Empowering Collaborative Filtering with Principled Adversarial Contrastive Loss

Summary

In this work, the authors focus on improving the generalization ability of the top-K recommendation model by a proposed principled Adversarial InfoNCE loss (AdvInfoNCE). Existing contrastive learning based methods usually lack considering the tailored inductive bias (such as hard negatives and false negatives) and sufficient theoretical understanding for the generalization ability. The proposed AdvInfoNCE loss could adaptively explore and assign hardness (weight) for each negative instance in an adversarial fashion. The theoretical guarantees and experiments demonstrate the effectiveness of the proposed AdvInfoNCE loss.

Strengths

1. The motivation that adaptively assigns hardness for each negative instance in an adversarial fashion to improve the generalization ability is justified. 2. The theoretical proof demonstrates the generalization ability of the proposed AdvInfoNCE loss. 3. The experiments on unbiased datasets and out-of-distribution settings demonstrate the effectiveness of the proposed AdvInfoNCE loss in terms of generalization ability.

Weaknesses

1. It would be better to add an intuitive example to show the adaptive learning process for hard negatives and false negatives in the min and max stages. 2. In terms of adaptively learning hardness for each negative instance, it would be better to add some baselines focusing on mining hard negatives or false negatives. 3. In terms of the generalization ability, it would be better to add some baselines focusing on out-of-distribution. Besides, OOD experiments are also performed on popularity-based distribution shift scenarios, what’s the difference compared with debiasing experiments?

Questions

1. Could you please give an intuitive description to show the adaptive learning process for hard negatives and false negatives in the min and max stage ? Especially for hard negatives and false negatives? 2. Why the corresponding item is identified as a hard negative rather than a general negative when sigma > 0 ? 3. In terms of adaptively learning hardness for each negative instance, could you please add some baselines focusing on mining hard negatives or false negatives [1,2] ? 4. In terms of the generalization ability, could you please add some baselines focusing on out-of-distribution [3,4] ? [1] Chen, Jin, et al. "Cache-Augmented Inbatch Importance Resampling for Training Recommender Retriever." In NIPS, 2022. [2] Chen, Jin, et al. "Learning Recommenders for Implicit Feedback with Importance Resampling." In WWW, 2022. [3] Hongyi Wen, et al. "Distributionally-robust Recommendations for Improving Worst-case User Experience". In WWW, 2022. [4] An Zhang, et al. "Invariant Collaborative Filtering to Popularity Distribution Shift". In WWW, 2023.

Rating

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

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors adequately point out the limitations and there is no negative societal impact of their work.

Summary

This paper is on collaborative filtering (CF) enhanced by contrastive learning (CL). The authors point out that the adoption of CL into CF is suboptimal due to challenges such as the issue of out-of-distribution, the risk of false negatives, and the nature of top-K evaluation. They also note that current CL-based CF methods lack consideration of the tailored inductive bias for CF and have limited theoretical understanding of their generalization ability. To address these limitations, the authors propose a principled Adversarial InfoNCE loss for CF that focuses on mining hard negatives and distinguishing false negatives from the vast unlabeled user-item interactions. The proposed method is compared with several state-of-the-art contrastive learning-based CF methods on both unbiased and synthetic datasets. The experiments show that the proposed method outperforms the baselines in terms of accuracy and robustness, demonstrating its potential for improving the performance of recommender systems.

Strengths

i) The paper proposes a novel approach to adaptive contrastive learning in collaborative filtering by adopting an adversarial approach. The proposed Adversarial InfoNCE loss addresses the limitations of existing methods and allows for the fine-grained assignment of hardness to each negative user-item pair, which enhances the recommender's generalization ability and empowers top-K recommendations via informative negative sampling. ii) The paper provides innovative theoretical insights into the benefits of adversarial hardness learning. It shows that the proposed hardness scores are correlated with the out-of-distribution problem of recommendation, and can thereby enhance the generalization ability of recommenders. iii) The study of hardness gives an in-depth analysis on the learned hardness scores, which uncovers the importance of learning correct hardness for the massive negative samples without observations.

Weaknesses

i) The proposed method can be viewed as an adaptive SSL method for recommendation [1-2]. Also, it can be a learnable negative sampling approach [3-4]. Literature review (baseline comparison would do better) should be done for these two very relevant research line. ii) Though a brief training cost experiment is given in the appendix. I would expect more detailed efficiency experiments or analysis to support the claim that the proposed AdvInfoNCE can serve as a foundation loss for future CF researches, as the proposed approach utilizes the adversarial training method. [1] Graph contrastive learning with adaptive augmentation [2] Automated Self-Supervised Learning for Recommendation [3] Personalized Ranking with Importance Sampling [4] AHP: Learning to Negative Sample for Hyperedge Prediction

Questions

It would be better for the authors to clarify the relations between the proposed method and the related research lines I mentioned above. And I would expect the authors to give a more detailed efficiency analysis or experiments.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and 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

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

N/A

Summary

This paper studies contrastive learning (CL) in collaborative filtering (CF) for top-k recommendation. In particular, it focuses on the CF-tailored challenges for CL, and then presents adversarial infoNCE (AdvInfoNCE) loss. This loss dynamically assigns hardness to negative instances and incorporates a fine-grained ranking criterion to improve the CF recommender’s generalization ability. Furthermore, this paper highlights theoretical properties of AdvInfoNCE. Experiments on both synthetic and real-world datasets are done to show the effectiveness of this loss, especially in out-of-distribution scenarios.

