Representation Bias in Data: A Survey on Identification and Resolution Techniques

1 INTRODUCTION

• A system may perform well in accuracy overall yet fail on less-populated regions with insufficient representation, which may matter because such regions frequently represent minority subpopulations in society.Source: §1 · pp. 1–2
• Representation bias occurs when training data underrepresents parts of the target population and can arise during collection or after it.The introduction also identifies selection bias, including sampling that reaches only part of a population or a mismatch between training and target populations.Source: §1 · p. 2
• A group rare in the underlying distribution may remain underrepresented even with random sampling.The introduction gives pregnant people as an example of a minority group that may receive fewer data points, including within age-based sampling.Source: §1 · p. 2
• Data scientists often lack control over collection and therefore use found data without guarantees that it is a representative sample.The introduction describes systematic collection as identifying subpopulations and ensuring sufficient samples from each.Source: §1 · p. 2
• The survey focuses on identifying and resolving representation bias as a property of datasets, regardless of their eventual use.It distinguishes this focus from work primarily concerned with fairness in machine-learning models.Source: §1 · p. 2
• It proposes a taxonomy for identification and resolution approaches, examines structured-data methods by objectives and capabilities, and reviews unstructured data including images, text, speech, and graphs.The survey also discusses less-noticed aspects of the literature as possible research directions.Source: §1 · pp. 2–3

2 AN OVERVIEW OF FAIRNESS LITERATURE

• Fairness has drawn substantial attention in ML, with surveys and tutorials preceding this survey’s focus on representation bias in data.The overview frames definitions and fairness techniques as context for specifying the survey’s scope.Source: §2 · p. 3

2.1 Definitions of Fairness

• Fairness has no clear agreed definition because it depends on the task being targeted, with perspectives including individual, group, and subgroup fairness.Subgroup fairness measures fairness when groups are defined by intersections of multiple sensitive-attribute values.Source: §2.1 · pp. 3–4
• Group fairness measures are classified as independence, separation, sufficiency, or causation, covering sensitive-attribute independence, outcome-conditioned fairness, prediction-conditioned parity, and counterfactual fairness [11, 19, 43, 79, 105, 143].The overview notes that these examples are not an exhaustive list of group fairness definitions.Source: §2.1 · pp. 3–4

2.2 Interventions to Achieve Fairness

• Pre-processing modifies data before model training through massaging, reweighting, sampling, feature representations, adversarial learning, and causal methods.Reweighting targets a discrimination-free distribution, while causal approaches examine dependencies between sensitive attributes and proxy attributes.Source: §2.2 · pp. 4–5
• In-processing promotes fairness by adding constraints or regularization to learning objectives, with other approaches including adversarial learning, re-weighing, and bandits [10, 21, 32, 74, 94, 165, 169].Constraints aim to ensure equal treatment of subpopulations under specified fairness measures.Source: §2.2 · pp. 5–6
• In-process fairness methods include penalizing predictable sensitive attributes, multi-constraint optimization, sample re-weighting, and bandit approaches that recognize fairness when observed.In-process re-weighting learns an unweighted classifier before using learned sample weights to retrain it.Source: §2.2 · p. 6
• Post-processing methods alter classifier outputs: Hardt et al. target equalized odds, while Pleiss et al. seek calibration and error-rate feasibility through label-flipping probabilities.Source: §2.2 · p. 6
• Reject option classification by Kamiran et al. uses posterior probabilities to label instances from deprived and favored groups.Source: §2.2 · p. 6

2.3 Scope of the Survey

• The survey studies representation bias as a data-set feature, independent of later use, across structured and unstructured data including images, graphs, text, and speech.Source: §2.3 · p. 6

2.4 Related Surveys and Tutorials

• The authors present this as the first survey focused on identifying and resolving representation bias across structured and unstructured data from a data-centric standpoint.Source: §2.4 · p. 6
• Unlike Stoyanovich et al. and Catania et al., this survey focuses on data-centric identification and resolution of representation bias across data types.Source: §2.4 · p. 7

