Universal Prompt Tuning for Graph Neural Networks

In recent years, prompt tuning has sparked a research surge in adapting pre-trained models. Unlike the unified pre-training strategy employed in the language field, the graph field exhibits diverse pre-training strategies, posing challenges in designing appropriate prompt-based tuning methods for graph neural networks. While some pioneering work has devised specialized prompting functions for models that employ edge prediction as their pre-training tasks, these methods are limited to specific pre-trained GNN models and lack broader applicability. In this paper, we introduce a universal prompt-based tuning method called Graph Prompt Feature (GPF) for pre-trained GNN models under any pre-training strategy. GPF operates on the input graph's feature space and can theoretically achieve an equivalent effect to any form of prompting function. Consequently, we no longer need to illustrate the prompting function corresponding to each pre-training strategy explicitly. Instead, we employ GPF to obtain the prompted graph for the downstream task in an adaptive manner. We provide rigorous derivations to demonstrate the universality of GPF and make guarantee of its effectiveness. The experimental results under various pre-training strategies indicate that our method performs better than fine-tuning, with an average improvement of about 1.4% in full-shot scenarios and about 3.2% in few-shot scenarios. Moreover, our method significantly outperforms existing specialized prompt-based tuning methods when applied to models utilizing the pre-training strategy they specialize in. These numerous advantages position our method as a compelling alternative to fine-tuning for downstream adaptations.

Paper

Similar papers

Peer review

Reviewer KwHM5/10 · confidence 4/52023-07-03

Summary

The paper introduces a universal prompt-based tuning method called Graph Prompt Feature (GPF) and its variation (GPF-plus) for pre-trained Graph Neural Network (GNN) models. GPF is a universal method that can be applied to any pre-trained GNN model under any pre-training strategy. It operates on the input graph's feature space and can achieve an equivalent effect to any form of prompting function. GPF introduces a learnable vector $p$ of dimension $F$, which is added to the node features, where $F$ corresponds to the dimensionality of the node features. The authors also purpose GPF-plus, which assigns an independent learnable vector $p_i$ to each node $u_i$ in the graph, instead of a single vector $p$. Experimental results show that GPF outperforms fine-tuning, with an average improvement of about 1.4% in full-shot scenarios and 3.2% in few-shot scenarios.

Strengths

1) Theoretical Guarantees: The authors provide rigorous derivations and theoretical analyses to demonstrate the universality and effectiveness of GPF. They prove that GPF can achieve an equivalent effect to any form of prompting function and can outperform fine-tuning in certain cases. This theoretical foundation strengthens the credibility of their proposed method. 2) Experimental Validation - Reproducibility: The paper includes extensive experiments conducted under various pre-training strategies, including both full-shot and few-shot scenarios. The experimental results consistently show that GPF outperforms fine-tuning, achieving an average improvement of about 1.4% in full-shot scenarios and 3.2% in few-shot scenarios. Furthermore, GPF surpasses existing specialized prompt-based tuning methods designed for specific pre-training strategies. Moreover, the availability of the source code enhances the reproducibility of the experiments.

Weaknesses

1) A weakness of the paper is the lack of a clear explanation or motivation regarding why the addition of the learnable vector p to the input features in the GPF model leads to better results compared to linear probing with a trainable layer in the final layer. While the Appendix demonstrates the superior performance of GPF, the paper does not provide a comprehensive analysis or reasoning behind this improvement. The absence of a clear explanation may leave readers questioning the underlying mechanisms and factors that contribute to the observed performance gain. Without a proper understanding of the motivations and justifications for the proposed approach, it becomes challenging to assess the significance and generalizability of the findings. 2) Theorem 2 assumes a simple linear 1-layer GNN (without activations). While the theorem provides theoretical insights into the convergence properties of this specific type of GNN, it may not accurately capture the behavior of more complex GNN architectures commonly used in practice. In real-world applications, GNNs often incorporate activation functions to introduce non-linearity and capture more intricate patterns in graph data. By focusing solely on a linear 1-layer GNN without activations, the theorem may limit the generalizability of its conclusions and overlook important aspects of GNN models commonly used in practical scenarios. Therefore, the applicability of the theorem's findings to more sophisticated GNN architectures with non-linear activations remains uncertain. 3) One weakness of the paper is the exclusive use of Graph Isomorphism Network (GIN) as the backbone GNN for finetuning. While GIN is a widely used GNN architecture, it may not be the most expressive or optimal choice for every graph-related task[1,2]. The paper does not provide a comprehensive exploration or comparison of different backbone GNN architectures in the finetuning process. Moreover, maybe a more powerful GNN would be able to achieve better performance in the fine-tuning stage. Therefore, it would be interesting for the authors to examine if their approach leads also to better results when a more powerful GNN is used. [1] Frasca, Fabrizio, et al. "Understanding and extending subgraph gnns by rethinking their symmetries." Advances in Neural Information Processing Systems 35 (2022): 31376-31390. [2] Morris, Christopher, et al. "Weisfeiler and leman go neural: Higher-order graph neural networks." Proceedings of the AAAI conference on artificial intelligence. Vol. 33. No. 01. 2019.

