acceptodds
Under review as a conference paper at ICLR 2027

Boosting Knowledge Graph Foundation Models via Enhanced Negative Sampling

Abstract

Knowledge graphs (KGs) are the core backbone of numerous downstream tasks such as question answering and recommender systems. However, despite all this, KGs are often very incomplete. To perform zero-shot link prediction in unseen KGs, which have different entity/relation vocabularies from those used for pre-training, KG foundation models (KGFMs) receive a wide range of attention. Existing KGFMs often perform training using random negatives, which are constructed by replacing the head or tail entity of a positive triple with a random entity. However, these negatives are often constructed with limited quality, providing weak supervision for KGFM training. In this paper, we propose a simple yet effective adaptive negative sampling approach, KMAS, to enhance existing KGFMs. KMAS constructs hard negative triples through the updated relation embeddings generated from the existing KGFM's relation encoder. To further adaptively align with the KGFM’s evolving capability to distinguish between positive and negative samples during the training process, KMAS adjusts the ratio of hard negative triples dynamically throughout the whole training process: after a warmup phase, it increases the ratio linearly and then decreases linearly. Extensive experiments are conducted over 44 data sets. Experimental results demonstrate that KMAS can enhance many SOTA KGFMs without requiring excessive additional memory consumption, and it incurs no additional time cost on most KGFMs.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Knowledge graph

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

This paper investigates negative sampling for knowledge graph foundation models (KGFMs) during cross-graph pretraining. The authors argue that negatives obtained by uniformly replacing entities are often too easy to provide sufficiently informative supervision, and propose KMAS to address this limitation.

KMAS uses the foundation model’s current relation encoder to compute similarities between relations and constructs candidate-entity sampling distributions from the head and tail entity sets associated with similar relations. During training, it combines uniformly sampled negatives with these relation-aware negatives and follows a warmup–increase–decrease schedule for the proportion of hard negatives. The sampling distributions evolve as the relation encoder is updated, whereas the mixture schedule is predetermined by training progress.

The main contribution is a training-time negative sampling module applicable to multiple existing KGFMs, rather than a new foundation model architecture or a theoretical advance in expressiveness.

The authors pretrain on three knowledge graphs and evaluate five foundation models on 44 datasets. Average MRR improves for all five models, with absolute gains ranging from 0.004 to 0.008. Ablation studies support the effectiveness of hybrid sampling, continuously updated relation representations, and the proposed mixture schedule. The paper also compares against existing negative sampling methods and reports cross-dataset significance tests for selected models, along with computational overhead. Overall, the experiments support average performance improvements, but do not fully establish the source of these gains or their robustness across random seeds and dataset families.

Critical strengths
  1. Compatibility with Existing Models without Changing Inference

KMAS reuses the existing relation encoder, requires no separate negative-generation network, and leaves the inference procedure unchanged. For models with inexpensive relation encoders, it can potentially improve performance with limited additional training overhead.

If the gains are robustly established, a training module that requires neither architectural redesign nor additional inference cost would have practical value. Simplicity is not itself a limitation: a simple method that consistently improves strong baselines can constitute a meaningful contribution.

However, the low-overhead claim should remain architecture-specific, as demonstrated by the substantially higher training cost for FLOCK.

  1. Evaluation across Multiple Foundation Models

Average performance improves for all five evaluated foundation models. The authors also provide ablations examining hybrid sampling, frozen relation representations, and alternative mixture schedules.

Since the central claim is that KMAS enhances existing KGFMs, effectiveness across architectures is more relevant than achieving the best result on an individual dataset. The reported results provide initial evidence of broad applicability and reduce the likelihood that the method’s success depends entirely on tuning for a single architecture.

The strength of this evidence is nevertheless limited by the absence of reported multi-seed results and the potential dependence among datasets.

Critical weaknesses
  1. The Added Value of the Core Mechanism over Simpler Alternatives Is Not Sufficiently Established

KMAS constructs candidate distributions from the head and tail entity sets of similar relations. This mechanism may primarily impose entity-type or relational-role constraints. The current evaluation does not adequately disentangle the benefit of these constraints from the additional benefit attributed to continuously updated, transferable relation representations.

The most relevant controlled comparisons would include:

  • Sampling from the head and tail entity sets of the target relation itself;
  • Sampling based on static relation–entity co-occurrence information;
  • Using the same mixture schedule without learned relation similarities.