3 AN OVERVIEW OF REPRESENTATION BIAS

• Representative data must include enough examples from less-popular regions, since underrepresented groups can face discriminatory outcomes in automated and machine-learning decisions.Matching the production distribution alone is insufficient; the Boston school-assignment example illustrates how overlooking the scarcity of top schools in underprivileged districts can cause problems.Source: §3 · pp. 7–8

3.1 Reasons for Representation Bias

• The survey groups causes into historical bias, skew in the underlying distribution, and selection, sampling, or self-selection processes, noting no consensus on an exhaustive taxonomy.Historical bias can reflect existing social conditions, while selection affects who is chosen, sampling affects who is likely to be included, and self-selection affects who participates.Source: §3.1 · pp. 8–9

3.2 Measuring Representation Bias

• The survey reviews proposed measures for evaluating representation bias in data.Representation rate is introduced as a metric for identifying bias with respect to base rates.Source: §3.2 · p. 9, §3.2.1 · p. 9

3.2.1 Representation Rate.

• Representation rate measures subgroup representation against base rates, with equal representation defined as equal numbers of objects from each subgroup.The passage notes that this condition may be difficult to achieve because real data sets rarely contain nearly equal numbers across all subgroups.Source: §3.2.1 · p. 9

3.2.2 Data Coverage.

• Data coverage measures whether each data-set object has enough similar entries, providing another way to assess representation bias.The survey introduces a generalized coverage definition over attributes of interest, such as gender, race, or salary.Source: §3.2.2 · pp. 9–10
• Data coverage requires sufficient representation of every meaningful subpopulation; thresholds are application-specific, and human review should exclude semantically invalid groups.The passages emphasize that coverage thresholds are inputs determined through statistical analyses and vary by context.Source: §3.2.2 · p. 10
• Representation rate gives stronger guarantees against downstream representation bias than coverage, but is more restrictive and harder to achieve, especially with skewed underlying distributions.Following the underlying distribution may then be incompatible with fully satisfying representation-rate requirements.Source: §3.2.3 · p. 10

3.3 Representation Bias Harms

• Figure 2 illustrates classifier performance on the task of determining whether a query point lies inside or outside the cat body across different under-represented regions.Source: Figure 2 · p. 11
• Representation bias is most problematic where a model cannot interpolate adequately; it may matter less far from a classification boundary or where regression targets fluctuate little.The section cautions that representation bias does not necessarily imply poor or groundless system decisions.Source: §3.3 · pp. 10–11
• For a different under-represented patch, the performance difference between the overall image and that region was around 4%, with the model performing well there.That patch did not help define the decision boundary in this classification task, so its under-representation had minimal impact.Source: §3.3 · p. 11

• Structured data organizes information in a structured manner to facilitate tasks such as storage and querying.Source: §4 · p. 12
• The section reviews proposed techniques for identifying and mitigating representation bias in structured data.Source: §4 · p. 12
• The proposed taxonomy classifies structured-data techniques by their objectives, capabilities, and assumptions.Source: §4 · p. 12, Figure 3 · p. 12

• The Adult Income Dataset predicts whether an individual’s annual income exceeds $50K using census data.Source: §4.1 · p. 12
• The example uses six attributes: gender, race, marital status, age, hours per week, and years of experience.Source: §4.1 · p. 12
• Gender, race, and marital status are categorical, while age, hours per week, and years of experience are continuous; the ground truth is binary.Source: §4.1 · p. 12

