Biased Aggregation Approach: Methods & Applications

• Biased Aggregation Approach is an umbrella term for procedures that embed explicit corrective biases to counter systematic distortions and align outputs with normative targets.
• It encompasses various implementations including fairness-constrained ranking, selection-bias correction, exposure-aware recommendation, and prior-guided federated learning.
• These methods intentionally trade off estimator neutrality for greater fairness and robustness, balancing accuracy with bias correction in diverse algorithmic settings.

“Biased aggregation approach” is best understood as an umbrella designation for aggregation procedures that do not treat combination as a neutral averaging step, but instead encode an explicit corrective or directional preference into the aggregation rule. In the recent literature, this includes rank aggregation with prefix-wise group fairness constraints, pairwise-comparison models that separate item quality from presentation effects, opinion aggregation that rebalances voter-attribute skew, recommendation pipelines that counter exposure or popularity bias, and dynamic-programming or federated-learning schemes that bias aggregation toward a value-function prior or a server-held foundation model (Chakraborty et al., 2023, Bugakova et al., 2019, Ueda et al., 2023, Dong et al., 2020, Bertsekas, 2019, Park et al., 2024). A plausible implication is that the term does not denote a single estimator class; rather, it denotes a design principle in which aggregation is deliberately altered so that the output better matches a normative target, a latent unbiased signal, or a trusted prior.

1. Scope and conceptual taxonomy

In the surveyed literature, biased aggregation appears in at least four technically distinct forms. The first is constraint-based aggregation, in which the aggregate must satisfy fairness or representation conditions. The second is selection-bias correction, in which the observed inputs are reweighted or modeled to undo observation, participation, or presentation effects. The third is exposure- or context-aware aggregation, in which the combination rule is modified to counter systematic distortions such as popularity bias, format bias, or neighborhood noise. The fourth is prior-guided aggregation, in which a known value function, belief approximation, or foundation model biases aggregation toward a more stable region of the hypothesis space.

Setting	Bias mechanism	Representative papers
Rank, vote, and label aggregation	Fairness constraints, weighting, post-processing	(Chakraborty et al., 2023, Ueda et al., 2023, Singer et al., 30 Jan 2026)
Pairwise, observational, and participation bias	Factor models, IPS, minimax correction, Bayesian constraints	(Bugakova et al., 2019, Ueda et al., 2023, Guo et al., 6 Feb 2025, Meertens et al., 2019)
Recommendation and graph aggregation	Rank fusion, inverse-propensity neighborhood weights, contextual standardization, biased walks	(Dong et al., 2020, Kim et al., 2022, Li et al., 15 Mar 2026, Pfeifer et al., 5 Mar 2026)
Dynamic programming, expert advice, and FL	Bias function $V$, belief-space correction, learned expert weights, FM-biased server aggregation	(Bertsekas, 2019, Li et al., 7 Jul 2025, Abels et al., 2021, Park et al., 2024)

The phrase is also polysemous. In the diffusion-limited aggregation literature, “biased” refers to particles being limited to move in a subset of possible directions, yielding the $k$-DLA family and complexity results such as P-completeness for Prediction in 2-DLA and NL-completeness in 1-DLA (Bitar et al., 2019). That usage is structurally different from the statistical and algorithmic senses above.

2. Fairness-constrained aggregation

The clearest formalization of bias-aware aggregation as a normative design choice appears in “Fair Rank Aggregation” (Chakraborty et al., 2023). Candidates are partitioned into disjoint groups $G_1,\dots,G_g$, and designers specify proportional lower and upper bounds $\alpha_i,\beta_i$ on group representation in top-$k$ prefixes. With $C_i(k,\pi)$ denoting the number of items from group $i$ in the top $k$ of ranking $\pi$, the framework imposes prefix constraints of the form $L_{i,k} \le C_i(k,\pi) \le U_{i,k}$. The paper distinguishes $k$0-fairness, block-fairness, and strict-fairness, and then studies the closest fair ranking problem under Kendall tau and Ulam distance. For Kendall tau, it proves an intra-group stability property and gives the exact linear-time AltGreedy algorithm for closest $k$1-fair ranking and the exact $k$2-time StrongGreedy algorithm for closest block-$k$3-fair ranking. For Ulam distance, it gives an exact dynamic program for strict-fairness when $k$4, with runtime $k$5. At the aggregation level, its meta-algorithm yields 3-approximation guarantees for center and median under both Kendall tau and Ulam, and it further gives a $k$6-approximation for fair median under Ulam with strong fairness.