The main technical contribution rests on using transferable relation representations to guide negative sampling. If simpler relational-role constraints combined with the same schedule achieve comparable performance, the current experiments would not establish the necessity of the proposed mechanism, and the contribution would appear to be a relatively incremental combination of existing ideas.

Comparisons with existing negative sampling methods do not fully substitute for controlled experiments that isolate the paper’s central mechanism.

  1. The Modest Average Gains Are Not Yet Supported by Sufficient Evidence of Robustness and Fair Comparison

The average MRR gains range from 0.004 to 0.008, but the paper does not clearly report means, variances, or confidence intervals across multiple random seeds. The per-dataset results also include noticeable regressions, while some of the larger gains are concentrated on individual datasets.

For example, SEMMA improves by 0.108 MRR on Metafam. Under an equally weighted average over 44 graphs, this single result contributes approximately 0.00245, or roughly 30% of the average gain of 0.008 reported in Table 1. This does not invalidate the improvement on Metafam, but it motivates examining how sensitive the aggregate conclusion is to individual datasets.

In addition, the random-sampling ablation KMASsub achieves an average MRR of 0.364 with ULTRA, whereas the original ULTRA baseline in Table 1 achieves 0.360. This discrepancy needs to be explained to isolate the contribution of the proposed sampling mechanism.

Questions for authors

A. Clarifications Regarding the Existing Experiments and Implementation

Question 1: Were the original baselines and their KMAS-enhanced counterparts evaluated using identical training, validation, and model-selection procedures?

Please explain the difference between the original ULTRA random-negative baseline and KMASsub. In particular, were the random seeds, training steps, loss weighting, checkpoint-selection procedures, and hyperparameter search budgets matched?

Question 2: How was overlap between the pretraining and evaluation datasets checked, and which triples were used for negative masking?

Please specify how entity, relation, and triple overlap was assessed. Please also clarify whether training-time negative masking used only facts permitted by the training protocol.

B. Additional Evidence Needed to Support the Current Claims

Question 3: With the mixture schedule and training configuration held fixed, does KMAS significantly outperform simpler relational-role-constrained sampling methods?

The most informative comparisons would be sampling from the target relation’s own head and tail entity sets and sampling based on static relation–entity co-occurrence information.

Question 4: Are the main gains reproducible across random seeds and robust at the dataset-family level?

Please provide multi-seed results for at least representative models. It would also be important to examine whether the conclusions hold when individual high-gain datasets are excluded and when results are aggregated by dataset family.

3 comments

We thank you for your careful assessment and constructive suggestions.

Weaknesses

W1: Thank you for these suggestions. Table 4 compares KMAS with three alternatives: KMASrelation−type\mathrm{KMAS}_{\mathrm{relation-type}}, which samples uniformly from the target relation's head or tail entity set; KMASco−occurrence\mathrm{KMAS}_{\mathrm{co-occurrence}}, which samples according to static relation–entity co-occurrence counts in the training graph; and KMASfrozen\mathrm{KMAS}_{\mathrm{frozen}}, which retains the relation embeddings obtained after warmup for negative sample construction while the KGFM itself continues to train. The first two alternatives use no learned relation similarities, while the frozen variant examines the benefit of updating the relation embeddings. All variants use the same mixture schedule, negative sample budget, masking, training configuration, and validation procedure. KMAS outperforms all three alternatives in both MRR and Hits@10 on ULTRA and SEMMA, demonstrating the benefits of learned relation similarities and continuously updated relation embeddings.

Table 4. Negative sample construction with the same adjustment strategy, averaged over 44 datasets.

Negative sample constructionULTRA MRRULTRA Hits@10SEMMA MRRSEMMA Hits@10
KMASrelation−type\mathrm{KMAS}_{\mathrm{relation-type}}0.3620.5190.3660.523
KMASco−occurrence\mathrm{KMAS}_{\mathrm{co-occurrence}}0.3540.5130.3650.527
KMASfrozen\mathrm{KMAS}_{\mathrm{frozen}}0.3470.5090.3730.532
KMAS0.3670.5320.3800.539

W2: We appreciate this concern. Table 2 in our response to Reviewer xpBs reports five-seed results for two representative KGFMs, ULTRA and SEMMA, and their KMAS-enhanced counterparts. Regarding the MRR difference between the original ULTRA and its KMASsub\mathrm{KMAS}_{\mathrm{sub}} variant in the submitted paper, the original sampler generates all random negatives in one random sampling, whereas KMASsub\mathrm{KMAS}_{\mathrm{sub}} uses two independent random sampling, replacing the relation-aware part with a second random group. Consequently, matching seeds do not guarantee identical sampled negatives or training trajectories. This sampling difference may lead to either higher or lower performance, as illustrated by SEMMA's Hits@10 decreasing from 0.529 to 0.527 with KMASsub\mathrm{KMAS}_{\mathrm{sub}} in the submitted paper.