Questions

1) Explanation of the relationship between GPF and linear probing: The paper compares GPF with linear probing, which involves a trainable layer in the final layer. It would be beneficial if the authors could discuss the relationship between these two approaches and explain why GPF, with its learnable vector added to the input features, achieves superior performance compared to linear probing. Is there any inherent advantage or characteristic of GPF that enables it to outperform linear probing? 2) Generalizability of Theorem 2: The paper presents Theorem 2, which focuses on a simple linear 1-layer GNN without activations. However, many practical GNN architectures incorporate activation functions to introduce non-linearity. It would be helpful if the authors could discuss the generalizability of the theorem's conclusions to more complex GNN models commonly used in real-world applications. Can the findings of Theorem 2 be extended to GNNs with non-linear activations? If not, what are the limitations or implications of the theorem in the context of practical GNN architectures? 3) Have you considered or experimented with using more powerful GNN architectures, such as subgraph GNNs or higher order GNNs, as the backbone GNN for the finetuning stage? If so, how does the performance of GPF and GPF-plus compare when utilizing these alternative GNN architectures? I will be happy to increase my score for the paper, if the authors adequately address the mentioned weaknesses and provide satisfactory explanations and improvements in response to the questions and suggestions raised. *** After rebuttal increased score to "Borderline Accept" **

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

yes

Reviewer dbJu6/10 · confidence 4/52023-07-04

Summary

This paper proposed a universal prompt-based tuning method, called GPF, for pre-trained GNN models. The idea is to operate on the feature space of the downstream input graph. The authors theoretically showed that GPF can achieve results equivalent to any prompting function, and is not weaker than full fine-tuning. Experiments showed the competitive results.

Strengths

1. This paper proposed a universal graph prompting approach, which basically operates on the input graph feature space. As far as I know, the idea of augmenting input graph node features for prompting is novel. 2. The authors provide theoretical analysis, which showed that GPF can achieve results equivalent to any prompting function, and is not weaker than full fine-tuning. 3. Experiments results showed the performance gains.

Weaknesses

1. The motivation for a universal prompting needs better explained, and advantages over specialized prompting approaches. 2. Comparison with other prompting meothds is only perfomed on the models pre-trained by edge predcition. Some minor comments: 1. It is not clear to me how the avg. in Table 1 and Table 2 obtained, and what does it mean. It is just the average across all datasets? 2. I suggest the authors include the detailed pretraining settings in experiments (e.g., datasets used) instead of in the appendix. 3. Fig. 1 is not informative. The authors should use a figure to better illustrate how GPF works, and the main differences with existing prompting approaches. 4. Table captions should be on top of tables insead of below.

Questions

See the weaknesses.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Potential limitations of the proposed GPF not discussed.

Reviewer Cfab5/10 · confidence 4/52023-07-06

Summary

This paper aims for efficient adaptation of pre-trained graph neural networks to downstream tasks. A simple prompt tuning method (i.e. GPF) is proposed for adaptation and is applicable to GNN pretrained with any objectives. The main idea of GPF is to add a learnable vector on all the node features in the input graph, while an improved version (i.e. GPF-plus) leverages a set of learnable vectors to compose the bias added to each node feature. Experiments show that GPF and GPF-plus achieve better results than the full finetuning.

Strengths

Strength 1. Figure 1 illustrates the difference between the proposed method and existing methods. 2. Provide a theoretical analysis of the proposed method. 3. The proposed method is simple and easy to understand.

Weaknesses

Major 1. The novelty of GSF may be limited because it just adds a bias term on the input of the pretrained network and there is nothing specific for the graph structure. However, the bias-term finetuning is not a new idea [a, b]. 2. The authors claim that the developed method is universal for all the graph learning. However, the theoretical proof is only valid for graph classification downstream tasks where the node features are sum-pooled as the representation. There is no evidence showing that GSF is applicable to other types of downstream tasks. 3. Baselines missing in Table 1 & 2. For example, Linear Probing is a standard baseline, and residual adapter in [c] can also be a baseline. 4. Given the variance shown in Table 1 & 2, the improvement of GPF-plus over GPF is marginal. Why is it an enhanced version? If GPF is theoretically feasible to address the problem (section 3.4), why there GPF-plus is needed? [a] BitFit: Simple parameter-efficient fine-tuning for transformer-based masked language-models. ACL’22. [b] Tinytl: Reduce memory, not parameters for efficient on-device learning. NeurIPS’20. [c] CLIP-Adapter: Better Vision-Language Models with Feature Adapters.