Strengths

1. The motivation of considering fine-grained hardness-aware ranking criteria is clear and reasonable. Technically, it is insightful and novel to transfer the distinguish between false and hard negatives into a margin learning problem. 2. As proved in Sec 3.3, the adversarial training framework of AdvInfoNCE is natural and somehow equivalent to solve a distributionally robust optimization (DRO) problem. I appreciate this theoretical guarantee that can endow a AdvInfoNCE-trained CF models with better generalization ability. 3. The experiments are done on four datasets with two CF base models, which are sufficient to demonstrate the effectiveness of the proposed loss. Moreover, the selected baselines are quite new, including some recent and strong works. 4. This proposed loss seems to be a general loss that can be applied in general recommendation models.

Weaknesses

1. The hardness is denoted as $\delta$ in Sec 3.2, while related to $p(j|(u,i))$. This is not explained clearly. More clarifications are needed here. 2. In Line 214, the limitation is stated as the ‘training instability’, which is not empirically shown in the experiments, such as indicated by ‘training loss variance’. It would be better to discuss more about this instability. 3. Although the proof of Theorem 3.1 seems correct, the DL-divergence In Line 195 misses a minus sign. Please double check and fix it.

Questions

1. The hardness is denoted as $\delta_{j}^{(u)}$ first and then $\delta_{j}^{(u,i)}$, which seems confusing. Is it related to a specific user $u$ or a user-item pair $(u,i)$?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and 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

4 excellent

Limitations

Please refer to the weaknesses and questions.

Summary

Current losses for collaborative filtering struggle to handle the issue of unobserved user-item pairs. Typically, the approach is to treat unseen pairs as negatives while seen pairs as positives, but this is somewhat problematic because unseen pairs could just be unobserved positives. The authors propose an adversarial InfoNCE-based loss that claims to address this problem in collaborative filtering. This loss works by minimizing the InfoNCE loss given that we have adversarially learned weights for the negative samples. They give a theorem that shows this proposed loss can be interpreted as a distributionally robust optimization problem. Finally, they give some empirical results showing the efficacy of their method over other CF baselines.

Strengths

1. This paper provides a method that appears to give solid gains across various collaborative filtering tasks. 2. They do make an attempt to try to interpret what the \delta (adversarially learned parameters) in their method are doing. 3. The graphs for the results and the pictures explaining the methods are good and helped me with understanding. 4. The authors try to tackle a hard problem: it is hard to think about how to best utilize unseen pairs in collaborative filtering due to their unseen nature. 5. The method is fairly novel as a nontrivial extension to InfoNCE to the collaborative filtering setting via adversarial training.

Weaknesses

1. I don't understand the role of the adversarial variables in the algorithm. In figure 2a there was some attempt at interpreting the values of the deltas, but it still does not make sense. I hope the authors can explain the role of the variables better. 2. I think the terminology of "hard negative" is confusing, because typically in the self-supervised learning literature people call negatives that are near the decision boundary "hard negatives". However, in this paper hard negatives are the opposite: negatives that are far away from the positives pair. I suggest rewriting the paper to make the message more clear. 3. In general, the paper is hard to understand and has many grammatical errors. The authors should fix this to make the paper easier to read. 4. From what I can understand, the loss should simply make the deltas as large as possible (positive) to increase the loss value given that there are no constraints (aside from the number of epochs trained, I guess). I have concerns about the usefulness and stability of this algorithm. 5. In theorem 3.1, we assume that the deltas imply a probability distribution. Is this true? As we train do the deltas for a user i add up to |N_i|? I'm not sure there is a constraint there enforcing this. In that case I'm not sure the theory applies to the algorithm as-is.

Questions

My questions are in 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

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Seems to be sufficient.

Summary

This paper proposes a principled AdvInfoNCE loss for CF methods to improve generalization ability. It utilizes a fine-grained hardness-aware ranking criterion to assign weights for unobserved user-item interactions. In this way, it can enable better distinction between different negative, thus mitigating the inductive bias in CF-based methods. It provides theoretical proof of the effectiveness of AdvInfoNCE loss and the experimental results compared with other popular loss used in recommenders look promising.

Strengths

It is well-written and easy to follow. Good motivation of improving the generalizability of CF-based methods. It provides thereotical guarantees for the loss design and conducts comprehensive analysis on the effectiveness of its method. Experimental results compared with other popular functions adopted in CF models look promising. Code is open.

Weaknesses

Experiments can be more extensive. The results on MF and LightGCN look promising. But I think it would be more convincing if the authors can consider more CF-based backbones like MultVAE [1] and DGCF [2]. [1] Liang et al. Variational Autoencoders for Collaborative Filtering. 2018 WWW. [2] Wang et al. Disentangled Graph Collaborative Filtering. 2020 SIGIR.

Questions

This paper proposes a novel constrastive loss for fine-grained ranking of negative samples to improve the generalizability of CF-based models, which is a significant contribution. It provides comprehensive theoretical analysis and experiments look promising. I just have minor concerns about the technical details. - Do you attempt other similarty measurements for calculating the hardness and whether the choice of similarity calculation matters? - What is the rate of negative samples over all the unobserved interactions used in your method and is there any result of using different negative sampling rate?

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

More CF-based backbones can be considered.

Thanks!

We appreciate your acknowledgment of our efforts in addressing the concerns. Your insightful comments have been instrumental in enhancing the quality of our work.