptation possible without updating the encoder.

  2. The node-classification results show a substantial system-level gain under the reported protocol. Average one-shot accuracy increases from 46.52% for BRIDGE to 52.95%, with the highest mean on six of seven targets (Table 1). Excluding the target graph from source pretraining and evaluating across several network types give this result more weight than a within-dataset comparison.

Critical weaknesses
  1. The target-label budget is substantially larger than the headline shot count. Appendix F uses 30 labeled validation examples per class in addition to the support set, selecting the SVD input rank and, for graph classification, the ego radius and readout. These labels do not enter the predictor fit, but they still guide target-specific decisions. The results therefore establish one-shot fitting with substantial additional labeled validation, rather than adaptation from one labeled example per class in total. It is also important to establish comparable validation access for the borrowed and rerun baselines.

  2. The benefit of graph-specific vocabulary induction is not well separated from using multiple feature transformations. A fixed four-program vocabulary already reaches 52.05% node accuracy, only 0.90 points below the induced vocabulary with a budget of ten (Section 4.3 and Table 5). The difference changes vocabulary size as well as selection, and no unpretrained-encoder control is reported to isolate the value of source pretraining. Thus, the full-system gains are stronger evidence for combining program views and support adaptation than for the specific vocabulary-induction and transfer claims.

Questions for authors
  1. What performance remains when input rank and graph readout are selected using source data only, with no extra target validation labels? Under the current protocol, do all baselines receive the same validation-label budget and comparable configuration selection?

  2. At the same ten-program budget and with the same support adaptation, how much does graph-specific selection improve over a fixed or random vocabulary? How much of the resulting gain remains if the source-pretrained encoder is replaced by an untrained encoder of the same architecture?

2 comments

Thank you for recognizing the executable-computation interface and the node-classification results. Your concerns are addressed below.

W1 & Q1:

Thank you for your valuable comment. We carefully checked the model-selection protocols used by the baselines. BRIDGE explicitly reports using validation data for model selection, and its released preprocessing code defines a validation split of 500 nodes, substantially more validation labels than we used. For baseline SAMGPT, it also evaluates its hyperparameters on the target graphs Cora, Photo, and Facebook before fixing them to the empirically robust values. Thus, the reported baselines likewise rely on target-domain model selection rather than on the labeled support set alone. Accordingly, the comparison remains fair, because GLEX does not use more labeled target data for model selection than the baselines.

Nevertheless, we agree that the term “one-shot” can be misleading when labels beyond the one support example per class are used for validation. We therefore added one-shot node- and graph-classification experiments that use no target validation labels. In both settings, we fix the SVD rank at 32. For graph classification, we additionally fix mean pooling and use a resource-bounded two-hop neighborhood with at most 10 first-hop neighbors and 4 second-hop neighbors per first-hop node. Removing target validation reduces average accuracy only modestly. Average node accuracy decreases by 1.24 percentage points, from 52.95% to 51.71%, and average graph accuracy decreases by 0.47 points, from 53.54% to 53.07%. Both fixed-setting averages remain above the reported baselines. Thus, target-validation-based rank selection does not change the overall performance conclusion. We have added these fixed-setting results to the corresponding tables in the revised manuscript.

One-shot node-classification accuracy (%).

Method / protocolCoraCiteSeerPubMedPhotoComputersFacebookLastFMAverage
GLEX (target-val selected)48.7838.9359.2073.3857.6745.3847.2952.95
GLEX (fixed; no target val.)48.5035.3358.6073.3156.9347.3541.9751.71
BRIDGE35.9944.3355.4063.6854.1343.0729.0446.52
SAMGPT47.8036.3850.2558.7148.2242.7033.3645.35

One-shot graph-classification accuracy (%).

Method / protocolCoraCiteSeerPubMedPhotoComputersFacebookLastFMAverage
GLEX (target-val selected)51.0943.0761.6773.3157.9349.2038.5453.54
GLEX (fixed; no target val.)44.2939.8755.4774.8659.6043.2754.1253.07
BRIDGE41.4346.2056.0068.8555.3343.9352.0651.97
SAMGPT55.3538.7548.6958.7548.7243.7148.2848.89

W2 & Q2: Vocabulary selection and source pretraining

Thank you for pointing out! We agree that the effects of program selection and source pretraining should be evaluated separately. We provide two matched controls below.

2.1 Influence of program selection

To evaluate the influence of program selection, we compare GLEX with four predefined ten-program banks under a fixed SVD rank of 32 and the same one-shot support/query splits. GLEX selects its bank from unlabeled program executions. All four predefined banks share standardized signals at depths zero, one, and two, but fill the remaining slots differently: the low-order multiscale bank applies low-order nonlinear transforms to signals at multiple propagation depths; the higher-order bank emphasizes higher-degree transforms; the contrast bank uses differences between propagation depths; and the transport bank further propagates nonlinear base signals. The best predefined bank averages 51.82%, only 0.11 percentage points higher than performance of auto-selected programs.

TargetGLEXLow-order multiscaleHigher orderContrastTransport
Cora48.5051.2548.9447.6951.59
CiteSeer35.3340.1338.7136.9835.42
PubMed58.6056.0054.1151.4953.38
Photo73.3171.4968.2063.9471.53
Computers56.9360.9658.7556.2061.02
Facebook47.3541.9141.2139.6642.17
LastFM41.9741.0343.9248.9840.58
Average51.7151.8250.5549.2850.81

The results reveal two complementary aspects of program selection. (1) Effectiveness. The predefined combinations vary substantially across targets: the contrast bank outperforms the transport bank by 8.40 points on LastFM but underperforms it by 7.59 points on Photo. Despite using no target-specific hand selection, GLEX achieves 51.71% on average, only 0.11 points below the best predefined bank, demonstrating competitive overall performance across heterogeneous targets. (2) More importantly, however, automatic selection is a methodological necessity rather than a mechanism introduced solely to maximize accuracy. The best program combination cannot be specified in advance for an unseen graph. The candidate combinations form a large search space, while the few labeled support examples provide too little evidence to choose among them reliably. Thus, the main purpose of automatic selection is not merely to exceed every predefined bank, but to make a principled choice from the unlabeled target graph when target-specific hand tuning is unavailable.

2.2 Influence of source pretraining

This controlled ablation concerns one-shot node classification. To isolate the effect of source pretraining, we use the same fixed-rank, no-target-validation GLEX configuration reported in our response to Question 1 and change only the encoder weights: the frozen source-pretrained encoder is replaced by a frozen random initialization of the same architecture.

EncoderCoraCiteSeerPubMedPhotoComputersFacebookLastFMAverage
Source-pretrained48.5035.3358.6073.3156.9347.3541.9751.71
Random initialization45.3134.0951.8461.8753.9141.6649.8448.36

Source pretraining improves the average accuracy from 48.36% to 51.71%, a gain of 3.35 percentage points, and improves six of the seven target datasets, with the largest gains on Photo (+11.44 points) and PubMed (+6.76 points). LastFM is the only exception: its mean decreases from 49.84% with random initialization to 41.97% after source pretraining. This target-specific decrease is consistent with negative transfer rather than a uniform benefit from pretraining. Overall, the six positive target-level changes and the 3.35-point average gain demonstrate the benefit of source pretraining.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.