Fairness without Sensitive attributes via Noise and Uncertain Predictions

First of all, we would like to thank you for all these inspiring reviews. All the ideas and hints are very much helpful for improvement of our work in this paper. Especially on the concerns of our motivations and the suggestions on the role of learnable noise. We appreciate the recognition of parts of this work from an expert. Thank you for your contribution to this community. Here are some of our thoughts according to your reviews:

1. In the first paragraph of Introduction, the evaluation to works that maximize the utility of the worst-case group is not precise. The Rawsian fairness (no sensitive attribute) based research leveraged side information (e.g., group ratio) to identify protected group, instead of relying on the correlation with observed features. Following this concern, DRO and ARL used as baselines in this paper are not very suitable, because they both highlighted that accuracy-related utility should be equal across different groups. Note that EO and DP are not criteria designed for these works, although you can conduct so.

1. Thank you for this correction. Indeed, DRO’s lower bound and upper bound are impacted by group proportion. In the new version, we change to “methods that focus on ensuring that accuracy-related utility is equal across various demographic groups.” DRO and ARL used as baselines in this paper are not very suitable
2. The reason we want to include these two baselines is because they are important methods to deal with fairness without sensitive attributes, though they are in a different branch with us. We self-identify our method as implicit method since we deal with fairness via representation (e.g. mitigating bias from input information). In the new version of paper, we add different marks to better indicate the different types of applied baselines in the table for clearance to our readers.

2. Given a threshold of 0.6 for splitting the training dataset and the resultant subsets have the distinct performance on EO and DP, which might be questionable. Recalling the definition of EO and DP, we see that they are both computed over predicted y. Since low-conf data tend to appear near to decision boundary, they certainly will yield small EO and DP values. Thus, letting low-conf data represent fairness is not very convincing for me. OOD data is NOT unbiased data.

Why showcase the values of EO and DP？The results presented in the paper regarding EO and DP values aim to provide readers with a more intuitive understanding of how the model tends to be misled by data with high differentiability. Our emphasis is not on highlighting fairness but rather on underscoring the severity of unfairness in the assessments of these data, considered easily differentiable. This becomes evident through a comparison of EO and DP values between these specific data points and the entire dataset. Building on this, our approach develops a framework to correct the model’s biases and alleviate unfairness. In the revised version of the paper, we have modified explanations for certain analytical results. Representing fairness with low-confidence data is inaccurate: Indeed, the values of EO and DP tend to decrease near the decision boundary. In the new version of paper, we will retain the EO and DP values from high-conf dataset and train_set, and present the number of samples comparison according to different small subsets e.g. # of white vs. # of black with positive labels to better demonstrate our motivations. Since the definition of EO and DP, presenting the number of samples also can help to connect fairness and confidence level.

Thank you for your time.

1. I have trouble understanding the high-level insights of the paper. The authors did not do a good job of presenting their method. e.g. why do they need VAEs? If the idea is to leverage low- and high-confidence data, why not just train two separate classifiers and then use ensemble? What is the purpose of adding learnable noise? I do not understand the author's explanation in 4.3.1. What is 𝛈 in Eq. (1)? What is the corresponding mathematical definition of "Pseudo-distribution" in Figure 1? Is it ℒL in Eq.(2)? In general, I think the technical part is poorly written, and would cause unnecessary confusion to readers.

Why not ensemble?: We actually considered this ensemble way in designing the proposed method, but it is not working in the way we want. It makes the prediction more unstable. Preserving the High-Conf’s weights, and adding Low-Conf’s weights, it is like a regulation on High-Conf’s weights, preventing it from getting biased on to those feature distribution patterns of the majority. Why vae?: We use vae for two main reasons: First reason is its effectiveness presenting through the experiment results and second reason is that we use the decoder’s output, which is the reconstructed input in the ablation study part. Observations through ablation study part help us to understand the reason of using learnable noise and pseudo-learning tricks together. We admit that there are other model can improve the performance, such as two-layer MLP, which we will also include the experiment results of using MLP in the new version of this paper. Please, have a look.

2. Can authors explain why the design can mitigate bias well when has no access to full sensitive attributes? The sensitive attribute S is rarely mentioned after the problem formulation in Section 3. How does the method connect to missing S ? I might miss it, but it shows the paper does not highlight how the method works.