Questions

The author is suggested to address the concerns in the weakness section.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

Limitation is not discussed in the paper.

Reviewer GUF85/10 · confidence 4/52023-07-06

Summary

This paper introduces the Graph Prompt Feature (GPF) approach, which aims to adapt pre-trained Graph Neural Networks (GNNs) for downstream tasks by appending tunable embeddings onto the frozen node embeddings. By doing so, the authors achieve a significant reduction in the number of parameters required for the downstream task compared to full fine-tuning. The experimental results demonstrate promising performance in binary classification tasks on Molecular and Protein datasets when compared to full fine-tuning.

Strengths

The paper is clearly written, the contribution is easy to understand and provide theoretical proof on effectiveness.

Weaknesses

There are several concerns that I would like to see addressed: 1) In the formulation of GPF, its relation to the concept of prompting seems unclear. It is challenging to connect this approach to the motivating concept of prompting in NLP. Instead, GPF appears to resemble task-specific adaptive fine-tuning in transformers, such as LoRA[1]. It would be helpful if the authors could elaborate more on where the similarity between GPF and prompting lies. 2) The choice of downstream tasks seems to be limited to small binary classification tasks, and the comparison is only made against fine-tuning as a baseline. Why are there no results presented for multi-class classification tasks? This decision weakens the potential of the proposed method. 3) Unlike previous graph prompt methods, GPF does not make any modifications to the graph's structure. It would be beneficial if the authors could explain how graph classification highlights the strength of GPF, as the additional results provided may seem somewhat redundant. 4) There is no detail provided on the number of features appended into the embedding space for each node. Since GPF-plus suggests a simple size range, does this mean that the authors used different settings for GPF-plus for each dataset? This is not discussed in the experimental settings and might make the comparison somewhat incomparable. [1] https://arxiv.org/abs/2106.09685

Questions

see weaknesses

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

The author did not address any limitations.

Reviewer aqDC7/10 · confidence 4/52023-07-06

Summary

This paper proposes the Graph Prompt Feature (GPF) to improve Graph Neural Networks (GNNs) performance amidst scarce labeled data and low out-of-distribution generalization. GPF, a universal prompt-based tuning method, operates on the input graph's feature space and is applicable to any GNN architecture. The authors also present a stronger variant, GPF-plus, providing diverse prompted features for different nodes. Both methods show better performance than fine-tuning, validated through theoretical analyses and extensive experiments.

Strengths

1. GPF and GPF-plus are universal, model-agnostic solutions that can be applied to any pre-trained GNN model, enhancing their general applicability in diverse scenarios. 2. The paper provides theoretical guarantees for the effectiveness of GPF and GPF-plus, strengthening the scientific rigor of the presented methods. 3. The authors demonstrate that their proposed methods outperform existing fine-tuning strategies and specialized prompt-based tuning methods in both full-shot and few-shot scenarios, showcasing their practical effectiveness.

Weaknesses

1. The effectiveness of GPF and GPF-plus largely depends on the quality and representation of the feature space, which may not be optimally prepared in all real-world applications. 2. While the paper provides theoretical analysis for the methods' effectiveness, more detailed explanation or examples of the theoretical proofs could enhance the clarity and comprehensibility. 3. The real-world applicability and performance of the proposed methods could be further substantiated with experiments on a wider range of tasks and datasets.

Questions

1. How do GPF and GPF-plus perform in scenarios where the feature space is noisy or inadequately represented? Are there strategies to mitigate these potential issues? 2. Could you provide more detailed explanation or real-world examples to further illustrate the theoretical guarantees of GPF and GPF-plus? 3. How would GPF and GPF-plus integrate with other advancements in GNNs or machine learning in general?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Not really. Dear authors, please enumerate some limitations of your work.

Reviewer aqDC2023-08-17

I read the authors' response, and maintain the same rating of 7.

Reviewer dbJu2023-08-16

Thanks for the rebuttal. I have read comments from other reviewers, as well as rebuttals, and I think my score is resonable and will keep my rating.

Reviewer KwHM2023-08-16

Response to Authors

I would like to thank the authors for their rebuttal and the new experimental results. I increase my score to "borderline accept".

Reviewer Cfab2023-08-18

Thank you for rebuttal

Thanks the author for providing the rebuttal. My concerns have been addressed and the score will be increased.

Reviewer GUF82023-08-18

I'd like to express my gratitude to the authors for their thorough response. After careful consideration, I'm convinced that the authors have adeptly addressed all the concerns I raised. In light of this, I will revise my score to a 5. Thank you for your diligent efforts.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC