A Simple Remedy for Dataset Bias via Self-Influence: A Mislabeled Sample Perspective

We sincerely appreciate the reviewer’s time and effort in carefully considering our response. --- The reviewer expressed concerns that the current quantitative evaluation and experiments do not test for debiasing per se. **However, the quantitative evaluation approach we have presented aligns with the primary and standard evaluation convention in the spurious correlation (debiasing) domain, which assesses the degree of debiasing using an unbiased test set or by measuring average group accuracy and worst-group accuracy [1-11].** Evaluating on an unbiased test set effectively tests for debiasing per se, as it determines whether the model's predictions are driven by genuine task-related features rather than a specific malignant bias. For example, if a model is biased towards a particular attribute, it will likely perform well only on subgroups with that attribute, resulting in lower accuracy on the unbiased test set. **As mentioned in the evaluation protocols in Sec 5.1, we also adhered to the convention by using worst-group accuracy in Waterbird and bias-conflicting accuracy in BFFHQ as the evaluation metrics, further reinforcing the test for debiasing per se.** To the best of our knowledge, there is currently no qualitative evaluation method in the spurious correlation community that is considered more reliable than these quantitative evaluations. While methods like Grad-CAM to visualize activation map changes or t-SNE to examine how class instances cluster in latent space post-debiasing are sometimes used, these techniques are typically employed as supplements to quantitative results and are not considered more accurate than the quantitative evaluations that we, along with most previous works, have presented using an unbiased test set. Additionally, **we conducted experiments on five benchmark datasets and further tested on CIFAR-10C by adjusting the severity of bias to levels closer to unbiased conditions.** **This extensive evaluation not only covers a broader range of datasets and scenarios but also provides a more rigorous and comprehensive validation of our method compared to other existing works.** We once again sincerely thank the reviewer for the thoughtful feedback and the time dedicated to reviewing our work. We hope our responses have addressed the additional concern raised and welcome any further questions. [1] Nam et al. "Learning from failure: De-biasing classifier from biased classifier." NeurIPS, 2020. [2] Sagawa et al. "Distributionally robust neural networks.", ICLR, 2020. [3] Lee et al. "Learning debiased representation via disentangled feature augmentation." NeurIPS, 2021. [4] Liu et al. "Just Train Twice: Improving Group Robustness without Training Group Information." ICML, 2021. [5] Seo et al. "Unsupervised learning of debiased representations with pseudo-attributes." CVPR, 2022. [6] Hwang et al. "Selecmix: Debiased learning by contradicting-pair sampling." NeurIPS, 2022. [7] Park et al. "Training debiased subnetworks with contrastive weight pruning." CVPR, 2023. [8] Lim et al., "Biasadv: Bias-adversarial augmentation for model debiasing." CVPR, 2023. [9] Deng et al. "Robust Learning with Progressive Data Expansion Against Spurious Correlation." NeurIPS, 2023. [10] Ahn et al. "Mitigating dataset bias by using per-sample gradient." ICLR, 2023. [11] Jung et al. "Fighting fire with fire: Contrastive debiasing without bias-free data via generative bias-transformation", ICML, 2023.