1. To be frank, we are supervised by receiving these requirements of explaining the role of sensitive attributes. We clearly state that sensitive attributes are omitted during the training phase in Section 3 - ‘Problem Definition’, and we also provide motivation for this omission throughout the paper, even the title itself indicates that sensitive attributes are not used in the proposed framework. While it is disappointing to receive such comments, we have added a clear statement at the beginning of the last item in the summarised main points in the Introduction section. We hope that this new statement will help prevent any misunderstanding. 2. We explain the working flow in the overview part in Part 4 and also in the caption of Figure 2. Please, have a look.

3. The experiment on COMPAS does not seem to outperform other methods in an obvious way. The two fairness measures improve but the accuracy also drops. It would be clearer if the results could be presented in an accuracy vs. fairness Pareto frontier style plot.

Thank you for the suggestion. Yes, it is more clear to present an accuracy vs fairness plot in the paper to help our readers, and we will include it in the new version of paper

4. The evaluation is done only on two tabular datasets. This is rare in fairness literature. Can authors justify why it is only tested on two tabular datasets?

We admit more datasets can help readers to understand the permanence of the proposed framework. Hence, we will include new experiments using one new dataset, CelebA for better present our proposed method’s generalisation on different datasets.

Thank you for the reviews helping to improve our work. We appreciate your suggestions on better data analysis for the paper’s motivation and the interesting experiment on learnable noise for better illustration of its effectiveness.

1. (main concern) The proposed method is based on the observation that “when data is close to the decision boundary, non-sensitive information associated with those data tends to be similarly distributed across demographic groups, leading to lower accuracy but increased fairness”. However, it is unclear why and when it happens. For example, does this just happen to be due to the data distribution nature of the COMPAS dataset? The author should provide more explanations and discussions about this, theoretically or empirically, since this key property affects the scope of application of the proposed method.

Thank you for this suggestion. Yes the data analysis is conducted on one dataset, COMPAS. If we have a look on experiment results, the most improvement is in another dataset, New Adult. We argue that this is evidence of good generalisation of our proposed method. However, we appreciate this suggestion, hence we will add data analysis on New Adult in the new version of the paper.

2. It is also unclear that why the learnable noises need to be added in this paper. Is it to block information of sensitive information? In addition to the final classification results, I suggest that the author give some relevant analytical experiments to prove the validity of the learned noises. For example, if we train a classifier for sensitive attributes on the data with learned noises, is it difficult to predict sensitive attributes accurately?

Thank you for the suggestion. It is a very interesting idea of training a classifier for sensitive attributes. May you help us to interpret this experiment better? We will be appreciated if we know why this can indicate learnable noises. In our assumptions, we argue the learnable noise is a sort of data augmentation technique, really focusing on how to retain the prediction-only information. Hence, it may also help to predict sensitive attributes and have good prediction results compared to vanilla ML models such as logistic regression. In the beginning, we wanted to use learnable noise because of one simple idea. If we have multiple sensitive proxies in the dataset and want to find them all out, one naive way is mask each attribute and train a classifier to see the performance and the fairness evaluation. Currently we have an ablation study result that helps us to understand the effectiveness of learnable noise. We ask the decoder to generate reconstructed input and compare it with the original input. We find if we use learnable noise, although it is very different from the original input, it has the most accurate predictions in all the experiments. Hence we argue the learnable noise is necessary for this proposed framework for accurate predictions.

3. (main concern) It seems that the compared baselines are not strong enough. Please note that some works have gone beyond the baselines used in this paper such as CvaR DRO, LfF, JTT. Moreover, only two tabular datasets are used. I encourage the author to perform experiments on more datasets to illustrate the effectiveness of the proposed method.

Thank you for the suggestion. We will include new experiments using one new dataset, CelebA for better present our proposed method’s generalisation on different datasets. For the suggested baselines, we intend to just use DRO and ARL in the comparisons with our work since these two focusing on worst-group performance, ours is not. We include DRO and ARL because we don’t want to miss the two mile stones in this area. Other baselines are chosen from recent research about fairness without missing sensitive attributes/demographic which are representing each different branch of methods. In the new version of paper, we will include proxy method and fairness-accuracy trade-off works.