acceptodds
Withdrawn from ICLR 2027

SAFE-Stream: Fairness-Aware Learning in Streaming Data without Sensitive Attributes via Latent Group Discovery

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Fairness-aware learning in streaming environments is challenging. As data distributions change over time, predictive performance may degrade and demographic disparities may increase. This challenge becomes harder when sensitive attributes are not available for continuous supervision. Most existing fairness methods assume static data or access to sensitive attributes, which limits their use in real-world streaming settings. In this work, we propose SAFE-Stream, an online fairness-aware learning framework that discovers latent groups and adapts to both performance and fairness drift with limited supervision. The first step is an encoder based on multilayer perceptrons that maps non-sensitive features into a latent space. Next, the Latent Group Discovery Mechanism (LGDM) finds dynamic groups by simultaneously considering the similarity of representations and the dissimilarity of prediction losses. These latent groups are then exploited by an online fairness to track Demographic Parity and Equal Opportunity over a sliding window. Further, two ADWIN-based detectors are used to monitor performance and fairness drift. When either type of drift is detected, SAFE-Stream adaptively reweights the training samples to update the online predictor and recover its performance and fairness. We evaluate SAFE-Stream on the UCI Adult Income dataset using multiple latent-group configurations and 5-fold cross-validation. To connect the discovered groups to real demographic attributes while keeping labeling costs low, only 2% of the streaming samples are manually checked for sex and race. Across the configurations, SAFE-Stream achieves competitive predictive performance while consistently reducing Demographic Parity Disparity for both sex and race compared to the true baseline. It also significantly reduces the Race Equal Opportunity Disparity. This demonstrates that latent groups can support meaningful fairness monitoring without the supervision of sensitive attributes.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Fairness-aware machine learning, streaming learning, sensitive-attribute, Demographic parity, Equal opportunity

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_5UanOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

To address fairness issues in streaming environments where sensitive attributes may be unavailable, this work proposes SAFE-Stream, a fairness-aware streaming framework based on latent group assignments. An empirical evaluation is conducted on the Adult dataset using a logistic regression algorithm.

Critical strengths
  1. Fairness is a critical and urgent problem in contemporary AI applications, particularly with respect to human demographic attributes.

  2. The research motivation is well justified for scenarios involving streaming data and unavailable sensitive attributes.

Critical weaknesses
  1. The exposition and structure of the paper require further refinement, especially in the Method section, many experimental settings should not be placed in the Method section.

  2. The proposed method is difficult to understand. The rationale for its design and the reasons for its expected effectiveness are insufficiently explained.

  3. The proposed method introduces too many hyperparameters, and the Method section only reports fixed settings without providing empirical or theoretical justification for their effectiveness or reasonableness.

  4. The experimental results in Table 2 do not demonstrate significant advantages over the baseline.

  5. The scope of the experiments is narrow: only one dataset, one model, and one baseline are used, with a single execution using seed 42. Consequently, the results lack statistical validity.

Questions for authors

See Weaknesses

Reviewer_K88AOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose an attribute-blind learning algorithm that seeks to control an online version of approximate DP and EOpp fairness. They do so by learning and updating latent group memberships from the streamed covariates without directly predicting an attribute group, and implementing a trigger update rule on the current batch of covariates if a fairness metric is violated. They entirely focus on the UCI adult benchmark dataset.

Critical strengths

This is an interesting problem to study, where we assume the number of groups is not known in addition to not knowing the protected group memberships. Their approach of using disparities of latent groups as a proxy to the true group disparities is novel, to the best of my knowledge.

Critical weaknesses

The scope of the method appears to be limited. The MLP network used to get embeddings of the covariates is not trained, so that control over the latent group disparities does not immediately imply control over the target, true group disparities. Similarly, it does not seem SAFE-Stream targets reductions in disparities when a fairness violation is triggered (Sec 4.5) -- the method further trains on the current batch that triggered the fairness violation, with no guarantees on how it will affect latent or true group disparities.

The empirical evidence seems limited as well. The single running experiment of the paper (UCI adult dataset) shows their method has difficulty meaningfully reducing the DP and EOpp to the trigger level of 0.05, even when K is close to the actual number of protected groups (Table 2).