Thank you for your comments, and we would like to provide a response to the reviewer's remaining concerns. --- ### [The clarity of definitions] ### **We carefully defined the terms in Sec. 3 using formulas following the conventions.** However, if any definitions remain unclear, we would be more than willing to clarify them further (e.g., by including additional images for bias-aligned and conflicting samples [3]) to improve readability. We would be grateful if you could specify which definitions are unclear, as your feedback will help us to further enhance the clarity and quality of our paper. --- ### [Explicit addressing of fairness metrics] ### The primary goal of our paper is to address spurious correlations in training datasets. Therefore, our primary evaluation focus is on assessing how well the trained model predicts on test datasets that are free from the spurious correlations present in the training set, using task-related features. **We have employed the standard and widely accepted benchmark datasets and evaluation metrics in spurious correlation studies [1-20]. It is important to note that the vast majority of spurious correlation studies[1-20] used evaluation methods the same as ours, such as unbiased accuracy and worst(or minority)-group accuracy, rather than the fairness metrics commonly used in the fairness domain.** However, we agree that incorporating fairness metrics would further enhance the evaluation section. Accordingly, we share experimental results evaluated using demographic parity (DP) and ~~equalized odds~~ equal opportunity (EOP) metrics on the Waterbird dataset. As shown in the table below, our approach demonstrates clear performance improvements, even when evaluated using fairness metrics. | Waterbird | DP | EOP | |:---:|:---:|:---:| | ERM | 0.1826 ± 0.0044 | 0.2731 ± 0.0187 | | SelecMix | 0.1146 ± 0.0004 | 0.1885 ± 0.0100 | | SelecMix+Ours | **0.0242 ± 0.0053** | **0.0099 ± 0.0064** | --- We appreciate the reviewer’s valuable feedback and the time spent reviewing our work. We hope that our responses have addressed the remaining concerns raised. --- [1] Wang et al. "Learning robust representations by projecting superficial statistics out." ICLR, 2019. [2] Bahng et al. "Learning de-biased representations with biased representations." ICML, 2020. [3] Nam et al. "Learning from failure: De-biasing classifier from biased classifier." NeurIPS, 2020. [4] Sagawa et al. "Distributionally robust neural networks.", ICLR, 2020. [5] Liu et al. "Just Train Twice: Improving Group Robustness without Training Group Information." ICML, 2021. [6] Kim et al. "Biaswap: Removing dataset bias with bias-tailored swapping augmentation." ICCV, 2021. [7] Lee et al. "Learning debiased representation via disentangled feature augmentation." NeurIPS, 2021. [8] Hong et al. "Unbiased classification through bias-contrastive and bias-balanced learning." NeurIPS, 2021. [9] Nam et al. "Spread spurious attribute: Improving worst-group accuracy with spurious attribute estimation."  ICLR, 2022. [10] Seo et al. "Unsupervised learning of debiased representations with pseudo-attributes." CVPR, 2022. [11] Idrissi et al. "Simple data balancing achieves competitive worst-group-accuracy." CLeaR, 2022. [12] Hwang et al. "Selecmix: Debiased learning by contradicting-pair sampling." NeurIPS, 2022. [13] Kim et al. "Learning debiased classifier with biased committee." NeurIPS, 2022. [14] Park et al. "Training debiased subnetworks with contrastive weight pruning." CVPR, 2023. [15] Lim et al., "Biasadv: Bias-adversarial augmentation for model debiasing." CVPR, 2023. [16] Deng et al. "Robust Learning with Progressive Data Expansion Against Spurious Correlation." NeurIPS, 2023. [17] Ahn et al. "Mitigating dataset bias by using per-sample gradient." ICLR, 2023. [18] Kirichenko et al.. "Last layer re-training is sufficient for robustness to spurious correlations." ICLR, 2023. [19] Liu et al. "Avoiding spurious correlations via logit correction." ICLR, 2023. [20] Jung et al. "Fighting fire with fire: Contrastive debiasing without bias-free data via generative bias-transformation", ICML, 2023. ---

By the way, I think Reviewer BCqo meant Equal Opportunity (EOP), not Equalized Odds (EO). These two metrics are slightly different. Please ensure the authors are using the correct metric for evaluation.