Questions

Q1: Each original baseline and its KMAS-enhanced counterpart use matching random seeds, training steps, loss weighting, checkpoint-selection procedures, and hyperparameter search budgets. Hyperparameters and checkpoints are selected only on the validation sets of FB15k-237, WN18RR, and CoDEx-Medium; the evaluation datasets are not used for model selection. Please see our response to W2 for an explanation of the difference between the original ULTRA sampler and KMASsub\mathrm{KMAS}_{\mathrm{sub}}.

Q2: We follow the dataset construction, splits, and pre-training and zero-shot evaluation protocol of previous KGFMs: pre-training on FB15k-237, WN18RR, and CoDEx-Medium and evaluating on the other benchmark datasets without training or fine-tuning on them. For training-time negative masking, strict_negative_mask excludes tails t′t' for which (h,r,t′)(h,r,t') is observed in the current pre-training graph when corrupting a query (h,r,?)(h,r,?); it analogously excludes observed heads for (?,r,t)(?,r,t). Thus, “the entire set of triples” in Section 3.3 refers to the observed triples in the current pre-training graph. Validation and test facts used for filtered ranking belong to the evaluation procedure and are not supplied to the training negative sampler.

Q3: Thank you for specifying the controlled comparison. Table 4 in our response to W1 compares KMAS with KMASrelation−type\mathrm{KMAS}_{\mathrm{relation-type}} and KMASco−occurrence\mathrm{KMAS}_{\mathrm{co-occurrence}}, using the same warmup, mixture schedule, and training configuration. The former assigns equal probability to distinct entities in the corresponding set; the latter weights them by their observed occurrence counts, without learned relation similarities or updates during training. Known positives are filtered in both variants; an empty candidate set falls back to the target relation's combined head and tail entity set under the same rule. The random-negative proportion, total negative sample budget, and model-selection procedure are unchanged. This comparison examines the additional contribution of learned relation similarities, while the frozen variant separately examines their continuous updating.

Q4: Table 2 in our response to Reviewer xpBs reports five-seed results for two representative KGFMs, ULTRA and SEMMA, and their KMAS-enhanced counterparts to assess variability across training runs. For ULTRA and SEMMA, Table 5 below reports MRR gains on all 44 datasets and by dataset family.

Table 5. MRR gains from adding KMAS to ULTRA and SEMMA: mean ± standard deviation over five seeds. Datasets within each family are equally weighted.

Evaluation collection or dataset familyNumber of datasetsULTRA MRR gainSEMMA MRR gain
All evaluation datasets44+0.0068 ± 0.0029+0.0091 ± 0.0030
CoDExSmall / CoDExLarge2+0.0058 ± 0.0056+0.0016 ± 0.0055
NELL995 / NELL23k2+0.0130 ± 0.0158+0.0062 ± 0.0116
WDsinger1+0.0197 ± 0.0130+0.0061 ± 0.0070
FB15k237(10) / FB15k237(20) / FB15k237(50)3+0.0061 ± 0.0053+0.0051 ± 0.0026
Hetionet1+0.0140 ± 0.0133+0.0061 ± 0.0290
WN18RRInductive4+0.0114 ± 0.0309+0.0013 ± 0.0070
FB15k237Inductive4+0.0044 ± 0.0037+0.0034 ± 0.0056
NELLInductive3+0.0050 ± 0.0101+0.0158 ± 0.0105
ILPC20222+0.0074 ± 0.0044+0.0035 ± 0.0035
HM3-0.0014 ± 0.0036+0.0009 ± 0.0038
FBIngram4+0.0060 ± 0.0026+0.0024 ± 0.0014
WKIngram4+0.0022 ± 0.0079+0.0028 ± 0.0035
NLIngram3-0.0070 ± 0.0091+0.0092 ± 0.0131
WikiTopics6+0.0053 ± 0.0090+0.0039 ± 0.0039
Metafam1+0.0713 ± 0.1191+0.1950 ± 0.1511
FBNELL1+0.0062 ± 0.0053+0.0164 ± 0.0145

