One to All: Individual Reweighting for User-Oriented Fairness in Recommender Systems

• The presentation of Section 4.1 and 4.2 could be improved. Authors only focus on how to presenting their framework while ignoring the motivation behind it. For example, it is hard to understand the intuition of Eq4 and the reason authors chose this formulation.

• The individual re-weighting methods to ensure UOF is not very novel. For example, a naive solution is to set the weight for each user as their individual popularity (levels of activity). Authors should discuss why this naive solution is not good. Additional, some existing papers, such as "Mitigating Popularity Bias for Users and Items with Fairness-centric Adaptive Recommendation" (https://doi.org/10.1145/3564286) also tackle the UOF issue from a individual perspective.

1. The paper's novelty is limited, as it merely assigns specific loss weights to each user to direct the model's focus toward disadvantaged groups. While this method may improve the prediction accuracy for some disadvantaged users, the fundamental issue remains that their low accuracy is primarily due to fewer interaction records. In my view, increasing weights does not address the root cause of User-Oriented Fairness (UOF), thus failing to resolve Limitation 1 identified by the authors.
2. The group fairness proposed by the authors is similar to group Distributionally Robust Optimization (DRO) in machine learning [1], so it is suggested to include related work on DRO and to add comparisons with some DRO models in the experiments.
3. There are some confusing aspects in the paper, such as in line 249, where the maximum loss weight is assigned a value of 1. Additionally, the authors do not explain why there may be individuals in the advantaged group who could also be overlooked by the model, given that these individuals have sufficient interaction records.
4. I agree that reweighting individuals in the advantaged group can improve their recommendation accuracy, but this may further exacerbate the differences between the two groups, leading to greater unfairness. However, based on the ablation study results presented in the paper (Table 2), reweighting the entire dominant group simultaneously improves both accuracy and fairness, which is a point of confusion for me.
5. My final concern is regarding the hyperparameter. As the authors mention in the paper (Line 257,500), this is inconsistent with the results shown in Figure 5. For example, when the NDCG in Figure 5 is approximately 0.32, while in Table 1, the NDCG is 0.483. This suggests that the conclusions drawn by the authors may be incorrect (When reaches a sufficiently large value, all samples will be assigned equal weights). [1] Sagawa, Shiori, et al. "Distributionally Robust Neural Networks." International Conference on Learning Representations.

(W1) The novelty of a proposed method is unclear.

The proposed objective in Eq. (9) is a trivial combination of group-wise and individual reweighting. Although the authors claim "The key motivation behind IR-UOF is to introduce an in-processing framework that employs an individual level optimization strategy" in line 213, there is no discussion on existing individual reweighting methods. For example, Shivaswamy and Garcia-Garcia [b] have proposed an individual reweighting strategy for recommender systems based on adversarial training [a]. They introduced adversarial training to ensure user-oriented fairness without demographics (e.g., user group or similarity). In addition, the formulation in Eq. (5) is quite similar to DRO. Togashi et al. [c] have proposed CVaR optimization (a dual of DRO) and discussed smoothing for adversarial individual weights, which is closely related to the regularization technique in Section 4.2 of the current manuscript.

[a] Lahoti, Preethi, et al. "Fairness without demographics through adversarially reweighted learning." Advances in neural information processing systems 33 (2020): 728-740.

[b] Shivaswamy, Pannaga, and Dario Garcia-Garcia. "Adversary or friend? an adversarial approach to improving recommender systems." Proceedings of the 16th ACM Conference on Recommender Systems. 2022.

[c] Togashi, Riku, et al. "Safe Collaborative Filtering." The Twelfth International Conference on Learning Representations. 2024.

(W2) The connection between the UOF metric in Eq. (2) and the proposed objective is unclear.

The authors introduced a UOF metric in Eq. (2) and then proposed to optimize Eq. (9). However, the connection between Eq. (2) and Eq. (9) is non-trivial. In the extreme case without any regularization proposed in Eq. (6), the sum of maximum loss values in advantaged/disadvantaged groups is minimized. Since loss values should negatively correlate with metric values, the sum of minimum metric values in each user group is maximized.
In my opinion, there is no trivial reason to believe the difference of means (UOF) is minimized by maximizing minimums. It would be helpful to provide a more rigorous theoretical analysis or justification for how optimizing Eq. (9) leads to minimizing the UOF metric in Eq. (2).

Limited Novelty: The approach of aligning features across different modalities is not particularly innovative, given existing literature on multimodal fusion.

Lack of Addressing Latent Space Differences: The authors fail to consider the significant differences in latent spaces between modalities, which could affect the model’s performance.

Outdated Baselines: The choice of baselines for comparison lacks relevance, as they are not based on current sequential recommendation methods, weakening the impact of the findings.

Weak Motivation: The paper does not adequately emphasize the significance of the problem it addresses, lacking persuasive preliminary evidence.

Insufficient Insights: The limited experimental exploration reduces the potential contributions of the work to the community.