We thank you for the detailed review of our work and for your help in improving it further. --- ### [Q1.] What about fairness metrics (DP, EOP)? ### As the reviewer recommended, we additionally evaluate our method on Waterbird using fairness metrics such as DP and EOP. Note that we evaluate ours solely on Waterbird since CMNIST, CIFAR10-C, and NICO have more than two classes, and that BFFHQ's test set contains only bias-conflicting samples. As shown in the table below, our approach demonstrates clear performance improvements, even when evaluated using fairness metrics. | Waterbird | DP | EOP | |:---:|:---:|:---:| | ERM | 0.1826 ± 0.0044 | 0.2731 ± 0.0187 | | SelecMix | 0.1146 ± 0.0004 | 0.1885 ± 0.0100 | | SelecMix+Ours | **0.0242 ± 0.0053** | **0.0099 ± 0.0064** | --- ### [Q3.] The response only discussed the cases of 70% and 90%. However, in Table 2, ERM performs better than the proposed method in most settings (30%, 50%, 70%, and 90%). ### We apologize for the insufficient response. In the case of CIFAR-10C, as the bias severity decreases from 30% to 90%, the dataset gradually transitions into the low-bias domain, ultimately approaching an unbiased state at 90%. As mentioned in our previous response, this reduction in bias severity undermines the assumption that the bias is sufficiently malignant, resulting in reduced effectiveness of previous debiasing methods and allowing ERM to achieve better performance. In this context, to improve the performance of our method when applied to ERM—which leverages a large number of conflicting samples—it is necessary to increase the size of the pivotal set, thereby expanding the number of conflicting samples that our method can utilize. As demonstrated in the table below, expanding the pivotal set can lead to performance improvements even in low-bias settings, achieving state-of-the-art (SOTA) performance. Furthermore, if we had access to information regarding bias severity (i.e., the proportion of bias-conflicting samples), we could further optimize performance by adjusting the top-k value. | | CIFAR10C-30% | CIFAR10C-50% | CIFAR10C-70% | CIFAR10C-90% | |---|---|---|---|---| | ERM | 65.64 ± 0.51 | 71.33 ± 0.09 | 74.90 ± 0.25 | 76.03 ± 0.26 | | ERM+Ours (topk=100) | 65.61 ± 0.77 | 70.61 ± 0.62 | 73.20 ± 0.35 | 73.57 ± 0.16 | | ERM+Ours (topk=2000) | **71.25 ± 0.34** | **74.46 ± 0.34** | **75.84 ± 0.33** | **76.14 ± 0.23** | We appreciate the reviewer for highlighting this point, which has contributed to enhancing the rigor of our paper. We will include a discussion of these findings in the revision. ---

We appreciate the reviewer for the valuable feedback. We have provided the requested experiments below and hope our responses address the concerns raised. We welcome any further questions. --- ### [Q2.] The authors should evaluate on a few more datasets from the spurious correlation domain. ### As the reviewer requested, we conduct experiments on the MultiNLI dataset, and experiments on the CivilComments dataset are currently running due to time constraints. As shown in the table below, our method effectively works on a text dataset as well. We will include the complete table in the revision. | MultiNLI | Avg Acc. | Worst-group Acc. | |---|---|---| | JTT | 80.0 | 70.2 | | JTT+Ours | 79.8 | **73.6** | --- ### [Q3.] If compute allows, the authors should compute the ablations (Figure 6) for all other datasets. ### As the reviewer recommended, we conduct additional ablation studies on the Waterbird, BFFHQ, and CIFAR-10C (1%) datasets. The overall results are consistent with those presented in the main paper. Specifically, in the additional results, unbiased accuracy across varying $k$ values showed that performance was not sensitive to changes in $k$. Although $k$ = 100 was used in all our experiments, it was not always the optimal choice, as we lacked prior knowledge such as bias severity, and some other $k$ values yielded slightly better performance. | SelecMix+Ours | k=50 | k=100 | k=150 | k=200 | |---|---|---|---|---| | BFFHQ | **68.73 $\pm$ 0.79** | 65.80 $\pm$ 3.12 | 68.53 $\pm$ 1.16 | 67.87 $\pm$ 1.16 | | Waterbird | 88.37 $\pm$ 0.70 | 89.67 $\pm$ 0.38 | **89.72 $\pm$ 0.45** | 89.25 $\pm$ 0.32 | For$\lambda$, using $\lambda$ > 0 has shown to yield robust performance across both low-bias and high-bias datasets; however, setting $lambda$ too high can result in performance degradation. For the $\lambda$, on the highly biased dataset, increasing $\lambda$ results in performance degradation, while on the low-bias dataset, $\lambda=0$ causes a significant performance drop. | SelecMix+Ours | lambda=0 | lambda=0.1 | lambda=0.2 | lambda=0.5 | |---|---|---|---|---| | BFFHQ | **68.67 $\pm$ 1.00** | 65.80 $\pm$ 3.12 | 68.53 $\pm$ 1.23 | 67.47 $\pm$ 1.23 | | Waterbird | 78.92 $\pm$ 4.16 | **89.67 $\pm$ 0.38** | 88.01 $\pm$ 0.48 | 85.72 $\pm$ 0.49 | For epochs, aside from the extremely short training with epoch = 1, we observe that performance is not sensitive to the number of epochs. | SelecMix+Ours | epoch=1 | epoch=3 | epoch=5 | epoch=7 | epoch=9 | epoch=11 | |---|---|---|---|---|---|---| | CIFAR10C-1% | 43.76 $\pm$ 0.67 | 44.79 $\pm$ 0.40 | **46.18 $\pm$ 0.33** | 45.43 $\pm$ 0.61 | 45.20 $\pm$ 0.61 | 45.69 $\pm$ 0.08 | | Waterbird | 87.90 $\pm$ 0.32 | 89.56 $\pm$ 0.09 | 89.67 $\pm$ 0.38 | 89.98 $\pm$ 0.21 | **90.03 $\pm$ 0.39** | 89.15 $\pm$ 0.63 | --- ### [Q9.] Why were some baselines omitted from Tables 2 and 3? ### We conduct evaluations of DCWP on the benchmark datasets presented in Tables 2 and 3. As shown in the table below, our method still outperforms DCWP. (Please note that "Ours (best)" refers to the highest performance achieved in Tables 2 and 3.) | | CIFAR10C-20% | CIFAR10C-30% | CIFAR10C-50% | CIFAR10C-70% | CIFAF10C-90% | Waterbird | NICO | |---|---|---|---|---|---|---|---| | DCWP | 63.37 $\pm$ 1.01 | 67.31 $\pm$ 0.54 | 69.61 $\pm$ 0.21 | 71.54 $\pm$ 0.10 | 71.85 $\pm$ 0.08 | 73.31 $\pm$ 1.78 | 44.98 $\pm$ 1.59 | | Ours (best) | **66.67 $\pm$ 0.43** | **68.13 $\pm$ 0.45** | **72.79 $\pm$ 0.38** | **73.56 $\pm$ 0.15** | **73.57 $\pm$ 0.16** | **89.67 $\pm$ 0.38** | **45.69 $\pm$ 1.12** | ---

We are grateful for your insightful feedback and the thoughtful suggestions that have significantly enhanced our work. --- ### [Q2.] The authors should evaluate on a few more datasets from the spurious correlation domain. ### We carried out further experiments on the CivilComments-WILDS dataset. As presented in the table below, the results demonstrate that our method performs well on CivilComments-WILDS. Consequently, our method has shown effective performance across CelebA, MultiNLI, and CivilComments, highlighting its effectiveness on text datasets. We will ensure that the complete table is included in the revision. | CivilComments | Avg Acc. | Worst-group Acc. | |---|---|---| | JTT | 92.6 | 63.7 | | JTT+Ours | 86.9 | **78.5** | ---

### I would encourage the authors to characterize (either theoretically, or empirically with synthetic noise) the behavior of their method under mislabeling, which is an important potential failure mode ### We sincerely appreciate the reviewer’s valuable suggestion regarding mislabeled samples, which is indeed an important issue in real-world scenarios. However, we would like to emphasize that the primary focus of our paper, as well as prior studies on spurious correlations, is to address spurious correlations within training datasets. While considering mislabeled samples in conjunction with spurious correlations is an important future direction, please note that the debiasing community is still facing significant challenges in effectively addressing spurious correlations alone. In fact, addressing both problems together introduces a new and complex challenge in this field. Therefore, we respectfully ask that if the reviewer’s last concern involves addressing the mislabeling problem simultaneously, we would greatly appreciate reconsideration of this given context. Once again, we sincerely thank you for your thorough review of our paper and for the valuable feedback provided. ---

Thanks for the author's detailed response to each of my questions and concerns. The use of the self-influence function to identify bias-conflicting samples is well-motivated. I think the fairness evaluation for this work is more align with Rawlsian Max-Min fairness, and I really appreciate the supplementary experiments that demonstrate this. However, I still believe this paper needs further improvements regarding the clarity of definitions, explicit addressing of fairness metrics, and the novelty of the work, so I keep the current rating score.