acceptodds
Under review as a conference paper at ICLR 2027

TFMLinker: Multi-Domain Link Predictor by Graph In-Context Learning with Tabular Foundation Models

Abstract

Link prediction is a fundamental task in graph machine learning with widespread applications such as recommendation systems, drug discovery, knowledge graphs, etc. In the foundation model era, how to develop multi-domain link prediction methods across datasets becomes a key problem, with some initial attempts adopting Graph Foundation Models (GFMs) utilizing Graph Neural Networks (GNNs) and Large Language Models (LLMs). However, the existing methods face notable limitations, including limited pre-training scale or heavy reliance on textual information. Motivated by the success of Tabular Foundation Models (TFMs) in achieving multi-domain prediction across diverse tabular datasets, we explore an alternative approach by TFMs, which are pre-trained on diverse synthetic datasets sampled from structural causal models and support strong in-context learning independent of textual attributes. Nevertheless, adapting TFMs for link prediction faces severe technical challenges such as how to obtain the necessary context and capture link-centric topological information. To solve these challenges, we propose TFMLinker (Tabular Foundation Model for Link Predictor), aiming to leverage the in-context learning capabilities of TFMs to perform link prediction across diverse graphs without requiring dataset-specific fine-tuning. Specifically, we first develop a prototype-augmented local-global context module to construct context that captures both graph-specific and cross-graph transferable patterns. Next, we design a multi-domain topology-aware link encoder to capture link-centric topological information and generate link representations as inputs for the TFM. Finally, we employ the TFM to predict link existence through in-context learning. Experiments on 7 graph benchmarks spanning diverse domains show that TFMLinker achieves the best average Hits@50 and ranks first on five datasets, without dataset-specific gradient-based fine-tuning.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Foundation Model, Tabular Foundation Model, Link Prediction, Graph Machine Learning

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

This paper focuses on adapting tabular foundation models (TFMs) for use for link prediction. This is due to the success of TFMs for node classification and regression tasks. However, existing models aren't adapted to the task of link prediction. As such, the authors of this work introduce TFMLinker, a TFM specifically designed for link prediction. It works by sampling K local samples and learning K' global prototypes for use in ICL. For a new query (i.e., link to be predicted), the query is encoded via DRNL along with a few local heuristics (the same is done for the K local samples). Then, the query, local, and global context are then rearranged as a table for use by the TFM which is trained to correctly predict the query link given the IC samples. The authors test the generalization capbilities of their method on multiple datasets showing some improvement over baselines.

Critical strengths
  1. The motivation of this paper is good as existing GFMs either don't work for link prediction or heavily rely on LLMs (or textual features).

  2. The paper is well written and very easy to follow.

  3. The results are generally good.

Critical weaknesses
  1. There is little novelty compared to UniLP. Both frameworks: (a) sample K positive and negative pairs; (b) encode each via DRNL; (c) perform ICL. The big difference in the inclusion of a: (a) global context; (b) heuristics; (c) TFM. However, as noted in Weakness 4, the global context seems to be limited in impact. As such, the real difference lies in the use of a TFM and heuristics which is really not much of a difference and is more of a design choice in my opinion.

  2. The authors don't consider the common datasets used for link prediction. Specifically, the common datasets used for link prediction are the OGB datasets [3], including ogbl-collab, ogbl-ddi, ogbl-ppa, and ogbl-citation2. However, none of these are used for the main experiments (See Q4 for a question on the ogbl-citation2 results in the appendix). I find this odd, as these datasets are the most realistic and accepted datasets for link prediction. To give an example why this matters, the train/valid/test splits for the datasets used in this study are randomly split (as per UniLP). However, this is clearly not realistic in most settings, where there is a clear pattern to different splits. E.g., the links in the ogbl-ppa datasets are biologicaly split. Furthermore, the datasets used are very small (the biggest evaluation dataset has ~22K nodes per Table 4). Real-world datasets are typically large (e.g., ogbl-ppa has ~576K nodes). As such, it's unclear to me how realistic the current setting is, raising some concerns on my end.

  3. Multiple strong baselines are not included in the GNN-based section. This includes [1] (which the authors cite) and [2]. I recommend including them to get the most accurate assessment.

  4. The ablation study seems to contradict the inclusion of a design settings. This includes the (a) global context, (b) GNN pre-train, (c) heuristics. Per Table 2, removing these components result in no statistically significant difference in performance, once accounting for the variance. The authors should seriously considers thew consequences of this, especially in regards to the performance w/o the global context which is a key part of their framework.

