An Efficient Memory Module for Graph Few-Shot Class-Incremental Learning

Summary

This paper tackles the problem of graph few-shot class-incremental learning, proposes a novel framework named Mecoin to address the challenges of catastrophic forgetting and overfitting. Mecoin includes two key componets: the structured memory unit (SMU) for storing and updating prototypes and the memory representation adaptive module (MRaM) for separating the learning of prototypes and class representations. Based on SMU, Mecoin efficiently maintain representative prototypes preventing the overfitting problem caused by few-shot setting, and rely on MRaM, Mecoin avoids heavily finetuning the model parameters which solved the catastrophic forgetting problem in class-incremental learning.

Strengths

1.Stores past knowledge efficiently with low storage space by using the SMU. 2.Avoids catastrophic forgetting by separating the learning of prototypes and class representations. 3.The experiments show significant improvement compared to other state-of-the-art methods.

Weaknesses

1.The abbreviations of the proposed method are confusing. For example, in some places, the Memory Representation Adaptive Module is abbreviated as MRaM, while in others, it is MRAM. Additionally, the full model name, Mecoin, is too similar to one of its submodules, the Memory Construction module (MeCo). 2.Fig. 1 is confusing; it doesn’t show the details and advantages of the SMU and MRaM structures. 3.In Sec. 3.1, Eq. 6 lacks explanatory information. How does Eq. 6 integrate edge information into the loss function? 4.The proof of Theorem 1 is too brief. I recommend the authors provide detailed derivation rather than citing a lemma from another paper. 5.In Sec. 3.2, Eqs. 8 and 9 are incomprehensible. I assume the GKIM maintains past knowledge and updates new knowledge with different elements 𝑝. For example, if the dataset has 100 classes, 𝑝1:60 would be maintained for past knowledge and 𝑝61:100 for new knowledge. Thus, 𝑁 should be 60. It’s challenging to explain how new knowledge is updated in Eq. 9. 6.In Sec. 4.1, the process of pretraining the GNN needs to be explained. Is the pretraining phase a standard graph node classification task or a few-shot node classification task? Additionally, in Tab. 3, the comparison method MAG-Meta is mentioned twice; I guess one should be Geometer. Also, if the proposed method is named Mecoin, why does Mecoin appear among the comparison methods in Tabs. 2, 3, and 4? 7.In Sec. 4.2, the toy dataset is too simple. Sampling only 4 classes makes it difficult to demonstrate the ability of different settings to handle overlapping problems. I recommend sampling at least 10 classes.

Questions

Please check the weaknesses

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

The paper is poor-writing and many places lack detailed explanations.

Summary

The authors focus on graph few-shot class-incremental learning. The authors first introduce Mecoin to efficiently construct and preserve memory. To avoid extensive parameter finetuning and forgetting, the authors introduce a memory representation adaptive module called MRaM to separate the learning of prototypes and class representations. Besides, the authors propose Graph Knowledge Interchange Module (GKIM) to injects past knowledge information into GNN. Additional analyses illustrate the effectiveness of the methods.

Strengths

1. The paper is well-written, and the motivations for each part are clear. 2. The reported performance out-performs many baselines. 3. The theoretical and experimental analyses successfully illustrate the effectiveness of the methods.

Weaknesses

1. The paper does not report error bars, standard deviations or provide detailed information about the statistical significance of the experiments, which is important for understanding the reliability and variability of the results.

Questions

See Weaknesses.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The paper has a section discussing several limitations in the appendix.

Summary

This paper mainly focuses on graph few-shot class-incremental learning. To alleviate the significant memory consumption and catastrophic forgetting of old knowledge, it proposes to store only old class prototypes and update class prototypes by considering the interaction among nodes and prototypes. Furthermore, it gives a theoretical analysis on the generalization error. Experimental results show that the proposed method can achieve better results.

Strengths

The proposed idea of storing old class prototypes and updating class prototypes by considering their interaction with node features seems reasonable, as it has been proved to be useful in few-shot class-incremental learning. Experiments are relative sufficient to show its efficacy.

Weaknesses

In general, this paper needs further polish, from my point of view. Just name a few, the abbreviations MRaM and MRAM are not consistent; the adjacency matrix A is incorrect; some sentences need careful attention. In addition, the motivation or the reason why the proposed method uses self-attention and MRAM is not clearly explained.

Questions

1.What are the meta-learning samples? There is no definition or explanation about them. It would be much better if more explanation would be given. 2.Why using Gaussian random projection to reduce the dimensionality, instead of others? 3.Why concatenating G_T and H_T^p? Could the authors give more explanation? 4.What is the difference between class prototype and class representation? It is very hard to tell from the paper.

Rating

5

Confidence

3

Soundness

2

Presentation

2

Contribution

2

Limitations

The authors have discussed the limitations in the paper.

Request for Detailed Accuracy Metrics to Evaluate Class Incremental Learning Performance

Dear Authors, Thank you for your thorough rebuttal, which has successfully addressed most of my concerns. However, I would like to raise an additional point regarding the evaluation of the few-shot class incremental learning approach presented in your paper. A well-recognized challenge in the field of class incremental learning is the model’s ability to maintain high accuracy for the initial session classes while often struggling with the classes introduced in subsequent sessions. The metrics you have provided give an overall picture of the model’s performance but do not sufficiently highlight this specific issue. To better assess the model’s capability in this regard, I kindly ask that you report the accuracy metrics in a more granular fashion. Specifically, I would like to see the accuracy for the classes of each session reported separately. For instance, at session 2, it would be beneficial to have the following: The accuracy for the classes introduced in sessions 0 and 1. The accuracy for the classes introduced in session 2. This division will allow us to more clearly evaluate the model’s performance on both the older and the new classes, providing a more comprehensive understanding of its incremental learning capabilities. Thank you for considering this request, and I look forward to your response.

Thanks for the Response

The authors response resolves my concerns, I decide to keep my rating.

The rebuttal has addressed part of concerns.

After seeing the response from the authors, part of my concerns has been addressed. Since the rest of the reviewers liked the paper and also considering my low confidence, I tend to raise my rating to borderline accept. But I do encourage the authors to better polish their paper, especially the Method section, to make it easier to follow.