Questions for authors

Can one attempt to learn latent groups via training a MLP (rather than using a randomized initialization)? What if we can not update models or predictions for large batches? (e.g., can only change the predictions on a candidate to candidate basis, as in, e.g., Cano et al. 2025)
It would be helpful to see how the true and latent group DP and EOpp change as a function of time, as oppose to seeing only the final latent group fairness metrics (Table 3). Can the authors include experiments on more datasets?

Reviewer_9FdxOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies fairness monitoring and adaptation in a streaming setting where sensitive attributes are not available to the learner. The authors present the framework SAFE-Stream, which combines latent group discovery using non-sensitive features, sliding-window fairness monitoring based on Demographic Parity (DP) and Equal Opportunity (EO), drift detection using ADWIN, and adaptive model updates triggered by detected fairness violations or performance changes.

The authors evaluate SAFE-Stream on the UCI Adult dataset using six latent-group configurations and five-fold cross-validation. Compared with an offline logistic regression baseline using sensitive attributes, SAFE-Stream achieves slightly lower predictive accuracy while reporting lower average DP gaps for sex and race and lower EO gaps for race. However, sex-based EO deteriorates, and the observed fairness differences are relatively small compared with their variability. A sparse post-hoc audit further suggests that discovered latent groups correlate more strongly with sex than race.

Overall, the paper explores the feasibility of using latent groups for fairness monitoring without sensitive attributes, but provides limited evidence that the proposed framework effectively mitigates demographic disparities or responds to fairness drift.

Critical strengths
  1. The studied challenge is interesting. This paper studies an important practical challenge: monitoring and mitigating fairness disparities in continuously evolving environments without direct access to protected demographic information. Combining these two constraints creates a meaningful research setting that is less extensively explored than conventional offline fairness learning.

  2. The problem formulation is clear, and the experimental report looks honest. The paper clearly separates the information available to the learning, monitoring, auditing, and evaluation components. The authors self-report degeneracies (Observations 1–3), confounds, and missing logs. They also explicitly acknowledge some limitations, including that the loss-based group signal is uniform, that adaptive updates are effectively batch-level, and that latent-group fairness should not be equated with demographic fairness.

Critical weaknesses
  1. The experiments cannot support the design of the proposed framework. To be more specific, the fairness components do not do what the framework claims.

Observation 1 shows the loss-based signal is identical for all groups. LGDM therefore reduces to soft nearest-center assignment on a fixed random embedding, plus a constant. Observation 2 shows the monitored latent gaps are shrunk toward the window mean by construction. This weakens the meaning of the ϵ\epsilon threshold. Observation 3 shows the reweighting gives every sample in a triggered batch the same repetition count. It cannot shift the predictor toward any group but only changes how many SGD steps a batch receives. Without a group-differentiating update, there is no mechanism by which latent-group monitoring could reduce demographic disparity.

  1. The experimental evaluation does not establish that SAFE-Stream improves fairness or effectively handles streaming drift.

The evaluation is restricted to the UCI Adult dataset, which does not contain an explicit temporal structure, and no controlled distributional or fairness drift is introduced. In addition, the only comparison is against an offline logistic regression model that includes sensitive attributes. Since the baseline differs from SAFE-Stream in multiple dimensions, observed fairness differences cannot be attributed to the proposed method. Essential comparisons against standard online learning without fairness adaptation and ablations removing latent-group monitoring or adaptive updates are missing. Finally, the observed fairness improvements are relatively small or highly variable, while sex-based EO becomes worse. Without stronger controls or statistical evidence, the results do not convincingly demonstrate a meaningful fairness benefit.

Questions for authors
  1. How does SAFE-Stream compare against an online logistic regression model trained on the same non-sensitive features, with identical preprocessing and training budgets but without latent-group monitoring or fairness-triggered adaptation? Ablations of latent-group discovery, drift detection, and adaptive updates would be particularly important for determining whether these components contribute meaningful improvements.

  2. Given that the loss-based group signal remains uniform and adaptive reweighting applies identical additional updates to every sample in a triggered batch, what mechanism specifically encourages reductions in demographic disparities? Can you demonstrate that the proposed adaptation produces fairness improvements beyond those achieved by ordinary repeated SGD updates?

  3. Can you evaluate the framework under controlled distributional or demographic fairness shifts and report whether changes in latent-group DP/EO track changes in true demographic DP/EO? Such evidence appears necessary to establish the central claim that latent groups provide a useful operational structure for fairness monitoring in streaming environments.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.