[1] "Lpformer: An adaptive graph transformer for link prediction." KDD 2024.
[2] "OCN: Effectively Utilizing Higher-Order Common Neighbors for Better Link Prediction." NeurIPS 2025.
[3] "Open graph benchmark: Datasets for machine learning on graphs." NeurIPS 2020.

Questions for authors
  1. Can you please clarify the novelty of the work? Specifically compared to UniLP? (Weakness 1)
  2. Would it be possible to test on one of the OGB datasets (Weakness 2)? For reference, ogbl-collab is the smallest.
  3. Could you include other baselines methods? (Weakness 3)
  4. Can you explain why the results for other models are so low for ogbl-citation2 in Table 5? For example the authors report a performance of 24.79 for NCNC which contradicts both the original paper [1] and subsequent replication [2].

[1] "Neural Common Neighbor with Completion for Link Prediction." ICLR 2024.
[2] "Evaluating Graph Neural Networks for Link Prediction: Current Pitfalls and New Benchmarking." NeurIPS 2023.

5 comments

Thanks for your constructive comments and here are detailed responses.

W1 & Q1. Novelty vs UniLP.

Thanks. We would like to clarify that, compared to UniLP, our novelty lies a new multi-domain LP formulation centered on TFM. TFMs are pretrained on diverse synthetic datasets sampled from structural causal models and make predictions by in-context learning. They have already shown strong performance on node classification [2] and graph anomaly detection [3]. To our knowledge, this is the first adaptation of TFMs to link prediction. Based on this motivation, we introduce several tailored modules: a prototype-augmented local–global context, and a multi-domain link encoder that injects topology. Relative to UniLP’s context, our prototype-augmented local–global context keeps graph-specific local demonstrations while adding shared learnable prototypes as a compact global context. This lets our method capture both target-graph patterns and cross-graph transferable link-prediction patterns, where UniLP’s purely local attention context cannot explicitly model them. We would revise our paper to further clarify these points.

W2 & Q2. Missing OGB datasets.

Thanks. Following your suggestion, we add results in ogbl-collab. The dataset uses the official split and the metric is Hit@50. TFMLinker achieves the highest metric. We will include the experimental results and analysis in the manuscript. More experiments on larger-scale graphs are left as important future works.

MethodCollab
SEAL59.26±1.29
GAE6.80±0.00
ELPH65.65±0.00
NCNC65.47±0.25
MPLP34.65±3.02
LPFormer64.55±0.00
OCN64.10±0.02
UniLP65.34±0.19
TFMLinker66.24±0.37

W3 & Q3. Missing baselines.

Thanks. Following your suggestion, we add LPFormer [4] and OCN [5]. The results show that TFMLinker greatly outperforms these methods.

MethodCelegansUSAirPBNSCSCoraFacebookAverage
LPFormer60.28±6.5581.15±2.7446.36±2.6882.26±2.4166.57±2.7260.79±1.7568.65±2.1466.58
OCN61.70±4.9576.89±3.3923.85±3.0572.76±2.3861.20±0.6352.42±2.1667.31±1.3359.45
LPFormer+FT61.49±7.0482.89±2.9046.83±2.7982.28±2.3865.73±3.0060.60±1.8166.17±3.7866.57
OCN+FT61.31±4.5975.60±4.0023.85±3.0573.23±2.4062.90±1.0353.40±1.8166.05±1.7959.48
UniLP65.20±4.4085.98±2.0048.14±2.9989.09±2.0564.59±2.6557.50±2.4065.49±2.0568.00
TFMLinker68.62±5.9791.67±1.3747.75±2.7690.46±2.4370.42±1.9162.63±1.7769.06±1.1271.52

W4. Ablation significance.

Thanks. We respectively‌ clarify that removing any of these components consistently hurts in terms of the overall performance across all seven datasets: removing global context leads to 2.29 a drop (71.52→69.23), removing GNN pre-train leads to 1.64 drop (71.52→69.88), removing heuristics leads to 0.85 drop (71.52→70.67). All these differences are statistically significant under a one-sided paired t-test at the 99% confidence level. Since our goal is universal multi-domain link prediction, the average over diverse datasets is the primary evaluation criterion. Fluctuations on individual datasets under reported variance do not contradict the usefulness of these components under this multi-domain average.