Thank you for the clarifications and additional experiments. They substantially address my main concern, and I now view the contribution of this article more positively. My remaining concerns mainly concern the methodological novelty. I view the novelty primarily as a useful mechanism design tailored to KGFM pretraining, rather than a fundamentally new negative sampling or curriculum learning principle. The contribution is valuable, but relatively incremental. Overall, I am updating my score from 2 (Weak rejection) to 3 (Weak acceptance), while retaining reservations about the strength of the novelty and the broader mechanistic claims.

Reviewer_xpBsOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes KMAS, a hard negative sampling strategy to improve KGFM performance. KMAS selects similar relations and samples negative entities from either the head or tail position. It then combines these hard negatives with randomly sampled negatives. The authors apply KMAS to five KGFMs and evaluate performance across 44 datasets in diverse settings.

Critical strengths

S1. The method is plug-and-play and is applied to five KGFMs.

S2. Training time and memory overhead are reported. The overhead is small for GNN-based KGFMs.

S3. The evaluation is broad. It covers 44 datasets across transductive, partially inductive, and fully inductive settings.

Critical weaknesses

W1. The novelty claims are insufficiently supported. The authors claim novelty in combining a relation encoder with negative sampling; however, existing methods such as self-adversarial sampling (Sun et al., 2019) and NSCaching (Zhang et al., 2019) also use the current model state to select negative samples. Furthermore, the authors never empirically evaluate the claim that the encoder and the negative sampling strategy mutually reinforce each other.

W2. Related work undermines the claimed novelty. The authors call the dynamic adjustment of the hard-negative ratio a clear departure from existing negative sampling techniques. Yet, in the same section, DANS is described as a curriculum-based strategy that transitions from "easy" to "hard" negative samples, and ARNS is presented as a method that adjusts the ratio of relation-aware to random negative samples during training.

W3. The results lack reliability because they use only a single random seed. Also, even if the Wilcoxon signed-rank test is assessed across datasets, it does not address the model's robustness to different random seeds.

Questions for authors

Q1. Could you report the mean and standard deviation over at least five seeds for both the base models and KMAS?

Q2. How were the hyperparameters selected? All 44 datasets are used for zero-shot evaluation, so selecting hyperparameters on them would leak test information.

Q3. In Table 3, SimKGC or BIGRAM lowers ULTRA’s MRR from 0.360 to 0.277 and 0.282. I think BIGRAM also samples from relation–entity co-occurrence, which is close to KMAS. Did you run these baselines with the same warm-up, random-negative mixing, and ratio schedule as KMAS? If not, this ablation would show whether the gap comes from the negatives themselves or the schedule.

3 comments

We thank you for your careful review and for recognizing the value and applicability of KMAS.

Weaknesses

W1: Thank you for your insightful comment. We will remove these overstated statements from the paper. It is worth noting that remove these contents does not change our contributions: (1)we conduct the first investigation of negative sampling for KGFMs; (2)we propose a simple yet effective negative sampling method that constructs hard negative samples with a heuristic hard negative ratio adjustment strategy and can be flexibly adapted to many KGFMs; (3)We conduct extensive experiments on 44 data sets. Experimental results demonstrate that our proposed negative sampling method can improve the performance of many SOTA KGFMs without requiring excessive additional memory consumption, and it incurs modest additional time cost on most KGFMs.

W2: Thank you for this comment. We would like to clarify that KMAS's increase-then-decrease hard negative ratio adjustment strategy differs from an easy-to-hard curriculum. Progressing from "easy" to "hard" corresponds to gradually increasing the proportion of hard negatives. In KMAS, after a warmup phase with random negatives, this proportion first increases linearly to a peak and then decreases linearly, so random negatives receive a larger proportion again in the later training stage. This subsequent decrease distinguishes our strategy from a schedule that only increases the hard negative ratio. ARNS adjusts mixture ratios using MRR feedback, whereas KMAS follows the piecewise linear schedule specified in Section 3.4. The comparisons with KMASincrease\mathrm{KMAS}_{\mathrm{increase}} and KMASdecrease\mathrm{KMAS}_{\mathrm{decrease}} in Section 4.5 examine the contribution of this specific adjustment strategy. A possible explanation is that including more easy negatives after the peak may help the model maintain its ability to distinguish them from positive samples, rather than focusing only on hard negatives.