Fairness-aware opinion aggregation adopts a different formal object: the target is not a ranking but a probabilistic soft label. “Mitigating Voter Attribute Bias for Fair Opinion Aggregation” (Ueda et al., 2023) defines fair aggregation as estimating the aggregate opinions that would arise from an ideally balanced voter population with attribute distribution $k$7, even when the observed pool is unbalanced. It criticizes majority voting and classical Dawid–Skene in subjective settings because the single-truth assumption and EM dynamics tend to produce overly sharp posteriors. Its Soft D&S model replaces a discrete latent truth with a soft ground-truth vector $k$8, uses Dirichlet priors on both $k$9 and worker confusion matrices $G_1,\dots,G_g$0, and optimizes the posterior by alternating constrained numerical optimization. The fairness layer is then supplied by weighted majority voting, data splitting, or GroupAnno. The paper reports that Soft D&S combined with data splitting is effective for dense data, whereas weighted majority voting is effective for sparse data.

Crowdsourced aggregation under explicit demographic parity constraints is developed further in “Optimal Fair Aggregation of Crowdsourced Noisy Labels using Demographic Parity Constraints” (Singer et al., 30 Jan 2026). In the binary case, the demographic-parity gap is $G_1,\dots,G_g$1. The paper analyzes Majority Vote and Optimal Bayesian aggregation, derives a small-crowd upper bound on the fairness gap of Majority Vote in terms of the annotators’ fairness gaps, and shows that under interpretable conditions the fairness gap of the aggregated consensus converges exponentially fast to the ground-truth fairness gap. Since the ground truth itself may remain unfair, it also generalizes a multiclass fairness post-processing method to the discrete setting, producing a randomized classifier that enforces strict demographic parity constraints for any aggregation rule.

3. Correcting presentation, participation, and observation bias

A second major family treats aggregation as a latent-variable estimation problem under biased observation. “Aggregation of pairwise comparisons with reduction of biases” (Bugakova et al., 2019) augments Bradley–Terry with a worker-specific mixture between an unbiased quality term and a factor-driven bias term. For worker $G_1,\dots,G_g$2, item scores $G_1,\dots,G_g$3, task-factor vector $G_1,\dots,G_g$4, sensitivity vector $G_1,\dots,G_g$5, and unbiased-mode propensity $G_1,\dots,G_g$6, the model is

$G_1,\dots,G_g$7

This makes presentation artifacts such as left-right position, branding cues, or ordering first-class components of the aggregation model. Parameters are fit by maximizing a regularized likelihood with virtual-node regularization and conjugate gradient descent. The reported empirical studies show that factorBT is more accurate and more robust to biased pairwise comparisons than BT, CrowdBT, HITS, and a linear position-bias model.

When the bias lies in which labels are observed rather than how they are presented, inverse-propensity weighting becomes central. “Mitigating Observation Biases in Crowdsourced Label Aggregation” (Ueda et al., 2023) introduces IPS into EM-style label aggregation. Let $G_1,\dots,G_g$8 indicate whether worker $G_1,\dots,G_g$9 responded to task $\alpha_i,\beta_i$0, and let $\alpha_i,\beta_i$1 be the propensity. The IPS lower bound reweights each observed contribution by $\alpha_i,\beta_i$2, yielding an unbiased objective with respect to a uniform observation process. The resulting E-step posterior is

$\alpha_i,\beta_i$3

and the M-step inherits the same weighting. Propensities are estimated from the binary response matrix via 1-bit matrix completion. The paper verifies improved aggregation accuracy under strong observation bias and increased robustness to both spam and colluding workers.

Participation bias in rating aggregation is addressed from a robust minimax perspective in “Mitigating the Participation Bias by Balancing Extreme Ratings” (Guo et al., 6 Feb 2025). Ratings $\alpha_i,\beta_i$4 are revealed with value-dependent probability $\alpha_i,\beta_i$5, so observed samples are missing-not-at-random. In the known-$\alpha_i,\beta_i$6 setting, the Balanced Extremes Aggregator estimates the unrevealed mean as a convex combination of the scale extremes and then blends that estimate with the observed mean. In the unknown-$\alpha_i,\beta_i$7 setting, the Polarizing-Averaging Aggregator computes lower and upper feasible means consistent with the revealed histogram and the bound $\alpha_i,\beta_i$8, then returns their midpoint. The paper proves that PAA is minimax optimal as $\alpha_i,\beta_i$9, gives a finite-sample near-optimality bound, and reports that PAA is closest to the survey ground truth on a hotel-rating dataset, whereas BEA can over-correct when real distributions are smooth rather than polarized.

The same correction principle appears in classification-based totals. “A Bayesian Approach for Accurate Classification-Based Aggregates” (Meertens et al., 2019) starts from the linear relation $k$0, where $k$1 is the confusion matrix and $k$2 is the vector of predicted class totals. Naive inversion of $k$3 can be unstable and can produce impermissible estimates such as negative counts or totals when $k$4 is estimated from a small labeled set. The paper therefore puts Dirichlet priors on the columns of $k$5, samples from the posterior, and truncates the parameter space to

$k$6

thereby ensuring feasible corrected counts. On company tax returns, the constrained Bayesian estimator substantially reduces mean squared error relative to unconstrained Bayesian correction when the labeled set is small.