Q4. citation2 baselines.

Thanks. We respectively‌ clarify that the difference comes from different settings: NCNC follows the supervised setting, where we follow UniLP [1] and evaluate under a pretrain then fine-tune protocol. Specifically, baselines are pretrained on the UniLP source graphs and then fine-tuned on the target dataset, rather than trained from scratch. This setting is substantially harder than the original OGB evaluation protocol, so many GNN4LP methods that excel in the supervised setting drop significantly in this setting. In contrast, TFMLinker adapts via in-context learning without target-graph fine-tuning and still reaches high MRR, highlighting its advantage for mutli-domain link prediction under the pretrain then finetune setting.

[1] Universal Link Predictor By In-Context Learning on Graphs. TMLR'25.

[2] Turning Tabular Foundation Models into Graph Foundation Models. GFM@ICML'26.

[3] Tabular foundation models are strong graph anomaly detectors. WWW'26.

[4] LPFormer: An Adaptive Graph Transformer for Link Prediction. KDD'24.

[5] OCN: Effectively Utilizing Higher-Order Common Neighbors for Better Link Prediction. NeurIPS'25

As we approach the end of the discussion period, we'd like to follow up to confirm if you have any additional questions regarding our responses. We're more than happy to provide further clarification and would greatly appreciate any other feedback you may have. Thank you for your valuable time and engagement throughout the review and rebuttal process.

Thanks for you reply. It partially addresses some of my concerns. However, I still have issues with the novelty and ablation. I expound below:

Novelty vs. UniLP

I understand it's the first TFM adapted for link prediction. However, it is still quite similar to UniLP. The authors highlight several modules they introduced, including a prototype-augmented local–global context, and a multi-domain link encoder. However the encoder uses DRNL just like UniLP. Furthermore, I'm not convinced the use of heuristics and global context is necessary (see my comment later in this reply). So my concerns wrt to the novelty still stand.

Missing OGB datasets. Missing Baselines

Thank you. I recommend including this in your revised manuscript.

Ablation

I don't think comparing the mean across 7 datasets is necessarily informative. How do you compute the t-test here? It needs to be done on the individual datasets.

Comparing against the heuristics, my back of the envelope calculations are that the increase is only statistically significant on 3/7 datasets. For the global context, it looks like only 1/7 datasets (CS). So my concerns still stand.

citation2

Thanks for the clarification. My apologies for missing it.

Reviewer_5VRNOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Paper introducesTFMLinker for link prediction across different graph domains. It combines a pretrained GNN, structural heuristics, and a tabular foundation model. The model uses labeled links from the target graph as local context and shared learnable prototypes as global context. After training on source graphs, it predicts links on target graphs without further parameter updates. TFMLinker achieves the best results on five of the seven main datasets, with an average Hits@50 of 71.52 compared with 68.00 for UniLP.

Critical strengths
  1. Paper combines GNN-based link representations, structural heuristics, and a tabular foundation model with shared learnable prototypes. This provides a useful approach to link prediction across graph domains without target-specific fine-tuning.

  2. TFMLinker performs best on five of the seven main datasets. The ablation study and experiments with different TFM backbones help support the method.

Critical weaknesses
  1. Some evaluation details are unclear. The paper does not clearly explain which validation graphs and labels are used to select checkpoints and hyperparameters, or what changes across the ten runs. Table 5 also lacks standard deviations. These details are needed to judge the reliability of the reported improvements.

  2. The ablation study does not compare combinations of frozen and trainable pretrained models. These comparisons would help explain how much of the improvement comes from pretrained representations and how much comes from jointly training the components on source graphs.

Questions for authors

Q1 Which validation graphs and labels are used for model selection, and what changes across the ten runs? Could you also provide standard deviations for Table 5?

Q2 Have you compared the full method with settings where the pretrained GNN, TFM, or both are frozen?

Q3 Are PolBlogs and PB versions of the same graph? If so, how do you prevent overlap between source training edges and target test edges?

5 comments

Thanks for your constructive comments and here are detailed responses.

W1 & Q1. Experiment details.

Thanks. We would like to clarify that we exactly follow the setting of UniLP [1]. Specifically, we use a merged validation set, which comprises 200 links from the validation set of each dataset. We only search for checkpoints, and hyperparameters are fixed following UniLP. The test dataset split changes across the ten runs, unless an official split is available. For Table 5 in the paper, we directly copied the results for baselines from the original paper. Following your suggestion, we re-run the experiments by repeating 10 times. The results are as follows. TFMLinker achieves the best metrics on both datasets.