4.2.1 Discrete Attribute Space.

• Underrepresented subgroups may lack enough records for suitable training data, and are at greater risk of unfairness in downstream algorithms [12, 88].The section describes groups using conjunctions of attribute-value assignments, such as race=black and gender=female.Source: §4.2.1 · p. 13
• For categorical attributes, coverage requires a threshold number of records per subgroup, and identification seeks maximal uncovered patterns (MUPs).No polynomial-time algorithm guarantees enumeration of all MUPs; proposed pattern-graph searches exploit parent-child relationships to prune patterns.Source: §4.2.1 · pp. 13–14
• Pattern-Breaker, Pattern-Combiner, and Deep-Diver search for MUPs, but their traversal strategies can be inefficient when uncovered patterns are sparse, numerous, or central in the graph.Deep-Diver uses depth-first search to find MUPs and prune nodes they dominate or are dominated by.Source: §4.2.1 · p. 14
• SliceFinder seeks the largest problematic data slices by balancing loss disparity, statistical significance, effect size, and slice size.Its decision-tree method searches interpretable partitions but may miss problematic slices and finds at most one among overlapping slices.Source: §4.2.1 · pp. 14–15
• To address decision trees' difficulty with overlapping slices, SliceFinder [45] also proposes Lattice Searching, where problematic slices can overlap.The section also notes that decision trees are optimized for classification and may not find every problematic slice.Source: §4.2.1 · p. 15
• Lattice search follows decision-tree slice search but can be more expensive; parallelization and sampling are suggested to address its scalability limits.The search space is exponential in the number of data slices.Source: §4.2.1 · p. 15
• SliceLine extends exact slice enumeration to find real top-k problematic slices using frequent-itemset mining, monotonicity pruning, and sparse linear algebra.Its scoring function ranks slices by score, while upper bounds enable lattice pruning.Source: §4.2.1 · pp. 15–16
• Pradhan et al. [139] identify bias-causing patterns by removing each pattern, measuring the classifier’s bias change, and ranking top-k patterns with a bottom-up search.Their interestingness measure incorporates the pattern’s data support and the difference between classifier bias before and after intervention.Source: §4.2.1 · p. 16
• Azzalini et al. detect representation bias through conditional functional dependencies, treating positive confidence differences with a sensitive attribute on the left as evidence of bias.They rank the remaining dependencies by support, impact difference, or a mean-based criterion.Source: §4.2.1 · p. 16
• Pastor et al. use divergence to compare subgroup and overall classification statistics, while DivExplorer extracts frequent attribute-value sets and computes their outcome-rate differences.They test high-divergence subgroups for statistical significance and use Shapley values to assess attribute contributions.Source: §4.2.1 · p. 16
• SSR-E ranks supplied slices by Shapley-based misclassification contribution, while FairVis clusters similar subgroups and uses feature entropy and fairness scores to surface underrepresented groups.For FairVis, entropy near zero indicates that a feature is concentrated in one value within a subgroup.Source: §4.2.1 · p. 17
• Lees et al. show that representative classifiers require adequate subgroup samples and model dimensionality aligned with subgroup distributions; otherwise, they recommend correcting representation bias through data collection.Sample-complexity bounds provide a lower limit on training examples needed from subpopulations to learn an approximately fair model with high probability.Source: §4.2.1 · p. 17
• Lin et al. address threshold-defined coverage identification across multiple relations, where joins and combinatorial COUNT queries make coverage analysis costly and require optimized query and MUP search.Pattern coverage is the number of records satisfying it in the equal join result across tables.Source: §4.2.1 · p. 17
• The MUP identification problem is NP-hard, so uses priority-based pattern-graph search and coverage-monotonicity pruning, with sampling for approximate coverage identification.The search prioritizes nodes with greater pruning efficiency; the approximate algorithm avoids needing exact pattern counts when only threshold membership matters.Source: §4.2.1 · pp. 17–18

4.2.2 Continuous Attribute Space.

• Binning continuous attributes into categories can simplify identification, but threshold-sensitive coarse groups can split similar cases, such as ages 35 and 36.The section motivates continuous-space methods with possible underrepresentation of age ranges such as age >40.Source: §4.2.2 · p. 18
• Asudeh et al. define continuous-space coverage by a vicinity radius ρ and threshold k: a query point is covered when at least k data points lie within radius ρ.The uncovered region comprises query points that fail this neighborhood-count condition.Source: §4.2.2 · p. 18
• For two-dimensional data, Uncovered-2D uses a k-th order Voronoi diagram and intersections of radius-ρ circles to identify uncovered areas outside the covered region.The paper illustrates covered and uncovered regions for its running example in Figure 6.Source: §4.2.2 · p. 19
• Extending exact search beyond two dimensions is impractical as the search space grows, so Uncovered-MD uses an ε-net learned from random query points labeled covered or uncovered.In adversarial cases, the sample count may grow exponentially with dimensions, though the section notes practical boundary complexity may be much lower.Source: §4.2.2 · pp. 19–20

4.3 Resolving Representation Bias

• Adding data can address sampling or selection bias when additional sources exist, while preventive warnings or query rewriting help users respond to historical bias when new data is costly or unavailable.Domain experts should semantically validate identified groups because some under-represented regions may be meaningless or invalid.Source: §4.3 · p. 20

4.3.1 Adding More Data.

• The greedy hitting-set approach for adding the fewest points to cover uncovered patterns has a logarithmic approximation guarantee, although the underlying problem is NP-complete.Asudeh et al. frame the task as selecting value combinations that hit uncovered patterns, targeting a desired coverage level or minimum pattern count.Source: §4.3.1 · pp. 20–21
• Azzalini et al. add tuples to repair CFDs, but adding to an advantaged group may increase discrimination; a correction step removes tuples linked to remaining CFDs.Their approach uses an improved Greedy Hit-Count algorithm [12] to select tuples before correcting remaining CFDs.Source: §4.3.1 · p. 21
• Selective data acquisition allocates examples across slices under a budget to minimize average loss and unfairness; Slice Tuner uses learning curves, while baselines can increase both.Learning curves may be unreliable when slices lack enough data or acquisition in one slice affects others; Slice Tuner updates them iteratively for dependent slices.Source: §4.3.1 · pp. 21–22
• Data augmentation enlarges datasets with altered duplicates or synthetic entries, including values for under-represented attribute groups to make their counts more equal.Adding data also incurs collection and processing costs, and collecting more examples for distribution-skew bias may violate the i.i.d. assumption.Source: §4.3.1 · pp. 20, 22
• Sharma et al. mirror tuples by flipping the sensitive attribute, then incrementally add synthetic tuples closest to the original distribution, balancing groups and improving fairness with nearly constant accuracy.The authors caution that adding too many synthetic entries may pollute the data set.Source: §4.3.1 · p. 22
• For representation bias, Iosifidis et al. propose SMOTE-based augmentation that generates minority instances by attribute, or by attribute and class to address subgroup class imbalance.Their oversampling baseline duplicates minority-subgroup instances to achieve balance.Source: §4.3.1 · p. 23
• Celis et al. learn a distribution close to the original using maximum entropy, combining reweighting and optimization to enforce sufficient-representation constraints through tuple weights.Their approach learns in polynomial time in the domain dimension.Source: §4.3.1 · p. 23
• Nargesian et al. tailor data distributions by integrating sources cost-effectively to meet user-specified subgroup counts, using dynamic programming to choose sources based on sampling and expected remaining costs.When source group distributions are known, tuples are kept only if they contribute to an unmet group count.Source: §4.3.1 · p. 24
• The dynamic-programming method becomes intractable for large group-count requirements, but for a binary sensitive attribute and similar source costs, querying the source most likely to yield a minority sample is optimal.Source: §4.3.1 · p. 24
• For the general case, Nargesian et al. [126, 127] also approximate source selection by modeling each group as a coupon-collector problem and bounding expected cost with a union bound.Source: §4.3.1 · p. 24
• For unknown group distributions, Nargesian et al. model data sources as costly bandit arms and use UCB to select sources, rewarding tuples from rarer groups.UCB balances exploration and exploitation by choosing the arm with the greatest upper confidence bound on expected reward.Source: §4.3.1 · p. 25
• Adaptive sampling [3] alternates between sampling the full population and under-represented groups, using p to trade off model performance loss against fairness loss.At p = 1 it minimizes performance loss; at p = 0 it minimizes fairness loss for the under-represented group.Source: §4.3.1 · p. 25
• Shekhar et al. allocate more of a fixed sampling budget to the hardest, worst-performing disadvantaged groups, but this and assume fair collection is attainable.Their optimism-based approach seeks data sets that converge to min-max fair solutions.Source: §4.3.1 · p. 25
• Niss et al. [130] test whether minority representation is feasible by reducing adaptive sampling to convex-hull feasibility over means of unknown distributions.They study Bernoulli and Multinomial settings and devise sampling algorithms for the problem.Source: §4.3.1 · p. 25

4.3.2 No More Data Available to Add.

• When a query point may belong to an under-represented subpopulation, a warning signal indicates whether the data cover it, leaving action to the user.For 2D data, coverage is checked within the query point’s Voronoi cell against a vicinity threshold; for multidimensional data, the last Uncovered-MD classifier determines coverage.Source: §4.3.2 · p. 26
• Data documentation and profiling help users assess dataset fitness: Gebru et al. [69] propose representativeness questions, while MithraLabel visualizes minority representation and coverage.The questions include whether a sample represents a larger set and how subpopulations are identified and distributed.Source: §4.3.2 · p. 26
• Moskovitch et al. design compact coverage labels that estimate counts for combinations of discrete attributes, trading label size against estimation error.Finding optimal labels is NP-hard, and the authors also present an optimized heuristic.Source: §4.3.2 · p. 26
• Accinelli et al. minimally rewrite preprocessing filter and merge operations so transformed data satisfy coverage constraints, using an approximate sample-based search.The approximation reflects discretization of the search space and errors in estimating cardinalities and constraint satisfaction from samples.Source: §4.3.2 · pp. 26–27
• CRBase rewrites selection conditions to find the nearest query meeting group coverage, addressing an example with 130 single and 13 married individuals but at least 70 needed in each group.It discretizes the search space and traverses candidate queries until it finds one meeting the coverage condition.Source: §4.3.2 · p. 27
• For approximate methods, Accinelli et al. measure discretization, sampling, and optimal-rewriting errors; extends the queries, constraints, and accuracy measures considered.The errors include those arising from grid discretization and sample use during preprocessing and processing.Source: §4.3.2 · p. 27
• Shetiya et al. use representation ratio as a fairness measure; SPQA uses linear-size jump pointers for sub-linear single-predicate range-query answering.SPQA seeks a fair range sufficiently similar to the original query.Source: §4.3.2 · pp. 27–28
• ≃80% Jaccard similarity is achieved by SPQA (Shetiya et al. 2021) while reducing the initial 20% male–female difference to below 5%.SPQA searches for the most similar fair range to the original query.Source: §4.3.2 · p. 28
• BFSMP searches a graph of queries whose outputs differ by one tuple, while IBFSMP uses an A*-inspired upper bound on Jaccard similarity to guide exploration.Both seek a most-similar fair range for multi-predicate queries.Source: §4.3.2 · p. 28
• Moskovitch et al. annotate tuples with query conditions, propagate provenance during evaluation, and use a minimal-changes table to generate minimal query relaxations that increase qualifying group counts.They derive candidate changes from provenance inequalities and traverse the table to retain or remove relaxations.Source: §4.3.2 · p. 28

4.4 Summary

• Attribute type records whether data occupies discrete or continuous space.Source: §4.4 · p. 28
• Relation model indicates whether the data is organized in one table or multiple tables.Source: §4.4 · p. 28
• Task distinguishes techniques that identify insufficient representation from those that resolve it.Source: §4.4 · p. 29
• Technique briefly describes the general idea of each proposed approach.Source: §4.4 · p. 29
• Figure 7 compares structured-data techniques side by side across the properties described in the section.The comparison associates each technique with its reference paper.Source: §4.4 · p. 28, Figure 7 · p. 30

5 REPRESENTATION BIAS IN UNSTRUCTURED DATA

• The paper notes extensive prior work on identifying and resolving representation bias in tabular datasets.Source: §5 · p. 31
• Representation concerns have also been investigated in unstructured data types such as images, text, and graphs.Source: §5 · p. 31
• This section reviews literature on identifying and mitigating representation bias in unstructured data.Source: §5 · p. 31

5.1.1 Identification of Representation Bias.

• Image-bias identification can convert images to tabular features or metadata, but automated recognition and labeling may perpetuate bias through errors.Dulhanty et al. [55] use face, gender, and age recognition to measure distributions in two ImageNet subsets containing human images.Source: §5.1.1 · p. 31
• REVISE audits annotated image sets using object-, person-, and geography-based metrics, but its findings are limited by available annotations.Its measures include object frequency and co-occurrence, person prominence and contextual representation, and geographic distributions.Source: §5.1.1 · pp. 31–32
• Karkkainen et al. assess subgroup diversity in image embeddings using pairwise Manhattan distances, where skew toward high distances indicates greater diversity.They visualize embeddings with t-SNE and compare a balanced face data set spanning age, race, and gender with prior work.Source: §5.1.1 · p. 32
• Cross-data-set generalization tests expose collection-related signature bias, but their interpretation assumes the training set represents the real world.Torralba et al. compare training and test sets and examine whether each set’s negative samples represent the wider world.Source: §5.1.1 · p. 32
• Khosla et al. model a shared visual world and data-set-specific biases; the combined biased model performs better on its target set but may not generalize to others.The visual-world model performs well on average but is not necessarily best on any individual data set.Source: §5.1.1 · pp. 32–33
• Attribution maps partly quantify image-classification bias, but can produce inconsistent results in some cases.Schaaf et al. evaluate methods including Grad-CAM and Integrated Gradients on artificially biased and unbiased data sets.Source: §5.1.1 · p. 33
• Hu et al. use crowd-sourced question generation, answer collection, and judgment to identify potential sampling biases in visual data sets.Crowd members inspect random image samples and judge whether statements generated from their questions and answers reflect the real world.Source: §5.1.1 · p. 33

5.1.2 Resolving Representation Bias.

• Adding targeted synthetic data improved cross-data-set generalization over original-data training at equal training-set size.Jaipuria et al. target task-specific noise factors; Georgopoulos et al. use GAN-based style transfer to create images reflecting race, gender, and age attributes.Source: §5.1.2 · p. 33
• One approach relaxes reliance on a single attribute label and adds tensor-based mixing to represent multiplicative interactions between attributes.Source: §5.1.2 · p. 33
• Yucer et al. use CycleGANs to transform facial images across racial domains while preserving identity-related traits, making race-related traits irrelevant to identity determination.Source: §5.1.2 · pp. 33–34
• CAMEL learns mappings between subgroup pairs with CycleGANs, generates transformed training examples, and uses them as augmentations to build a more robust classifier [73].The approach targets class information shared across subgroups while being oblivious to their differences.Source: §5.1.2 · p. 34
• REPAIR assigns weights to instances a classifier can more easily penalize, then formulates mitigation as a minimax problem that alternately updates classifier coefficients and resampling weights.It uses a deep neural network for feature extraction and an independent linear classifier; mitigation maximizes the ratio of reweighted-data loss to ground-truth-label uncertainty.Source: §5.1.2 · p. 34

5.2.1 Identification of Representation Bias.

• Gender-swapped evaluation and task-specific Gender Bias Evaluation Test Sets (GBETs) help assess group performance differences because standard NLP test sets may themselves underrepresent groups.Evaluation can compare false-positive or false-negative rate differences after gender swapping.Source: §5.2.1 · pp. 34–35
• For toxicity classification, Dixon et al. examine hand-curated terms, while Badjatiya et al. add skewed class occurrences and predicted class probabilities as bias signals.Terms disproportionately occurring in toxic samples or receiving high non-neutral class probabilities can encourage stereotyped predictions.Source: §5.2.1 · p. 35
• Embedding analyses identify sensitive-attribute directions: Bolukbasi et al. use gender seed words and an SVM, while Manzini et al. [120] extend the approach to non-binary gender and multiple attributes.Papakyriakopoulos et al. quantify bias using cosine distance to an inter-group direction.Source: §5.2.1 · pp. 35–36

5.2.2 Resolving Representation Bias.

• Removing explicit demographic indicators can slightly reduce bias toward minority groups, but data imbalances leave a significant gap; generated GPT-3 stories can also retain gender stereotypes without cues.Li et al. report stereotypes in generated narratives even without gender indicators or stereotype-related cues; prompt design is a possible but not universally feasible workaround.Source: §5.2.2 · p. 36
• Corpus augmentation includes training on original, gender-swapped, and named-entity-anonymized data [173], or gender-tagging sentences to preserve speaker gender for machine translation [160].Gender-tagging addresses male-dominated corpora when translating into gender-specific languages where masking gender is not an option.Source: §5.2.2 · p. 36
• Bias fine-tuning transfers from a bias-free data set to biased data, allowing models to use the former while remaining sufficiently capable on the assigned task [135].Source: §5.2.2 · p. 36
• Embedding adjustment can mitigate similarity to sensitive-attribute subspaces without retraining, but complete bias elimination is not considered feasible.Methods include projecting gender-neutral vectors away from the gender axis or making them equidistant to gender-specific words.Source: §5.2.2 · p. 37
• Manzini et al. [120] show that the preceding solution extends to non-binary sensitive attributes.Source: §5.2.2 · p. 37
• Cheng et al. [40] note that bias across sensitive attributes can be correlated or independent, and propose a sentiment direction to hard-neutralize neutral-word vectors.They form contrasting word sets, apply PCA to derive the sentiment direction, then make neutral-word vectors orthogonal to it.Source: §5.2.2 · p. 37
• Zhao et al. [174] propose embeddings that isolate sensitive information in one dimension and retain neutral information in others, allowing sensitive information to be used or ignored on demand.The method requires retraining the embeddings.Source: §5.2.2 · p. 37

5.2.3 Representation Bias in Speech Recognition.

• Speech-recognition bias studies span gender, race, and age; identification compares subgroup error rates using demographic information from annotations or automatic methods.With demographics available, identification becomes a tabular-data bias problem; mitigation recommends diversifying training datasets by adding data across demographic groups.Source: §5.2.3 · p. 37

5.3.1 Identification of Representation Bias.

• Assortative mixing evaluates graph-level correlation with a sensitive attribute: values near −1 or 1 indicate stronger correlation, while zero indicates no bias.The coefficient ranges from −1 to 1.Source: §5.3.1 · p. 38
• At the embedding level, a sensitive-attribute classifier’s weighted one-vs-rest AUC measures bias; values closer to 0.5 indicate a more nondiscriminatory graph.The survey distinguishes this embedding-level Representation Bias (RB) from representation bias as its broader topic.Source: §5.3.1 · pp. 38–39

5.3.2 Resolving Representation Bias.

• Other graph repairs align node distributions across sensitive groups with optimal transport or modify the adjacency matrix to balance sensitive-attribute homophily.The adjacency-matrix method removes edges between same-attribute nodes using randomized response during training.Source: §5.3.2 · p. 39
• Unbiased-embedding methods constrain representation learning, including a Bayesian approach that uses a strongly biased prior to reduce sensitive-attribute information.The methods place a fairness constraint on the learning objective or seek embeddings with minimal sensitive-attribute information during training.Source: §5.3.2 · p. 39
• Random-walk fairness repairs modify transition probabilities or reweight cross-group links, making generated walks more likely to include nodes from different groups.The reweighting approach assigns greater weight to cross-group links to provide a higher chance of discovery in extreme cases.Source: §5.3.2 · p. 39

5.4 Summary

• Figure 11 provides a side-by-side comparison of reviewed techniques for identifying and resolving representation bias in unstructured data.The figure’s properties are organized by data type, task, and technique.Source: §5.4 · p. 39, Figure 11 · p. 40
• The data-type dimension specifies which kind of unstructured data each work targets.Source: §5.4 · p. 39
• The task dimension distinguishes identification from resolution of insufficient representation, while the technique dimension summarizes each approach’s general idea.Source: §5.4 · p. 39

6 CONCLUSION

• The survey reviews representation-bias definitions, causes, measurement, and techniques across structured and unstructured data, using taxonomies and side-by-side comparisons.It also discusses algorithms to illustrate different challenges and the problems they address.Source: §6 · p. 41
• The authors identify addressing representation bias in other types of data sets as a promising research direction.Source: §6 · p. 41
• New data types such as streaming and spatio-temporal data raise challenges, and current solutions may not extend directly to them.Source: §6 · p. 41
• Coverage and representation rate each have potential shortcomings, leaving room for additional representation-bias metrics.Source: §6 · p. 41