betaFeedback

WorksComputer Science

Uniqueness-Aware Feature Engineering for Representing Humans in Algorithmic Decision Making

Abstract

Fairness analyses of algorithmic decision systems typically treat the feature representation of individuals as an a priori fixed input. How well a feature representation captures what is actually decision-relevant for different people and their unique backgrounds, however, is typically left unexamined. We address this gap by modeling feature engineering as a stochastic selection process over the pooled characteristics as candidate features for a finite training population. Based on this model, we develop a new framework for uniqueness-aware feature engineering which operates in two layers. The first layer measures population heterogeneity using a multivariate hypergeometric model of feature prevalence. It shows how well an individual is represented by any given input representation and which individual of the population is "least represented" compared to their personal, reference characteristics. The second layer acts on this assessment using set-similarity measures to search for an optimum between representational fit of the population and predictive accuracy. Without assuming prior knowledge about unseen individuals, our framework grants represented humans agency over what counts as decision-relevant in their case, gives deploying institutions accountability through a publishable representation fit report, and provides each person subject to a decision transparency about their personal representation fit. We operationalize our framework as a six-step pipeline and show a proof-of-principle using synthetic data. Our approach offers a blueprint for making algorithmic decision systems accountable for how well they represent individuals during development and in production.

Introduction

Eightynine-year-olds need walking aids. High-school dropouts won’t earn much later in their career. Disinterested first-year students won’t graduate with a PhD. Such and similar observations can plausibly result from group-level data analyses, but individual observations can reverse each conclusion: an 89-year-old is an active gymnast, a dropout built a successful company, and an unmotivated student was just lacking a challenge.

Group-level representations like these lead to unjustified and incorrect decisions for individuals, i.e., decisions based on incomplete knowledge or ill-fitting observations.1 Algorithmic decision about humans, such as loan granting or college admission, is particularly prone to this problem: they are built on the assumption that a single, fixed input representation captures all humans equally well. We call this the individual representation problem and illustrate it in Figure 1 (left): as individuals differ in which characteristics are genuinely predictive for their case, no single fixed input representation can faithfully capture an a heterogeneous population. This creates an inherent tension between the breadth of representational coverage and the accuracy of decision outcomes.

Figure 1. Left: The individual representation problem: individuals i1i_1 to i3i_3 each hold a unique set of task-relevant features (their reference representations did_i). A single fixed input representation xx grounded in the population covers individual reference representations to different degrees: It captures many features of individual i1i_1 and only some of i2i_2 and i3i_3. Right: The proposed six-step pipeline for uniqueness-aware feature engineering, organized in three phases: (I) Observe—collect reference representations and characterize group heterogeneity provides humans with agency towards their representation, (II) Select and Train—choose xx from the observed population and train maximizing representation coverage and prediction accuracy gives deploying institutions a tool to account for their choice of input representation, and (III) Deploy—publish a representation fit report for the observed population and inform individuals in production about their coverage to uphold transparency over time.

In this work, we ask “How can we design feature engineering strategies that account for human uniqueness and group heterogeneity?” We inspect feature engineering strategies that yield representations of humans used as input to algorithmic decision making systems (we call them input representations). In contrast to traditional feature engineering strategies, uniqueness-aware feature engineering takes into account unique circumstances and individual characteristics by gathering individual reference representations, i.e. representations of humans that document task-relevant features from their perspective. Combined with a set of measures that assess the validity, i.e., the fit, of the input representation with respect to group heterogeneity, and similarity to reference representations, uniqueness-aware feature engineering provides a framework for affected individuals to detect and contest unjustified decisions, as well as for developers to optimize representation fit and decision accuracy.

We address the individual representation problem with our uniqueness-aware feature engineering pipeline illustrated in Figure 1. It entails six steps: (1) collecting reference representation data that reflects the internal perspective of individuals on what is decision-relevant for them; (2) characterizing the heterogeneity of the observed group using the measures we introduce; (3) selecting from the most prevalent features in the pooled population while excluding features too rare to support reliable decision making; (4) calibrating the optimum between coverage and accuracy explicitly rather than absorbing it silently into model training; (5) publishing any deployed model together with a structured account of the individual representation problem; and (6) in production, informing each new individual of how well the chosen input representation fits their specific reference representation.

The pipeline is organized in three phases; each addresses a core desideratum of trustworthy algorithmic decision-making by design: Collecting reference representations grants individuals agency over what counts as decision-relevant in their case. Selecting and training with explicit coverage calibration gives deploying institutions a quantitative basis on which to be held accountable for their choice of input representation, and publishing the representation fit report alongside informing individuals at deployment upholds transparency over time.

In summary, our contributions are (1) a framework for uniqueness-aware feature engineering for representing human individuals (cf. Figure 1), (2) a set of measures to quantify the faithfulness of each individual representation, (3) a proof-of-principle analysis on synthetic data illustrating the framework’s potential, and (4) a discussion of practical implications for feature engineering processes and auditing of algorithmic decisions with human decision targets.

Related Work

Algorithmic decisions about people require abstract representations of people. These representations reduce individuals to a set of observable features and thereby necessarily omit aspects of their unique reality. The central question is therefore not only whether a feature set enables a useful prediction, but whether it adequately represents the people to whom the resulting decisions apply.

Feature engineering, feature selection, and representation. Feature engineering converts unstructured and often high-dimensional raw data to computationally processable features that are more structured and often of lower dimensionality. Verdonck et al. [40] survey and classify various feature engineering strategies across domains, differentiating two major strategies: adjusting existing features using transformations or creating new features. Traditional feature engineering strategies include dimensionality reduction techniques, such as PCA. Reducing raw data to the most frequent and predictive features allows computation of more accurate results over the entire population but disregards individual features that only apply to a few, reinforcing the assumption that those reductions have the same effect on all individuals.

A structured account of this problem is offered by Friedler et al. [16], who distinguish between idealized construct spaces for features and decisions and their imperfect observed counterparts used as proxies in formal decision systems. Jacobs and Wallach [23] refine this further through measurement modeling, showing that construct–proxy mismatches lead to systemic errors and inequities when observable features fail to operationalize task-relevant constructs. Suresh and Guttag [37] identify three sources of such measurement bias: construct oversimplification, variation in feature measurement across prediction targets, and variation in measurement accuracy across groups. Taken together, these works establish that feature selection is not a neutral technical choice but a locus of potential harm.

Grgic-Hlaca et al. [19] show that people assess features as unfair when they are perceived as irrelevant or unreliable with respect to the decision at hand. This highlights the role of perceived relevance in decision justification and motivates approaches that measure misalignment between individuals’ relevant traits and the features used by the algorithm. Work considering fair use of features from the perspective of represented individuals shows that each person would select different sets to represent them, even within a closed set [19]. However, these approaches still operate within a predefined feature space that hinders representation of unique entities such as humans.

This is related to what the psychometric tradition addresses through construct validity, convergent validity, and measurement invariance [11, 32] — formal frameworks for assessing whether a measurement instrument tracks the latent construct it is intended to reflect, though with a focus on groups rather than individuals.

Zając et al. [42] find that omitting background knowledge on a patient from an x-ray labeling task in an attempt to reduce annotator bias paradoxically produces noisier labels, because expert annotators are deprived of information they would rely on in the real clinical task. Thus, omitting features from classification makes outcomes noisy, even when humans are the annotators.

Bengio et al. [5] surveyed over 200 papers on representation learning in 2013. Almost all of this body of work focuses on representing objects that are either born-digital or digitized, and on compressing them by selecting the most relevant features from a finite set. Only a small fraction considers representations of humans [4, 34, 43], and these primarily focus on human physiology. The question of whether a feature representation is valid for a given human individual, in the sense of capturing what that individual considers relevant about themselves, remains largely unaddressed.

This fixed feature set creates a circular validation problem: outcomes are evaluated against the same features from which they were constructed, so characteristics outside the set can never challenge the representation. It may appear adequate simply because the population has already been reduced to it.

Representational harm and measurement. Prior work demonstrates that harms can originate in representation before a model computes a decision outcome. Failures of closed-set representations have been documented in education [27, 35], employment [6, 12, 15], finance [2, 26, 30], and public administration [3, 7, 8, 17, 29, 31, 33]. For example, pandemic-era grading systems relied on teacher-derived rankings that failed to capture individual circumstances [27].

This concern is formalized in work on measurement and representation. Friedler et al. [16] distinguish constructs from their observational representations, while Jacobs and Wallach [23] and Suresh and Guttag [37] show how mismatches between constructs and observations can produce error and inequity. In information retrieval, Tobia et al. [38, 39] similarly show that standardized user models can poorly represent children. These works establish that representation matters, but generally seek better representations for an entire group rather than measuring the fidelity of a representation to each individual.

Work by Lu et al. [28] and Katzman et al. [24] emphasizes the importance of self-identification and dynamic, evolving representations, but does not provide methods for comparing these against existing system representations.

Stereotypes and the problem of validation. The problem is related to the concept of stereotypes. Allport describes a stereotype as an “exaggerated belief associated with a category” used to justify conduct toward that category [1 p. 191]. Group-level representations are not necessarily stereotypes in the social sciences sense, but they share the relevant property of reducing heterogeneous individuals to a single limited set of characteristics. Aside from fitting individuals differently well, the validity of representations for the population itself is questionable: if characteristics outside the group-level set are never observed, they cannot challenge the representation. A representation may thus be treated as adequate without ever being tested against the full diversity of characteristics present in the population.

The German Ethics Council has identified “Risk[s] to the individual posed by statistical stratification” [13]. Recently, Höltgen and Williamson [20] challenge the assumption, prevalent in machine learning practice, that data originates from a pre-existing underlying distribution, arguing that the causal direction is in fact reversed. They call attention to the underexplored role of representational choices, specifically, how the input space and its covariates are defined, as a foundational yet neglected determinant of model behavior. Our work fills this research gap by proposing a validation strategy for representations of individuals that does not assume individuals originate from an underlying distribution.

Human representation validation is a blind spot in fairness research. Algorithmic fairness research provides extensive methods for evaluating decisions across groups and individuals [9, 10, 36]. Many approaches represent individuals as points in a feature space and evaluate distances, similarities, or outcomes within that space [14, 18, 21, 25, 41]. These methods are essential for assessing how decisions are made from representations, but generally take the representation itself as given and assume the same degree of faithfulness for each individual.

Individual fairness, introduced by Dwork et al. [14], posits that similar individuals should be treated similarly with respect to a given decision task. Rather than grappling with the potentially large set of possible representations of individuals, Dwork et al. [14] model individuals as single instances in a metric space. Subsequent research has largely focused on defining, learning, or circumventing the need for a universal distance metric between such individuals [18, 21, 25, 41]. This line of work assumes that the features supplied to the metric are adequate proxies for the individuals’ actual characteristics. Our work questions precisely this assumption: whether the features supplied to the decision process adequately capture the individual being represented. Thus, a decision can satisfy a fairness criterion relative to a feature space while that feature space fails to capture important characteristics of the people within it.

Heterogeneity of a population and Jaccard. We draw on Paul Jaccard’s observation that “probably no two square metres of vegetation in the whole Alps, possess exactly the same floristic composition” [22], projecting square metres to individuals and plants to predictive features. We use the Jaccard Index to quantify representational similarity and individual rarity within a population; formal definitions follow in § 3.

In what follows, we introduce uniqueness-aware feature engineering, focusing on reducing harm to human individuals through group-level inaccuracies in representation.

A Framework for Uniqueness-Aware Feature Engineering

We formalize the individual representation problem and lay out the framework that addresses it. The framework has two parts: a set of entities that define the setting precisely, and a six-step pipeline that operationalizes the framework from data collection through individual-level transparency of representation fit at deployment.

Setting

Individuals. We understand individuals i∈Ii \in I as real humans with unique and fluctuating traits, experiences, and identities. The individuals in scope are subjects of an algorithmic decision, such as loan-granting or college-admission.

Task. The decision task tt specifies what the algorithmic system is asked to predict or decide for each individual, such as the probability of credit default, academic success, or criminal re-offense. The task determines which features are potentially relevant and defines what counts as a decision-relevant characteristic in the following.

Reference representations. A reference representation di\mathbf{d}_i is a per-individual, task-specific characterization of which features are decision-relevant from ii’s own perspective. For example, reference representations may be self-reported by the individual, provided by a knowledgeable proxy such as a caseworker or teacher. We treat di\mathbf{d}_i as a binary feature vector over a possibly open feature space: a feature is either reported (1) or not (0), with no duplicates per individual. Where multiple reference representations exist for one individual, we assume they can be aggregated under deduplication to a single vector that contains each relevant feature exactly once.

Reference representations are a novel construct we introduce in this work. They are not features selected by a data scientist or observed passively in a dataset; they are task-relevant characterizations generated or validated by or for the individual themselves from an internal and subjective perspective. Taking access to reference representations as given is a prerequisite of the framework; how they are collected is a subject of ongoing work (§ 5.1).

Pooled population. The pooled population D\mathcal{D} is the multiset of all features reported across all reference representations in the observed group of individuals during training of the algorithmic decision system:

D=⋃i∈Idi.\mathcal{D} = \bigcup_{i \in I} \mathbf{d}_i.

Some features appear in many reference representations; others appear in only one. This heterogeneity is the central object the framework measures and acts on.

Training population. The training population is set of individuals and their reference representations that are observed during the feature engineering process, including selecting the input representation and training the algorithmic decision system. Before production, the pooled population contains only the reference representations in the training population (and the held-out set).

Input representations. An input representation x\mathbf{x} is a fixed feature set used by an algorithmic decision system to represent every individual. It is the same for all individuals and is derived from D\mathcal{D} through feature selection. The individual representation problem arises precisely from the conflict between the population-level x\mathbf{x} and the individual di\mathbf{d}_i: There is a conflict between the coverage of all individually decision relevant features in a heterogeneous population and the population-level useful features to compute a statistically grounded algorithmic decision with.

The Uniqueness-Aware Feature Engineering Pipeline

We operationalize the framework as a six-step pipeline, illustrated in Figure 1. Steps 1–2 characterize the population before any feature is selected. Steps 3–4 govern feature selection and model training. Steps 5–6 fulfil transparency obligations at deployment and in production.

Step 1: Collect reference representations. Gather di\mathbf{d}_i for all individuals in the training population. The collection process should be task-specific: individuals (or knowledgeable proxies like case workers) document what they consider decision-relevant for for this individual and the particular task at hand, despite of what is observeable in an entire group. We assume the resulting reference representations are complete and truthful to the best of each individual’s knowledge; limitations arising from imperfect self-knowledge are discussed in § 5.

Step 2: Characterize population heterogeneity. Compute the measures introduced in § 3—full coverage probability, similarity (here evaluated pairwise across the population), and rarity—for every individual in the training population. This step answers the question: How diverse is the pooled population, and are minority subgroups systematically rarer? A population in which most individuals report nearly the same features is tractable for uniform feature selection; a population in which individuals report highly idiosyncratic feature sets signals that any fixed input representation will leave many poorly covered, and alternative strategies to algorithmic decisions, such as human oversight, or human-only decisions, should be considered.

Step 3: Select features by prevalence. Choose a prevalence threshold τ∈[0,1]\tau\in[0,1] and retain in the input representation x\mathbf{x} all features whose prevalence p(f)p(f), i.e. the fraction of individuals who report it, in the pooled population meets or exceeds τ\tau:

x(τ)={f∈D∣p(f)≥τ}.\mathbf{x}(\tau) = \{f \in\mathcal{D} \mid p(f) \geq\tau\}.

Features below τ\tau are excluded on the grounds that too few reference representations support reliable estimation of their predictive effect. The choice of τ\tau is calibrated in Step 4.

Step 4: Calibrate the coverage–accuracy optimum. Raising τ\tau prunes rare features, which can improve predictive accuracy up to the point where the group is underspecified but also reduces the fraction of each individual’s reference representation that xx covers. Low coverage leads to poor representational fit of affected individuals. This optimum is not incidental: it is a structural consequence of the individual representation problem, and it does not only fall on individuals but aggregates on minority subgroups whose relevant features are precisely the ones most likely to fall below any positive threshold.

We make this optimization problem explicit by sweeping τ\tau and computing, for each candidate threshold, mean individual coverage and held-out predictive accuracy (cf. § 3). A minimum coverage floor is set (e.g. by the deploying institutions and society). τ\tau is chosen so that accuracy is maximized within this admissible region above the coverage floor. Additionally reporting worst-group coverage makes baldly represented subgroups visible, which is a direct consequence of the threshold choice.

Step 5: Publish a structured representation report. Before deployment, publish a structured account of the chosen input representation: which features are included and which are excluded, the full calibration curve (coverage and accuracy as a function of τ\tau), the chosen threshold and its justification, and the group-level coverage gaps at that threshold. This report is the transparency artefact that enables external auditing, regulatory review, and institutional accountability. It makes explicit what any feature selection decision implies for different individuals and groups: Information that is currently absorbed silently into model training with no external record.

Step 6: Inform each individual of their fit. During production (at decision time), compute the individual coverage (§ 3) for each new individual against the deployed input representation and potentially assign them coverage classes for simpler communication — e.g., poorly, partially, or well represented — defined by percentile thresholds of that score in the training population. The individual fit report also identifies any features in did_i that are absent from the training population (never-seen features): such features systematically lower the individual coverage and if repeatedly reported during production may signal that the deployed representation may need recalibration to include emerging characteristics not present when the system was trained.

Measures

We introduce a set of measures that together assess population heterogeneity, guide feature selection, and quantify individual and system-wide representation fit. Prevalence and rarity operate at the feature level and characterize the pooled population, full coverage probability characterizes how distinctive each individual reference representation is within that population, similarity and coverage measure the fit between a chosen input representation and an individual reference representation. Aggregates of these measures to system-level and subgroup-level quantities are used during calibration and at deployment.

Prevalence and Rarity

Let k(f)k(f) denote the number of reference representations in the pooled population that contain feature ff, and let N=∣I∣N = |I| be the total number of individuals. The prevalence of feature ff is the fraction of individuals who report it:

p(f)=k(f)N.p(f) = \frac{k(f)}{N}.

The rarity of feature ff is the complement of its prevalence—the fraction of individuals who do not report it:

Rare⁡(f)=1−p(f)=1−k(f)N.\operatorname{Rare}(f) = 1 - p(f) = 1 - \frac{k(f)}{N}.

A feature with Rare⁡(f)≈1\operatorname{Rare}(f) \approx1 appears in almost no reference representations; a feature with Rare⁡(f)≈0\operatorname{Rare}(f) \approx0 is reported by nearly everyone.

The rarity of an individual’s reference representation di\mathbf{d}_i summarizes how uncommon their reported features are in the pooled population. We use the median rarity over the individual’s reported features:

Rare⁡(di)=median⁡f∈diRare⁡(f).\operatorname{Rare}(\mathbf{d}_i) = \underset{f \in\mathbf{d}_i}{\operatorname{median}} \operatorname{Rare}(f).

Prevalence drives feature selection (Step 3): features with p(f)≥τp(f) \ge\tau are retained in the input representation x(τ)\mathbf{x}(\tau); features below the threshold are excluded. Rarity characterizes which individuals are most affected by that exclusion.

Full Coverage Probability

Full coverage probability PiP_i is a special case of the multivariate hypergeometric distribution, adapted for our binary setting: each feature in the pooled population is either reported by an individual or it is not. This corresponds to an urn model in which each ball can be drawn at most once. Hence, PiP_i answers “Given the pooled population, how likely is it that a randomly drawn feature set exactly matches individual ii’s reference representation di\mathbf{d}_i?” A low PiP_i indicates that ii’s feature combination is highly distinctive and unlikely to arise by chance from the population distribution.

Sampling without replacement, the probability that the features actively reported by individual ii are drawn is:

Pi=∏j:di[j]=1aj(∑j=1∣di∣aj∥di∥1).P_i = \frac{\prod_{j:\mathbf{d}_i[j]=1} a_j}{\binom{\sum_{j=1}^{|\mathbf{d}_i|} a_j}{\|\mathbf{d}_i\|_1}}.

where aja_j is the count how often a specific feature at position jj in ii’s reference representation appears in the pooled population given that it was actively reported by individual ii (i.e., where di[j]=1\mathbf{d}_i[j] = 1).

Because PiP_i reflects both the number of reported features and how common those features are, it serves as a compact measure of how easy or difficult it is to match any specific individual with a uniform input representation. In heterogeneous populations, PiP_i is typically vanishingly small for all individuals, and is best interpreted on a log⁡10\log_{10} scale or relative to the population distribution of log⁡10Pi\log_{10} P_i.

The most demanding scenario for feature engineering arises when PsystemP_{\mathrm{system}} is low and the distribution of PiP_i is (approximately) identical. In this case, no single input representation is likely to adequately cover most individuals, and there is no exploitable variation in individual representability to guide the decision on which features should compose x\mathbf{x}. This configuration of the full coverage probability distribution is a signal that a single pooled feature set may be fundamentally insufficient, and that alternative strategies—individualized decisions, or explicit flagging of poorly covered instances at deployment—should be considered before proceeding with a uniform feature selection approach.

Coverage and Similarity

The coverage of input representation xx for individual ii’s reference representation did_i is their Jaccard index:

Ji(x,di)=∣x∩di∣∣x∪di∣J_i(x,d_i)=\frac{\lvert x\cap d_i\rvert}{\lvert x\cup d_i\rvert}

Ji=1J_i=1 means the input representation is a perfect match for ii’s reference representation; Ji=0J_i=0 means no overlap.

We suggest to implement thresholds (e.g. as percentile fit classes as implemented in [5]) of poorly represented individuals that in order to flag those for human review rather than automated processing.

Three system-level aggregates of JiJ_i are used during calibration and auditing:

Population Similarity (JpopJ_{\mathrm{pop}}). The mean pairwise Jaccard index of reference representations in the pooled population:

Jpop=2N(N−1)∑i<i′J(di,di′).J_{\mathrm{pop}}=\frac{2}{N(N-1)}\sum_{i<i'}J(d_i,d_{i'}).

JpopJ_{\mathrm{pop}} measures how similar individuals are to one another in their reference representations, independently of any chosen input representation. A low JpopJ_{\mathrm{pop}} indicates high population heterogeneity, which signals that uniform feature selection will inevitably leave many individuals partially covered.

System Coverage (JsystemJ_{\mathrm{system}}). The mean coverage across all individuals in the reference population at a given threshold τ\tau:

Jsystem(τ)=1N∑i=1NJi(x(τ),di).J_{\mathrm{system}}(\tau)=\frac{1}{N}\sum_{i=1}^{N}J_i(x(\tau),d_i).

JsystemJ_{\mathrm{system}} is the primary quantity used to calibrate τ\tau (Step 4): it measures what fraction of each individual’s reference features the chosen input representation captures on average.

Worst-group coverage (JworstJ_{\mathrm{worst}}). The minimum of JsystemJ_{\mathrm{system}} computed separately for each subgroup g⊆Ig\subseteq I:

Jworst(τ)=min⁡g1∣g∣∑i∈gJi(x(τ),di).J_{\mathrm{worst}}(\tau)=\min_{g}\frac{1}{\lvert g\rvert}\sum_{i\in g}J_i(x(\tau),d_i).

JworstJ_{\mathrm{worst}} surfaces subgroup-level representational disparities that are averaged away in JsystemJ_{\mathrm{system}}, and is published as part of the structured representation report (Step 5).

These measures decompose the individual representation problem along four axes: which features the population can reliably support (prevalence), how unusual an individual’s features are within that population (rarity), what a specific chosen input representation actually achieves for each individual (coverage), and how likely it is to find an input representation with perfect coverage given the population structure (full coverage probability). Together they cover population characterization and individual assessment, with prevalence and rarity grounding feature selection for the input representation in the training population’s specific heterogeneity, and full coverage probability and coverage and similarity providing the before-and-after view of representational fit for each individual and deploying institution.

Prototype on Synthetic Data

We illustrate the six-step pipeline on a controlled synthetic dataset in which ground truth is fully known. The goal is not to validate the measures on real individuals—that requires empirical work discussed in § 5.1—but to trace the pipeline end to end and make its outputs concrete.2

Synthetic population

Data generation (Step 1). We draw a population of Nref=400N_{\mathrm{ref}} = 400 individuals whose reference representations are binary feature vectors over F=570F = 570 reported (1) or not reported (0) possible features. The feature space is partitioned into three kinds: Shared features (150) are reported by most individuals regardless of subgroup; Subgroup features (120) are specific to one of four subgroups A–D, with population shares 60 %, 25 %, 10 %, and 5 %; and Idiosyncratic features (300) are sampled from a long-tailed distribution; many are reported by only one or two individuals.

Each individual’s feature count follows the same long-tailed distribution (median 29 features). A binary outcome yiy_i is assigned to each individual. An additional Nnew=100N_{\mathrm{new}} = 100 individuals are held back as new arrivals to simulate representation reports in production in Phase III; arrivals also report brand-new features—feature keys that no training individual reported—at an average rate of 1.5 per arrival.

Phase I: assessing population heterogeneity (Step 2)

Population-level measures. Table 1 (a)-(Phase I) summarizes the pooled population after computing the measures for every individual in the training data. Note that every individual has a unique feature vector: no two individuals in the training population share exactly the same set of reported features. Full coverage probabilities quantify how unlikely it is that a randomly drawn input representation would exactly match an individual’s reference representation. Because every individual has a unique feature vector, these probabilities are vanishingly small for everyone: median log⁡10Pi\log_{10} P_i ranges from −34.9 for the majority subgroup A to −41.5 for subgroup C (corresponding to probabilities of roughly 10−3510^{-35} to 10−4210^{-42}). The more negative values for minority subgroups reflect that their subgroup-specific features appear in only a small fraction of the pooled population, making their exact feature combinations rarer still. Rarity scores are uniformly high (Rare(did_i) = 85.2 %), meaning the typical individual reports features that fewer than 15 % of the population also report. Population-level Jaccard similarity (JpopJ_{\mathrm{pop}}, the mean pairwise Jaccard index across all reference representations) is 0.132: any two individuals share, on average, only 13 % of their reported features.

Phase IPhase IPhase IPhase IPhase IPhase IIPhase IIPhase IIPhase IIIPhase IIIPhase III
GroupnnFeatureslog⁡10Pi\log_{10} P_iJpopJ_{\mathrm{pop}}Rare(did_i) %JsystemJ_{\mathrm{system}}AccuracyWell repr.Partially repr.Poorly repr.
A23729−34.9-34.90.13884.80.9350.73432310
B9829−37.1-37.10.12788.00.9390.63317100
C4331−41.5-41.50.12088.60.8800.814160
D2228−41.0-41.00.11788.30.7080.727012
AllΣ\Sigma 400Avg 29−35.7-35.70.13285.20.8660.727Σ\Sigma 50482
τ\tauFeat.JsystemJ_{\mathrm{system}}JworstJ_{\mathrm{worst}}Acc.
0.0005390.9890.9870.688
0.0202360.9180.7080.717
0.100950.7000.5400.670
0.450100.2020.1990.518
UnseennnJiJ_i (mean)Percentile (med.)WellPartialPoor
0210.93867.51830
1330.90734.122110
2310.86715.610192
3+150.8389.40150

Table 1. Prototype evaluation results across all phases (Section 5). (a) Population-level uniqueness measures (Phase I), per group coverage and accuracy at τ=0.02\tau= 0.02 (Phase II), and per group coverage classes for 100 new arrivals (Phase III). (b) Coverage–accuracy optimization across prevalence thresholds (Phase II). (c) Effect of never-seen features on fit score and coverage classes (Phase III).

Minority subgroups are systematically rarer. Subgroup D’s individuals have a median rarity of 88.3 % and a median log⁡10Pi\log_{10} P_i of −41.0, compared with 84.8 % and −34.9 for the majority subgroup A (Table 1). This reflects the design of the synthetic population—subgroup-specific features are, by construction, uncommon across the full pooled population—but the same dynamic is expected in real-world populations where minority characteristics are underrepresented in data collection. Figure 2 visualizes the measure distributions by subgroup.

Population heterogeneity by subgroup

Figure 2. Distribution of the measures by subgroup (Phase I). Full coverage probabilities are shown on a log⁡10\log_{10} scale after normalization. Minority subgroups (C, D) exhibit lower full coverage probabilities and higher rarity scores, reflecting their subgroup-specific and idiosyncratic feature sets.

Phase II: feature selection and calibration (Steps 3–4)

Prevalence threshold sweep. We select an input representation x\mathbf{x} by prevalence, retaining all features that meet or exceed a prevalence-threshold τ\tau in the pooled population. Prevalence-based selection represents the best-case scenario for representational coverage: because x\mathbf{x} is grounded in what the training population actually reports, most individuals are well-served by it. An input representation chosen without reference to the observed population—in the worst case, one containing features nobody would ever report—would yield poor coverage for nearly everyone, making prevalence the natural baseline against which more targeted selection strategies should be measured. For each τ\tau, we compute via five-fold cross-validation: (i) JsystemJ_{\mathrm{system}}, the mean share of each held-out individual’s reference features that the system retains—our measure of representational coverage—and (ii) held-out predictive accuracy using logistic regression.

Table 1 and Figure 3 show the resulting optimization function. A low τ\tau keeps almost the entire feature space (τ=0\tau= 0: 539 features, Jsystem=0.989J_{\mathrm{system}} = 0.989) but yields accuracy of only 0.688. Raising τ\tau prunes rare features, reducing coverage while initially increasing accuracy. The chosen threshold τ=0.02\tau= 0.02 selects 236 features, achieves Jsystem=0.918J_{\mathrm{system}} = 0.918 and accuracy 0.717 ±\pm 0.046, and is the lowest threshold that keeps Jsystem≥0.60J_{\mathrm{system}} \ge0.60—our minimum coverage requirement. Above τ=0.10\tau= 0.10 coverage and accuracy decline together: the representation no longer captures enough signal to be predictive.

Coverage–accuracy optimization and coverage by subgroup

Figure 3. Coverage–accuracy optimization as a function of the prevalence threshold τ\tau (Phase II). The shaded region marks Jsystem<0.60J_{\mathrm{system}} < 0.60, below the minimum coverage requirement. The chosen threshold (τ=0.02\tau= 0.02, dotted line) maximises held-out accuracy within the admissible region. Per-group JsystemJ_{\mathrm{system}} curves show that coverage degrades fastest for minority subgroups as τ\tau increases.

Subgroup coverage gaps. Even at the chosen threshold, representational coverage is unevenly distributed (Table 1) (a)-(Phase II). Majority subgroup A keeps 93.5 % of its features on average; minority subgroup D keeps only 70.8 %. This gap arises because subgroup-specific features of smaller groups fall below the prevalence threshold: a feature reported by 5 % of the population meets τ=0.02\tau= 0.02 in expectation, but when the subgroup itself constitutes only 5 % of individuals, most of its distinctive features fall below τ\tau. The worst-group coverage JworstJ_{\mathrm{worst}} drops from 0.987 at τ=0\tau= 0 to 0.708 at the chosen threshold—a direct, measurable consequence of applying a minimum prevalence requirement.

Phase III: deployment (Steps 5–6)

Representation report (Step 5). The deployed system uses 241 of the 570 features from the pooled population: 149 of 150 shared features, 91 of 120 subgroup features, and 1 of 300 idiosyncratic features. The single retained idiosyncratic feature meets the prevalence threshold because it happens to be reported by a sufficient fraction of the training population; the remaining 299 idiosyncratic features are pruned. The structured representation report packages these facts alongside the calibration summary and the per-group coverage gap for public release.

Individual fit scoring at arrival time (Step 6). Each of the 100 new arrivals is scored against the deployed system using their reported feature names. Table 1 (a)-Phase III summarizes the results. Of the 100 arrivals, 50 are well represented (Ji≥0.90J_i \ge0.90), 48 are partially represented (0.75≤Ji<0.900.75 \le J_i < 0.90), and 2 are poorly represented (Ji<0.75J_i < 0.75). All poorly represented individuals belong to subgroup D.

Never-seen features as a signal. Seventy-nine of the 100 arrivals report at least one feature that no individual in the reference population reported. Table 1 (c) shows that the presence of never-seen features consistently lowers the fit score (JiJ_i): arrivals reporting zero unseen features have a mean fit score of 0.938 and a median percentile of 67.5; those reporting three or more unseen features have a mean fit score of 0.838 and a median percentile of 9.4, placing them in the bottom decile of the pooled population’s coverage distribution. 85 distinct never-seen features appear across the 79 affected arrivals; several are reported by multiple arrivals (e.g. two keys each appear in four arrival reports), suggesting that some represent emergent subgroup characteristics the training data did not yet capture. These features are flagged as candidates for the next recalibration pass.

Examplary individual fit report. A new arrival receiving a low fit score illustrates Step 6. Individual new_074 (subgroup D) is poorly represented: Ji=0.69J_i = 0.69 at the 2nd percentile of the reference population. The system keeps 20 of 29 reported features; the nine omitted features include six subgroup-specific features, one idiosyncratic feature, and two features the system has never seen. Their rarity score is Rare⁡(di)=87.3%\operatorname{Rare}(d_i) = 87.3\% and log⁡10Pi=−43.4\log_{10} P_i = -43.4. This individual is flagged by the pipeline as candidate for human review rather than automated processing.3

Summary of prototype findings. The prototype confirms three structural properties of the framework: uniqueness is universal — every individual in the training population has a distinct feature vector and vanishingly small full coverage probability — minority subgroups are systematically disadvantaged at any positive prevalence threshold, and the coverage–accuracy optimization is empirically non-trivial, with the threshold that maximizes accuracy entailing measurable representational loss concentrated in the least-covered groups. Never-seen features additionally serve as a recalibration signal: their systematic degradation of individual fit scores provides an operational mechanism for detecting when the deployed representation has drifted from the population it serves.

Discussion

Heterogeneity within a population is not merely a statistical property to be acknowledged and set aside; it is a fundamental determinant of which input representation best serves that population. Feature engineering must therefore be grounded in an understanding of the population a model is trained on, and the variation of individuals with respect to any chosen representation must be measured and communicated. This assessment cannot be performed once at training time and assumed to hold indefinitely: it must extend to individuals newly entering the system in production.

Our feature engineering framework and associated prototype build on the notion that uniqueness of the represented individuals is the norm, not the exception. It does not overcome harm through misrepresentation, but in contrast to traditional feature engineering strategies, makes this harm visible and quantifies it. Trustworthy algorithmic decision-making requires structures that communicate to affected individuals and deploying institutions how appropriate it is to apply a formalized decision to any specific person, and that allow the public to assess for how whom algorithmic treatment could be appropriate and for how whom human oversight is warranted.

When is uniqueness-aware feature engineering necessary? The approach is most clearly warranted when the subjects of an algorithmic decision are human individuals: they are unique, they hold an internal perspective on what is predictive for them, and they can be structurally harmed when that perspective is ignored. The boundary is less sharp for non-human subjects. At first glance, an artifact like a vehicle lacks moral standing and is not exposed to structural bias. On closer inspection, however, uniqueness resurfaces: individual driving behavior and usage patterns may carry high predictive value yet remain invisible to conventional feature engineering, and the stakes of misclassification can be severe. A principled guideline delineating when uniqueness-aware feature engineering is warranted—and when the added complexity is not—remains an open task.

Benefits of uniqueness-aware feature engineering. At the individual and group level, the framework provides a principled basis for quantifying how well an input representation captures what is genuinely predictive for a specific person or group, enabling transparency grounded in features the individuals themselves would recognize as relevant. At the system level, aggregate measures allow institutions and auditors to assess not merely whether a model is accurate on average, but whether its accuracy is fairly distributed across the population it governs. For individuals and groups whose features fall outside the chosen feature set, the framework provides a mechanism to flag cases for human review rather than allowing an unjustified automated decision to stand unchallenged.

Challenges of uniqueness-aware feature engineering. The framework assumes that each person’s reference representation is complete and correct—that it accurately captures all characteristics predictive of the decision outcome for that individual. In reality, individuals may forget relevant features, may not know which of their characteristics are predictive, or may misjudge their own relevance. We believe this does not undermine the framework’s value: if an individual regards their reference representation as adequate, our measures can quantify their input representation’s adequacy from their perspective. At the system level, aggregated subjective assessments—even if individually skewed—carry a meaningful signal about a system’s faithfulness.

A deeper structural concern is different access to realistic reference representations. For example, people differ in their ability to articulate what is decision-relevant for them, and this variation may systematically disadvantage individuals or groups. The framework must be carefully adapted to mitigate this risk.

Practical challenges center on constructing reference representations. It is not clear which modality best conveys decision-relevant information for a given task, nor what should be extracted—feature keys, values, or both—and how similar types of information should be grouped. Each extraction method introduces its own systematic distortions, which compound across the pipeline. Methods that make these distortions legible to researchers, institutions, and affected individuals are as urgently needed as the extraction methods themselves.

Future Work

Before the framework can be applied at scale, empirical and methodological groundwork is needed to establish how reference representations are created—who produces them, in what medium, in what format—and how the resulting distributional properties vary across domains and tasks. Reliable conversion of raw representations into structured feature vectors is a prerequisite. We regard natural language texts, such as task-specific self-descriptions of individuals, processed with NLP methods, as the most promising direction. A specific challenge here is semantic equivalence: two people sharing a relevant feature may describe it in entirely different words, and mapping such surface variation onto a common feature space requires robust paraphrase detection and semantic normalization [22].

An important extension is to ask whether representational adequacy can degrade over time. Our measures offer a detection mechanism: monotonically declining similarity scores or full coverage probabilities for individuals entering the system in production suggest the representation is drifting out of alignment with lived reality. Repeated reporting of the same features that have not been observed in the training population but accumulate in production may reveal emergent subgroups absent from the training data and be also a signal for retraining the decision system.

Conclusion

This work addresses the individual representation problem: no group-level input representation can faithfully capture all individuals, and the resulting gap between what group-level features the system needs and what is decision-relevant in an individual level carries substantial consequences for the justifiability of resulting decisions. We address it with a framework and prototype operating in two layers: the first diagnoses population heterogeneity, revealing where the difficulty of the individual representation problem for the training population; the second acts on this diagnosis by choosing the input representation with optimal representational coverage and accuracy. Operationalized as a six-step pipeline, the framework grants individuals agency over what counts as decision-relevant, gives institutions a quantitative basis for accountability through a publishable representation fit report, and upholds transparency by informing each person of their representational fit at deployment.

Endmatter

Generative AI Usage Statement

The authors used the following generative AI tools during the preparation of this manuscript:

Claude Sonnet 4.6 (Anthropic, accessed via Claude desktop application) was used for English language editing and style improvements, brainstorming paper titles, developing names for proposed concepts (e.g., the “individual representation problem”), generating a mock peer review to anticipate reviewer feedback, and suggesting ways to shorten and tighten the manuscript.

Claude Code with Opus 5.5 (Anthropic, accessed via Visual Studio Code) was used to implement the final prototype based on a code draft and specification of the 6-step pipeline written by the authors.

Gemini 3.5 Flash (Google, accessed via Google Colab) was used to assist with coding the initial draft notebook.

The authors retain full responsibility for the originality and integrity of all content, including the correctness of all claims, references, and citations. All AI-generated suggestions were reviewed, verified, and revised by the authors prior to inclusion.

Ethical Considerations

The proposed approach introduces several potential risks. First, collecting and processing real reference representations (as opposed to abstract, synthetic ones in the paper) raises concerns about privacy and informational self-determination, as individuals may disclose sensitive attributes that could be misused. In the event of data breaches, such information could be exploited to train systems for harmful or discriminatory purposes.

Second, same as input representations, reference representations may lead to unequal representation fit: individuals differ in their ability of expression, which may systematically disadvantage certain groups. This can result in biased or inaccurate measurements that affect downstream decisions and fit reports.

Finally, improving the apparent trustworthiness and interpretability of decision systems may create a false sense of reliability. Systems that appear transparent or aligned may be less critically scrutinized, making harmful behavior harder to detect and potentially increasing their impact.

Positionality Statement

Our perspective is shaped by a Western, privileged academic context with a background in Computer Science and Arts. Accordingly, the problems addressed in this work reflect concerns emerging in relatively well-resourced settings and should not be prioritized over more immediate threats to safety, rights, and well-being in less privileged contexts.

We take an individual-centered view on fairness, focusing on alignment between generalized system representations and individuals’ unique reality. This choice reflects awareness of the limits of our own perspective—we cannot assume knowledge of all relevant or marginalized groups.

Our work is further informed by a background that emphasizes the uniqueness and irreducibility of human experience. Motivated by concerns about the simplification inherent in computational systems, we aim to identify and question reductions of complex human realities, while recognizing that our perspective is partial and that broader impact requires engagement beyond it.

References

References

  1. [1]Gordon W. Allport. 1954. The Nature of Prejudice. Addison-Wesley Publishing Company, Cambridge, MA. 537 pages. https://ia601500.us.archive.org/17/items/allport.dli.2015.188638/2015.188638.The-Nature-Of-Prejudice.pdf
  2. [2]Edmund L. Andrews. 2021. How Flawed Data Aggravates Inequality in Credit. https://archive.ph/jVn7d
  3. [3]Julia Angwin, Jeff Larson, Surya Mattu, and Lauren Kirchner. 2016. Machine Bias. https://archive.ph/NgH9S
  4. [4]Nikos Athanasiou, Mathis Petrovich, Michael J. Black, and Gü l Varol. 2022. TEACH: Temporal Action Composition for 3D Humans. In 2022 International Conference on 3D Vision (3DV). 414–423. doi:10.1109/3DV57658.2022.00053DOI
  5. [5]Yoshua Bengio, Aaron Courville, and Pascal Vincent. 2013. Representation Learning: A Review and New Perspectives. IEEE Trans. Pattern Anal. Mach. Intell. 35, 8 (Aug. 2013), 1798–1828. doi:10.1109/TPAMI.2013.50DOI
  6. [6]Miranda Bender. 2019. All the ways hiring algorithms can introduce bias. https://archive.ph/pcw8a
  7. [7]Janna Brancolini. 2022. Italy Turns to AI to Find Taxes in Cash-First, Evasive Culture. https://archive.ph/2JwNb
  8. [8]Eva Constantaras, Gabriel Geiger, Justin-Casimir Braun, Dhruv Mehrotra, and Htet Aung. 2023. Inside the Suspicion Machine. https://archive.ph/DlmMa
  9. [9]Sam Corbett-Davies, Johann D. Gaebler, Hamed Nilforoshan, Ravi Shroff, and Sharad Goel. 2023. The measure and mismeasure of fairness. J. Mach. Learn. Res. 24, 1, Article 312 (Jan. 2023), 117 pages.
  10. [10]Emily Corvi, Hannah Washington, Stefani Reed, Chad Atalla, Alexandra Chouldechova, P Alex Dow, Jean Garcia-Gathright, Nicholas Pangakis, Emily Sheng, Dan Vann, Matthew Vogel, and Hanna Wallach. 2025. Taxonomizing representational harms using speech act theory. (2025). arXiv:2504.00928 [cs.CL]
  11. [11]Lee J. Cronbach and Paul E. Meehl. 1955. Construct Validity in Psychological Tests. Psychological Bulletin 52, 4 (1955), 281–302. doi:10.1037/h0040957DOI
  12. [12]Jeffrey Dastin. 2018. Amazon scraps secret AI recruiting tool that showed bias against women. https://archive.ph/MWiCy
  13. [13]Deutscher Ethikrat. 2023. Mensch und Maschine – Herausforderungen durch Künstliche Intelligenz. Technical Report. Deutscher Ethikrat, Berlin. https://www.ethikrat.org/fileadmin/Publikationen/Stellungnahmen/deutsch/stellungnahme-mensch-und-maschine.pdf
  14. [14]Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. 2012. Fairness through Awareness. In Proceedings of the 3rd Innovations in Theoretical Computer Science Conference (Cambridge, Massachusetts) (ITCS ’12). Association for Computing Machinery, New York, NY, USA, 214–226. doi:10.1145/2090236.2090255DOI
  15. [15]Alexander Fanta. 2021. Jobcenter-Algorithmus landet vor Höchstgericht. https://archive.ph/varKJ
  16. [16]Sorelle A. Friedler, Carlos Scheidegger, and Suresh Venkatasubramanian. 2021. The (im)possibility of fairness: different value systems require different mechanisms for fair decision making. Commun. ACM 64, 4 (March 2021), 136–143. doi:10.1145/3433949DOI
  17. [17]Gabriel Geiger. 2023. How Denmark’s Welfare State Became a Surveillance Nightmare. https://archive.ph/lePUL
  18. [18]Stephen Giller, Christopher Jung, Michael Kearns, and Aaron Roth. 2018. Online Learning with an Unknown Fairness Metric. In Advances in Neural Information Processing Systems, Vol. 31. Curran Associates, Inc. https://proceedings.neurips.cc/paper/2018/hash/50905d7b2216fbecc b5b41016357176b-Abstract.html
  19. [19]Nina Grgic-Hlaca, Elissa M. Redmiles, Krishna P. Gummadi, and Adrian Weller. 2018. Human Perceptions of Fairness in Algorithmic Decision Making: A Case Study of Criminal Risk Prediction. In Proceedings of the 2018 World Wide Web Conference (Lyon, France) (WWW ’18). International World Wide Web Conferences Steering Committee, Republic and Canton of Geneva, CHE, 903–912. doi:10.1145/3178876.3186138
  20. [20]Benedikt Höltgen and Robert Charles Williamson. 2026. The Costs of Pretending That There Are Data-Generating Probability Distributions in the Social World. Association for Computing Machinery, New York, NY, USA, 5219–5236. https://doi.org/10.1145/3805689.3806478
  21. [21]Christina Ilvento. 2020. Metric Learning for Individual Fairness. (2020), 11.arxiv.org/abs/1906.00250
  22. [22]Paul Jaccard. 1912. The Distribution of the Flora in the Alpine Zone. New Phytologist 11, 2 (1912), 37–50. doi:10.1111/j.1469-8137.1912.tb05611.x
  23. [23]Abigail Z. Jacobs and Hanna Wallach. 2021. Measurement and Fairness. In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (Virtual Event, Canada) (FAccT ’21). Association for Computing Machinery, New York, NY, USA, 375–385. doi:10.1145/3442188.3445901DOI
  24. [24]Jared Katzman, Angelina Wang, Morgan Scheuerman, Su Lin Blodgett, Kristen Laird, Hanna Wallach, and Solon Barocas. 2023. Taxonomizing and measuring representational harms: a look at image tagging. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence (AAAI’23/IAAI’23/EAAI’23). AAAI Press, Article 1601, 9 pages. doi:10.1609/aaai.v37i12.26670DOI
  25. [25]Michael Kim, Omer Reingold, and Guy Rothblum. 2018. Fairness Through Computationally-Bounded Awareness. In Advances in Neural Information Processing Systems. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett (Eds.), Vol. 31. Curran Associates, Inc. https://proceedings.neurips.cc/paper/2018/file/c8dfce5cc68249206e4690fc4737a8d-Paper.pdfarxiv.org/abs/1803.03239
  26. [26]Will Knight. 2019. The Apple Card Didn’t ’See’ Gender—and That’s the Problem. https://archive.ph/ypqKA
  27. [27]Tom Lamont. 2021. The student and the algorithm: how the exam results fiasco threatened one pupil’s future. https://archive.ph/GNC3e
  28. [28]Christina Lu, Jackie Kay, and Kevin McKee. 2022. Subverting machines, fluctuating identities: Re-learning human categorization. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency (Seoul, Republic of Korea) (FAccT ’22). Association for Computing Machinery, New York, NY, USA, 1005–1015. doi:10.1145/3531146.3533161DOI
  29. [29]Aaron Mak. 2019. Facing Facts. https://archive.ph/yDwmp
  30. [30]Emmanuel Martinez and Lauren Kirchner. 2021. The Secret Bias Hidden in Mortgage-Approval Algorithms. https://archive.ph/sakVy
  31. [31]Morgan Meaker. 2023. The Fraud-Detection Business Has a Dirty Secret. https://archive.ph/bikFx
  32. [32]Samuel Messick. 1987. VALIDITY. ETS Research Report Series 1987, 2 (1987), i–208. arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1002/j.2330-8516.1987.tb00244.x doi:10.1002/j.2330-8516.1987.tb00244.xDOI
  33. [33]Jonathan Oosting. 2023. Big fines compound false fraud claims in Michigan. https://archive.ph/AVVu0
  34. [34]Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. 2019. Expressive Body Capture: 3D Hands, Face, and Body from a Single Image. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR). 10975–10985. doi:10.1109/CVPR.2019.01123DOI
  35. [35]Juan Carlos Perdomo, Talon Britton, Moritz Hardt, and Rediet Abebe. 2025. Difficult Lessons on Social Prediction from Wisconsin Public Schools. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’25). Association for Computing Machinery, New York, NY, USA, 2682–2704. doi:10.1145/3715275.3732175DOI
  36. [36]Cynthia Rudin. 2019. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence 1, 5 (May 2019), 206–215.DOI
  37. [37]Harini Suresh and John Guttag. 2021. A Framework for Understanding Sources of Harm throughout the Machine Learning Life Cycle. In Proceedings of the 1st ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (–, NY, USA) (EAAMO ’21). Association for Computing Machinery, New York, NY, USA, Article 17, 9 pages. doi:10.1145/3465416.3483305DOI
  38. [38]Diletta Micol Tobia, Hrishita Chakrabarti, Maria Soledad Pera, and Monica Landoni. 2026. Using Anti-Personas to Model Children: How to Represent User-System Mismatches. In Proceedings of the 34th ACM Conference on User Modeling, Adaptation and Personalization (UMAP ’26). Association for Computing Machinery, New York, NY, USA, 282–291. doi:10.1145/3774935.3806168DOI
  39. [39]Diletta Micol Tobia, Isabella Possaghi, Hrishita Chakrabarti, Maria Soledad Pera, and Monica Landoni. 2026. Are These the Ways? Children’s Search Practices in Classroom Online Information Seeking Tasks. In Proceedings of the 25th Annual ACM Interaction Design and Children Conference (IDC ’26). Association for Computing Machinery, New York, NY, USA, 1005–1010. doi:10.1145/3773077.3812169DOI
  40. [40]Tim Verdonck, Bart Baesens, María Óskarsdóttir, and Seppe Vanden Broucke. 2024. Special issue on feature engineering editorial. Machine Learning 113, 7 (2024), 3917–3928. doi:10.1007/s10994-021-06042-2
  41. [41]Gal Yona and Guy Rothblum. 2018. Probably Approximately Metric-Fair Learning. In Proceedings of the 35th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 80), Jennifer Dy and Andreas Krause (Eds.). PMLR, 5680–5688. https://proceedings.mlr.press/v80/yona18a.html
  42. [42]Hubert Dariusz Zając, Natalia Rozalia Avlona, Finn Kensing, Tariq Osman Andersen, and Irina Shklovski. 2023. Ground Truth Or Dare: Factors Affecting The Creation Of Medical Datasets For Training AI. In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society (Montréal, QC, Canada) (AIES ’23). Association for Computing Machinery, New York, NY, USA, 351–362. doi:10.1145/3600211.3604766DOI
  43. [43]Wenjie Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. 2022. MotionBERT: A Unified Perspective on Learning Human Motion Representations. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) (2022), 15039–15053. https://api.semanticscholar.org/CorpusID:259982676arxiv.org/abs/2210.06551

Paper details

vlnc:W_besrqv55
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 11

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Claude Sonnet 4.6 (bare), Claude Code, Gemini 3.5 Flash

View statement
Looking for related papers…
Contents