MethodPhotoCitation2
SEAL27.27±2.0677.29±1.88
GAE4.57±0.358.83±1.04
ELPH35.25±2.9576.82±2.08
NCNC41.65±1.5983.35±0.80
MPLP34.05±2.0440.08±1.37
UniLP33.90±2.2381.41±0.93
TFMLinker43.09±3.8186.59±0.82

W2 & Q2. Freeze ablation.

Thanks. Following your suggestions, we add comparison with settings where the pretrained GNN, TFM, and both are frozen. Experimental results show that adapting both the GNN and the TFM has an effect, but the TFM is much more important. Freezing the TFM is worse than freezing both the GNN and the TFM, indicating that the TFM plays a foundational role during adaptation: when the TFM is frozen, further adapting the GNN is unhelpful or even harmful. We will include the experimental results and analysis in the manuscript.

MethodCelegansUSAirPBNSCSCoraFacebookAverage
TFMLinker68.62±5.9791.67±1.3747.75±2.7690.46±2.4370.42±1.9162.63±1.7769.06±1.1271.52
freeze GNN66.81±6.2790.33±1.5046.20±3.4289.58±1.9167.67±2.6760.81±1.3469.84±0.9370.18
freeze TFM54.85±3.9185.18±2.8733.97±4.9787.30±2.3062.84±6.1959.84±1.5148.15±9.3161.73
freeze GNN & TFM55.76±5.5585.18±3.1234.19±4.8187.81±2.5064.07±4.7460.35±2.1948.92±8.2262.33

Q3. PolBlogs / PB overlap.

Thanks. We would like to clarify that we exactly follow the setting of UniLP [1], which uses PolBlogs as a pretraining dataset and PB as a target test dataset. Following your suggestion, we take a closer look and find that these two graphs are indeed extracted from the same underlying Political Blogs network, but with different preprocessing:

  • PolBlogs: the original directed crawl with 1,490 nodes and 19,025 directed edges, plus political-leaning labels; no connected-component filtering.

  • PB: convert to an undirected simple graph and keep the largest connected component, yielding 1,222 nodes and 16,714 undirected edges.

Therefore, we acknowledge that edge overlap may indeed exist. Following your suggestion, we exclude PB and re-plot the main comparison on the remaining six targets as follows. Compared with the second-best method (UniLP), both the improvment in the average score and average-rank are larger if we remove PB from the test set. We will include these results in the revised manuscript.

MethodCelegansUSAirNSCSCoraFacebookAverageAve. Rank
CN46.88±12.2882.75±1.5474.03±1.5956.84±15.5633.85±0.9358.70±0.3558.8413.00
AA61.07±5.1686.96±2.2474.03±1.5968.22±1.0833.85±0.9367.80±2.1265.328.17
RA62.80±4.8487.27±1.8974.03±1.5968.21±1.0833.85±0.9368.84±2.0365.837.50
PA43.85±4.1277.69±2.2935.35±3.016.49±0.6122.09±1.5212.95±0.6333.0716.83
SP0.00±0.000.00±0.0080.00±1.1141.34±35.5852.97±1.530.00±0.0029.0515.42
Katz58.86±6.4884.64±1.8678.96±1.3566.32±5.5952.97±1.5360.79±0.6067.099.42
SEAL61.28±3.7686.00±1.5684.07±1.9662.82±1.6256.21±2.2454.57±1.4867.498.00
GAE44.71±3.3976.12±2.2715.20±2.145.08±0.4824.22±1.536.65±0.5728.6617.17
ELPH59.23±4.5084.42±2.2284.27±1.4370.69±3.6356.91±1.4361.80±2.4669.557.17
NCNC48.07±4.7975.44±4.2280.07±1.4334.27±2.2852.51±2.5419.28±1.5851.6113.67
MPLP56.74±5.3182.94±2.3080.33±1.5424.26±1.2846.71±2.2548.06±2.0956.5112.50
SEAL+FT64.45±4.1488.49±2.1684.84±2.3261.54±3.0962.19±3.2758.70±2.7870.035.33
GAE+FT44.71±4.0774.47±2.9618.34±2.274.95±0.4425.31±1.486.11±0.3928.9817.67
ELPH+FT60.51±5.7284.52±2.0886.08±0.6971.10±3.4857.18±1.8963.31±3.6370.455.83
NCNC+FT64.45±5.1085.85±2.4688.25±2.2558.75±5.7460.00±2.5059.32±6.9369.445.83
MPLP+FT62.56±4.7985.08±1.5480.16±1.1850.35±1.0156.02±2.0956.72±1.2065.159.33
OpenGraph55.94±4.3372.38±7.3385.57±1.5037.68±4.8359.17±2.8313.52±17.4254.0411.50
UniLP65.20±4.4085.98±2.0089.09±2.0564.59±2.6557.50±2.4065.49±2.0571.314.33
TFMLinker68.62±5.9791.67±1.3790.46±2.4370.42±1.9162.63±1.7769.06±1.1275.481.33

[1] Universal Link Predictor By In-Context Learning on Graphs. TMLR'25.

As we approach the end of the discussion period, we'd like to follow up to confirm if you have any additional questions regarding our responses. We're more than happy to provide further clarification and would greatly appreciate any other feedback you may have. Thank you for your valuable time and engagement throughout the review and rebuttal process.

Thank you for the clarifications and additional experiments. The frozen-model comparisons address my ablation concern, and the added standard deviations are helpful. I also appreciate the acknowledgment of the shared source of PB and PolBlogs.

I have a few remaining questions:

Does the merged validation set include target graphs, including PB? After excluding PB from evaluation, was checkpoint selection also performed without PB validation links?

What changes across the ten runs for datasets with official splits?

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

This paper tackles the problem of link prediction, by leveraging the strength of TFM (Tabular Foundation Model) aiming to be able to do link prediction across diverse ranges of graphs (as a generic foundation models for graph tasks). To tackle a few challenges presented by such an approach, the proposed TFMLinker method uses three key components: 1) the prototype-augmented local-global context module, 2) the multi-domain topology-aware link encoder, 3) and the TFM.

Critical strengths

This adoption of TFM for graph tasks such as link prediction is a novel idea. The authors claimed this is a first study of using TFM for graph link prediction.

The authors' motivation is very clearly made. The challenges faced by adopting TFM, and why the 3 key components contribute to handling these problems are well explained.

Writing clarity is of excellent quality, making the paper very easy to read.

Literature is also very comprehensive and related works are well cited.

RQs are clearly defined, and backed by result analysis from experiments.

Critical weaknesses

If you do complete pre-training of your TFMLinker, how do you get data labels easily (ie., ground truth data)? If this is difficult or expensive to collect, what would be the implications?

Tables for presenting results show average values and standard deviations, but I cannot find in the paper where you mentioned how many runs you use to calculate the average values and as well as standard deviations. Please clarify!!

Can you explain in more details how 2-hop ego-subgraph help capture "global" information across multiple graphs?

Questions for authors

Q1: If you do complete pre-training of your TFMLinker, how do you get data labels easily (ie., ground truth data)? If this is difficult or expensive to collect, what would be the implications?

Q2: Tables for presenting results show average values and standard deviations, but I cannot find in the paper where you mentioned how many runs you use to calculate the average values and as well as standard deviations. Please clarify!!

Q3: can you explain in more details how 2-hop ego-subgraph help capture "global" information across multiple graphs?

Q4: How did you choose your data sets for your experiments, eg., Table 1?

Q5: It may seem that one day GFD may be a more natural fit than TFD. Here you use several encoders to create corresponding rows in a table format, but what would be the result of training and inferencing using such encoded data? Any potential drawbacks?

3 comments

Thanks for your constructive comments and here are detailed responses.

W1 & Q1. Ground-truth labels.

Thanks. We would like to clarify that, similar to other graph foundation models like UniLP [1] and RiemannGFM [2], and tabular foundation models like TabPFN [3], our method needs a small amount of context for inference. We analyze the sensitivity of our method with respect to the number of in-context links (i.e., labels) in Figure 2. Our method maintains good results with a reasonbly number of links. Achieving multi-domain link prediction with no labels is our future work.

W2 & Q2. Number of runs.

Thanks. We clarify that we use ten runs as explained in Appendix B.2. We will add a detailed description in the main text of the manuscript.

W3 & Q3. 2-hop global info.

Thanks. We first clarify that the 2-hop ego-subgraphs are served as the local context, which empower our method to capture specific patterns within individual graphs. To capture "global" information across multiple graphs, we maintain a single set of learnable and shared prototypes across all link prediction tasks. By sharing the same prototypes across diverse tasks, we encourage them to learn transferable patterns.

Q4. Dataset selection.

Thanks. We follow the datasets in UniLP [1]. These datasets are selected from diverse domains and exhibit varied graph statistics. They are intentionally chosen to ensure that evaluation encounters a broad spectrum of link prediction connectivity patterns.

Q5. GFM vs TFM.

Thanks. Following your suggestion, we compare with a naive variant that replaces the TFM with UniLP-style attention ICL, randomly initialized without pretrained ICL weights. The results are shown as follows. TFMLinker with the TFM is better on 6/7 datasets and on average. We attribute the gap to the naive attention ICL having weaker in-context learning ability than the TFM due to smaller capacity and no large-scale pretraining.

MethodCelegansUSAirPBNSCSCoraFacebookAverage
TFMLinker68.62±5.9791.67±1.3747.75±2.7690.46±2.4370.42±1.9162.63±1.7769.06±1.1271.52
Attention ICL66.67±6.5988.87±2.2346.00±2.3987.65±2.1463.99±1.9944.32±2.1170.09±1.3266.80

[1] Universal Link Predictor By In-Context Learning on Graphs. TMLR'25.

[2] RiemannGFM: Learning a Graph Foundation Model from Riemannian Geometry. WWW'25.

[3] Accurate predictions on small data with a tabular foundation model. Nature'25.

As we approach the end of the discussion period, we'd like to follow up to confirm if you have any additional questions regarding our responses. We're more than happy to provide further clarification and would greatly appreciate any other feedback you may have. Thank you for your valuable time and engagement throughout the review and rebuttal process.

Thanks for the clarification. I have no further questions. And I'd like to maintain my score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Author Rebuttal

We thank the reviewers for carefully reading our paper and for their constructive feedback, which helped us improve the manuscript. We summarize our responses as follows.

1. Additional Experiments

We added seven sets of experiments, as follows:

  • We added the ogbl-collab benchmark, where TFMLinker achieves the best result.

  • We added the LPFormer and OCN baselines, and TFMLinker achieves the best average and rank.

  • We added ablations that freeze the GNN, the TFM, or both. Adapting the TFM matters most, and freezing it prevents GNN adaptation from helping.

  • We added an attention-ICL variant (UniLP-style, randomly initialized, no pretrained ICL weights), and TFMLinker outperforms the variant on average. The response attributes the gap to the weaker in-context learning ability of a smaller, non-pretrained attention model.

  • We added repeated 10-run results on Photo and Citation2, where TFMLinker is best on both.

  • We added results that we only use pretraining graphs for validation, where TFMLinker achieves the best average and rank.

  • We clarify that removing global context (69.23, −2.29), GNN pretraining (69.88, −1.64), or heuristics (70.67, −0.85) consistently lowers the 7-dataset average. We added a one-sided paired t-test at the 99% level over the 7-dataset averages, under which the response reports these differences as significant. Per-dataset fluctuations are argued not to contradict an average-based criterion, since the goal is multi-domain link prediction.

2. Novelty Relative to UniLP

We would clarify our novelty from the following aspects:

  • This is the first adaptation of TFMs to link prediction, to the best of our knowledge. At a higher level, our goal is framed as bridging tabular and graph data in one foundation model, with link prediction as the validation scenario.

  • The prototype-augmented local–global context captures both graph-specific and transferable patterns: local demonstrations from each target graph capture dataset-specific structure, while shared learnable prototypes provide a compact global context that captures cross-graph link-prediction patterns, which UniLP's purely local attention context cannot model explicitly.

  • The multi-domain link encoder injects topology through heuristic signals, yielding richer link representations than UniLP.

3. Regarding Evaluation Protocol

  • We clarify that our experiments exactly follow UniLP's protocol. The merged validation set contains 200 links from each target dataset's validation split and only checkpoints are selected, and hyperparameters are fixed.

  • We acknowledge that PolBlogs and PB come from the same underlyding network with different preprocessing. We added results by excluding PB from the evaluation and TFMLinker averages 75.48 with rank 1.33, versus the second-best method at 71.31 with rank 4.33. The margin over the second-best method is even larger.

  • We clarify that the gap to OGB-reported numbers is originated from the protocol. OGB adopts the supervised setting. We adopt the pre-training then fine-tuning setting for baselines. TFMLinker adapts by in-context learning without target fine-tuning.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.