W3: Thank you for raising the issue of reliability. During the rebuttal period, we have made our best effort to conduct the tests across datasets and evaluation across training seeds. Table 2 reports mean ± standard deviation over five random seeds for two representative KGFMs, ULTRA and SEMMA, and their KMAS-enhanced counterparts.

Questions

Q1: Please see Table 2. For each seed, we first average MRR and Hits@10 over the same 44 datasets used in the main comparison, then calculate the mean and sample standard deviation of these averages across the five runs. The same seed values are used for each pair of models, with matching training steps, loss weighting, validation procedures, and hyperparameter search budgets. Thus, the reported standard deviation measures variation across independently trained runs, rather than variation between datasets.

Table 2. Zero-shot link prediction for ULTRA and SEMMA with and without KMAS on 44 datasets: mean ± standard deviation over five seeds.

ModelMRR (mean ± std.)Hits@10 (mean ± std.)
ULTRA0.3606 ± 0.00120.5229 ± 0.0028
ULTRA + KMAS0.3674 ± 0.00240.5302 ± 0.0019
SEMMA0.3704 ± 0.00210.5287 ± 0.0035
SEMMA + KMAS0.3796 ± 0.00130.5371 ± 0.0015

Q2: We apologize for the incomplete description in Section 4.1. All hyperparameters and checkpoints were selected using only the validation sets of the three pre-training KGs: FB15k-237, WN18RR, and CoDEx-Medium. Neither the validation sets nor the test sets of the 44 evaluation datasets were used for hyperparameter or checkpoint selection. The selected configuration and checkpoint were then used for zero-shot evaluation without target-dataset tuning or fine-tuning. Each base model and its KMAS-enhanced counterpart use the same validation and model-selection procedures.

Q3: Thank you for identifying this issue. The methods in Table 3 of the submitted paper used their original negative sampling and, where applicable, adjustment strategies, rather than KMAS's adjustment strategy. Table 3 below provides the controlled comparison on ULTRA and SEMMA. For each method, we use the same total number of negatives and the same αm\alpha_m to determine the proportions of uniformly sampled negatives and negatives generated by that method. Warmup, peak position, maximum ratio, head/tail replacement, negative masking, training steps, loss weighting, and validation are held fixed. For DANS and ARNS, their internal curriculum or feedback-based adjustment is replaced by KMAS's strategy, retaining negative sample construction with fixed internal mixture weights. This comparison evaluates the sampling components under a shared adjustment strategy, separately from the original method comparisons in Table 3 of the submitted paper. Under the same adjustment strategy, KMAS achieves the highest MRR and Hits@10 on both ULTRA and SEMMA, supporting the effectiveness of its negative sample construction.

Table 3. Negative sampling methods with KMAS's adjustment strategy on ULTRA and SEMMA, averaged over 44 datasets. Random sampling remains uniform throughout.

Negative sampling methodULTRA MRRULTRA Hits@10SEMMA MRRSEMMA Hits@10
Random sampling0.3600.5240.3720.529
SIMKGCneg\mathrm{SIMKGC}_{\mathrm{neg}}0.3640.5220.3670.532
BIGRAMneg\mathrm{BIGRAM}_{\mathrm{neg}}0.3530.5120.3630.524
SATKGCneg\mathrm{SATKGC}_{\mathrm{neg}}0.3530.5070.3670.524
DANSneg\mathrm{DANS}_{\mathrm{neg}}0.3610.5220.3720.533
ARNSneg\mathrm{ARNS}_{\mathrm{neg}}0.3640.5270.3740.538
KMAS0.3670.5320.3800.539

Thank you for the additional experiments and clarifications: the five-seed results on ULTRA and SEMMA, the clarified model-selection protocol, and the comparisons under a shared schedule substantially address my experimental concerns. However, while these results strengthen the evidence of practical benefit, I still find the methodological novelty relative to existing negative sampling methods and curriculum strategies limited. I therefore maintain my rating.

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

This paper proposes a hard negative sampling methods for knowledge graph foundation Model Training. The hard negative sampling method relies on relation embeddings to select semantically similar relations, and regards the entities connected by the semantically similar relations as candidate hot negative samples. These methods is general and can be applied to the training of a few existing knowledge graph foundation models. Experiments prove that applying the new negative sampling method KMAS, the trained knowledge graph foundation models performs better.

Critical strengths
  1. Improving the performance of knowledge graph foundation Model Through better negative sampling methods makes sense, just like in the area of knowledge graph embedding.
  2. The zero shot link prediction results show that KMAS is general enough that can be applied to improve knowledge graph foundation models including ULTRA, SCR, MOTIF, SEMMA and FLOCK.
Critical weaknesses
  1. During the experiment of zero-shot link prediction, 44 datasets are applied during test. But in the original data sets for knowledge graph foundation model, There are 54 KGs. The reason of not the all the 54 datasets are used and only 44 of them are selected is not explained.

  2. Since the number of data sets are different from the standard setting, The experiment results of knowledge graph foundation model baselines are different from the original paper. For example, the original MRR results of FLOCK from the paper on the zero-shot entity prediction task is 0.391. While in table 1, The MRR results of FLOCK on 44 graphs is 0.380 and with FLOCK+KMAS is 0.386.

  3. In table 3, the other negative sampling baselines applied to ULTRA makes the performance worse than random sampling. Is there any reason for this phenomenon?

Questions for authors

See weakness section.

2 comments

We thank you for your careful review and for recognizing the value and applicability of KMAS.

Weaknesses

W1: We apologize for not explaining the reason of dataset selection. The evaluation on 44 datasets was limited by the available computational resources and evaluation time. In particular, FLOCK evaluation takes approximately 150 hours in our environment. During the rebuttal period, we have made our best effort to conduct the tests. Here, we evaluate ULTRA, SEMMA, and FLOCK, together with their KMAS-enhanced counterparts, on all 54 datasets (Table 1).

W2: Because the number of datasets and the experimental environments differ, the results of the same method differ between the 54 datasets in the original paper and the 44 datasets in this paper. Table 1 of the submitted paper reports our reproductions using the official implementations, averaged over 44 datasets. Our experiments use NVIDIA 4080 GPUs, whereas the original KGFMs experiments, FLOCK for example, used H100 GPUs. It is worth noting that, for fairness,within our study each base model and its KMAS-enhanced counterpart use the same hardware, datasets, training steps, loss weighting, and model-selection procedure.

Table 1. Zero-shot link prediction on all 54 datasets.

ModelMRRHits@10
ULTRA0.3600.520
ULTRA + KMAS0.3670.526
SEMMA0.3720.525
SEMMA + KMAS0.3770.532
FLOCK0.3820.552
FLOCK + KMAS0.3870.556

W3: Thank you for raising this question. As discussed in Section 4.6, their poor performance may stem from their limited exploitation of cross-KG transferable invariances during negative sampling. Negatives that are hard within a pre-training KG are not necessarily informative for generalization to unseen KGs. If negative selection emphasizes graph-specific statistics or local structures, it may encourage the model to focus on patterns that transfer poorly, making these negatives less useful than uniform random negatives for zero-shot prediction. KMAS uses the cross-KG transferable invariances encoded in the KGFM’s continuously updated relation embeddings to identify similar relations and construct hard negatives from their head or tail entity sets. This may provide more informative supervision for the KGFM to further learn these invariances.

Reviewer_WArFOpenReview ↗

After reading the feedback from the authors

Thanks authors for the feedback. I prefer to keep my score.

Comments on the paper

1 comment

Dear AC,

Thank you and the reviewers for the time and effort devoted to evaluating our paper. We have addressed all weaknesses and questions raised by the reviewers and reviewer gbWt has increased the rating from 2 to 3. We respectfully emphasize that the concerns raised by the reviewers do not affect the core contributions of our work:

  1. We conduct the first investigation of negative sampling for knowledge graph foundation models (KGFMs).
  2. We propose a useful mechanism design tailored to KGFM pretraining.
  3. We conduct extensive experiments on 44 data sets. Experimental results demonstrate that our proposed mechanism can improve the performance of many SOTA KGFMs without requiring excessive additional memory consumption, and it incurs modest additional time cost on most KGFMs.

All experimental results presented in our rebuttal, will be incorporated into the revised paper, together with the corresponding experimental settings, analyses, and clarifications.

We sincerely thank you and the reviewers again for your careful consideration. We hope our responses and the additional evidence assist your assessment of the paper.

Kind regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.