4. Exposure-aware recommendation and graph aggregation

In recommender systems, biased aggregation often means counteracting popularity-driven exposure concentration. “Alleviating the recommendation bias via rank aggregation” (Dong et al., 2020) introduces the Two-Way Rank Aggregation framework, which linearly combines a user-oriented forward rank and an item-oriented backward rank: $k$7 The empirical premise is that forward ranks are strongly correlated with item degree, whereas backward ranks are approximately uniform across degrees. The paper evaluates the framework using Precision, Hamming Distance, and the Gini coefficient, and reports that on MovieLens, TWRA(P3) with $k$8 reaches Precision $k$9, HD $C_i(k,\pi)$0, and Gini $C_i(k,\pi)$1, compared with P3 at Precision $C_i(k,\pi)$2, HD $C_i(k,\pi)$3, and Gini $C_i(k,\pi)$4; on Netflix, $C_i(k,\pi)$5 yields analogous fairness gains together with higher Precision.

Bias can also enter at the message-passing stage. “Debiasing Neighbor Aggregation for Graph Neural Network in Recommender Systems” (Kim et al., 2022) argues that inverse propensity weighting should not be confined to the loss; it should modify the neighborhood aggregation operator itself. It estimates propensities from relative item popularity,

$C_i(k,\pi)$6

uses edge weights $C_i(k,\pi)$7, and then applies Laplacian normalization in either random-walk or symmetric form. In LightGCN form, this yields $C_i(k,\pi)$8. Across Alexa, Yahoo Music, and MovieLens, Navip improves ranking quality, with gains up to $C_i(k,\pi)$9.

A broader contextual version of exposure correction is presented in “MBD: A Model-Based Debiasing Framework Across User, Content, and Model Dimensions” (Li et al., 15 Mar 2026). There, the aggregation problem is not rank fusion but the construction of a value model from heterogeneous behavioral signals such as watch time, loop rate, like probability, and comment probability. MBD learns contextual means and variances for each signal with respect to a chosen partial feature set $i$0, then standardizes the signal into a z-score or percentile before aggregation. For example,

$i$1

This makes “unbiasedness” itself configurable through the cohort definition. The paper reports offline NLL reductions of more than 50% versus cluster baselines and near-zero correlation between the debiased watch-time signal and duration, together with positive online effects on watch time, likes, and ecosystem efficiency.

A structurally different, but conceptually similar, use of biased aggregation appears in “Robust Node Affinities via Jaccard-Biased Random Walks and Rank Aggregation” (Pfeifer et al., 5 Mar 2026). TopKGraphs anchors a random walk at a start node $i$2, biases transitions toward neighbors with high Jaccard overlap with $i$3, converts each walk into a partial first-visit ranking, and then aggregates the rankings with a penalized Borda mean: $i$4 The resulting affinity matrix is interpretable and non-parametric. Across SBM, LFR, k-nearest-neighbor graphs, CORA, and a high-confidence PPI network, the method is reported to be competitive or superior to Jaccard, Dice, personalized PageRank, and Node2Vec, especially in sparse, noisy, or heterogeneous networks.

5. Prior-guided and model-based biased aggregation

In approximate dynamic programming, “Biased Aggregation, Rollout, and Enhanced Policy Improvement for Reinforcement Learning” (Bertsekas, 2019) introduces a bias function $i$5 that shifts classical aggregation from direct approximation of $i$6 to approximation of $i$7. With aggregation map $i$8 and low-dimensional correction $i$9, the final value approximation is

$k$0

The aggregate Bellman equation is identical in structure to classical aggregation except for the shaped stage cost $k$1. When $k$2, the method reduces to classical aggregation; when $k$3 and there is a single aggregate state, it is equivalent to rollout; with multiple aggregate states and $k$4, it yields a stronger policy-improvement mechanism for approximate policy iteration. Error bounds are expressed in terms of the Bellman residual of $k$5 or the intra-aggregate variation of $k$6.

The same idea is transferred to POMDP belief space in “Feature-Based Belief Aggregation for Partially Observable Markov Decision Problems” (Li et al., 7 Jul 2025). The paper first aggregates unobservable states into feature states, then aggregates feature beliefs into a finite representative set $k$7, solves an aggregate MDP, and uses the solution for one-step lookahead in the original POMDP. Its biased variant again introduces a bounded bias function $k$8, replaces the stage cost by

$k$9

and proves an error bound of the form

$\pi$0

where $\pi$1 is now the footprint-set oscillation of $\pi$2. This makes the quality of the bias function directly measurable through the reduction of local oscillation.

A data-driven version of prior-guided aggregation appears in collective decision-making with expert advice. “Dealing with Expert Bias in Collective Decision-Making” (Abels et al., 2021) maps the advice vectors of $\pi$3 experts into a contextual bandit whose meta-context for arm $\pi$4 is $\pi$5. The aggregation rule learns weights $\pi$6 so that $\pi$7 approximates the expected reward. In the linear case, it maintains

$\pi$8

A crucial consequence is that experts worse than random can receive negative weights, so the aggregation process can effectively invert biased advice. The paper gives regret bounds for Meta-CMAB variants and reports that the approach outperforms best-expert selection baselines, especially when expert quality degrades or the group is heterogeneous.

In federated learning, prior-guided aggregation is instantiated as server-side mixing with a concealed foundation model. “FedBaF: Federated Learning Aggregation Biased by a Foundation Model” (Park et al., 2024) computes the standard client aggregate $\pi$9, then biases it toward a masked projection of the foundation model $L_{i,k} \le C_i(k,\pi) \le U_{i,k}$0 via

$L_{i,k} \le C_i(k,\pi) \le U_{i,k}$1

where $L_{i,k} \le C_i(k,\pi) \le U_{i,k}$2 and $L_{i,k} \le C_i(k,\pi) \le U_{i,k}$3 is randomized on the server. The update is the closed-form minimizer of a strictly convex proximal objective, and the randomization is used to prevent recovery of $L_{i,k} \le C_i(k,\pi) \le U_{i,k}$4. The paper reports that FedBaF often matches or surpasses weight initialization by a foundation model while preserving model confidentiality, with gains up to $L_{i,k} \le C_i(k,\pi) \le U_{i,k}$5 in IID settings, up to $L_{i,k} \le C_i(k,\pi) \le U_{i,k}$6 in non-IID settings, and perplexity reductions up to $L_{i,k} \le C_i(k,\pi) \le U_{i,k}$7.

6. Limitations, trade-offs, and open directions

A recurring limitation is that biased aggregation almost always trades one error source against another. In fair rank aggregation, fairness constraints can increase distance from the input rankings, the exact Ulam algorithm is only polynomial for a constant number of groups, overlapping groups are not addressed, and breaking the factor 3 barrier for Kendall tau remains open (Chakraborty et al., 2023). In presentation-bias correction, factorBT uses a smooth but non-convex objective, may overfit when the number of factors grows, and becomes difficult to identify when factors are strongly collinear (Bugakova et al., 2019). In opinion aggregation, only a binary attribute is studied, GroupAnno can be unstable in low-data regimes, and multi-attribute or continuous attributes are left for future work (Ueda et al., 2023). In demographic-parity-constrained crowdsourcing, strict DP can be too strong for some applications, and intersectional extensions remain open (Singer et al., 30 Jan 2026).

Selection-bias correction methods are also sensitive to their nuisance models. The participation-bias framework assumes homogeneous reveal probabilities bounded below by $L_{i,k} \le C_i(k,\pi) \le U_{i,k}$8, and BEA can over-correct when the true rating distribution is smooth rather than polarized (Guo et al., 6 Feb 2025). Navip estimates exposure propensities only from item popularity, so user- or context-dependent exposure remains outside its model class (Kim et al., 2022). IPS-based aggregation inherits the usual variance inflation of inverse-propensity weighting, so small estimated propensities must be regularized or stabilized (Ueda et al., 2023). Contextual debiasing with MBD depends heavily on the chosen partial feature set $L_{i,k} \le C_i(k,\pi) \le U_{i,k}$9: too few features leave residual bias, too many create sparse cohorts, unstable variance estimates, and the risk of over-normalizing genuine preferences (Li et al., 15 Mar 2026).

A final caution is that aggregation can itself be the source of inferential bias even when no explicit correction is intended. “Data aggregation can lead to biased inferences in Bayesian linear mixed models and Bayesian ANOVA” (Schad et al., 2022) shows that by-subject aggregation can make Bayes factors too conservative for contrasts with small random-slope variance and too liberal for contrasts with large random-slope variance, while ignoring item random slopes yields liberal Bayes factors. That result places an important boundary on the broader idea of biased aggregation: not every aggregation scheme that appears statistically efficient is inferentially well calibrated.

Taken together, these works suggest that biased aggregation is becoming a general methodological pattern rather than a niche trick. The pattern is consistent across ranking, crowdsourcing, recommendation, graph mining, reinforcement learning, POMDPs, and federated learning: aggregation is modified so that the final estimator is representation-aware, observation-aware, prior-aware, or fairness-aware. The central technical question is therefore not whether aggregation should be biased, but how the bias should be parameterized, justified, and controlled for a given objective.