BiFair: Bi-Level Fairness Framework

• BiFair is a framework that uses bi-level optimization to address dual sources of bias—prior unfairness and training unfairness—in LLM-enhanced recommender systems.
• It employs an adaptive inter-group balancing mechanism to reduce disparities in group utility metrics, achieving lower coefficient of variation (CV) and higher minimum utility (MIN).
• In stochastic bandits, BiFair decomposes fairness into group exposure and meritocratic allocation, ensuring anytime fairness guarantees with UCB-style algorithms.

Searching arXiv for papers explicitly using or closely associated with the term “BiFair”. arxiv_search(query="BiFair", max_results=10, sort_by="submittedDate") arxiv_search(query="Bi-Level Fairness stochastic bandits meritocracy exposure fairness", max_results=10, sort_by="relevance") BiFair most directly denotes a fairness-aware training framework for LLM-enhanced recommender systems built around bi-level optimization and an adaptive inter-group balancing mechanism. In a related use, the term also describes a bi-faceted fairness objective in stochastic bandits, where fairness is enforced both across groups and within groups. Across adjacent literatures, the same conceptual pattern reappears as two-level optimization, two-sided fairness constraints, or biologically informed fairness, although those works typically use different names rather than “BiFair” itself (Zhang et al., 6 Jul 2025, Pokhriyal et al., 2024).

1. Conceptual scope

In the recommender-systems literature, BiFair is defined at the level of model training. The central claim is that unfairness in LLM-enhanced recommender systems has two sources: prior unfairness in LLM-generated item representations and training unfairness introduced when the recommendation model is optimized. BiFair addresses both simultaneously by treating the representation layer and the trainable projector as two coupled optimization targets.

In the stochastic-bandit literature, an explicitly related idea appears under the name Bi-Level Fairness. There, fairness is decomposed into a macro-level exposure guarantee across groups and a micro-level meritocratic allocation rule within each group. The resulting framework is interpretable as a “BiFair” formulation because it imposes fairness at two coupled levels rather than along a single axis.

This yields a useful general characterization. BiFair is not a single fairness metric. It is a structural design principle in which fairness is split into two interacting components: representation and training, across-group and within-group, or one-side and bi-side constraints, depending on the task domain.

2. BiFair in LLM-enhanced recommender systems

The paper "BiFair: A Fairness-aware Training Framework for LLM-enhanced Recommender Systems via Bi-level Optimization" formulates BiFair for a two-stage LLM-enhanced recommendation pipeline. In Stage I, a frozen LLM or embedding model maps item textual metadata $x_i$ to semantic vectors $z_i = f_{\mathrm{LLM}}(x_i)$, and user representations are formed by aggregating interacted item vectors, $z_u = \frac{1}{|I_u|}\sum_{i\in I_u} z_i$. In Stage II, a trainable projector $g_{\boldsymbol{\theta}}$ maps semantic vectors into the recommendation latent space, $e_i = g_{\boldsymbol{\theta}}(z_i)$ and $e_u = g_{\boldsymbol{\theta}}(z_u)$, after which a scoring function $s(e_u,e_i)$ predicts interaction likelihood $\hat y_{u,i} = s(e_u,e_i)$.

The framework targets item-side group fairness. Items are partitioned into groups $G_I=\{G_1,\dots,G_N\}$ by attributes such as popularity or genre. Group utility is defined as

$$\text{G\_Utility}(n)@K = \frac{1}{|U|}\sum_{u\in U}\mathrm{utility}(u,n)@K,$$

and $z_i = f_{\mathrm{LLM}}(x_i)$0-Item-side Fairness requires

$z_i = f_{\mathrm{LLM}}(x_i)$1

To summarize group disparities, the framework uses $z_i = f_{\mathrm{LLM}}(x_i)$2 and $z_i = f_{\mathrm{LLM}}(x_i)$3, the average utility of the bottom $z_i = f_{\mathrm{LLM}}(x_i)$4 of groups. Lower CV and higher MIN indicate better fairness.

BiFair operationalizes its two-source unfairness decomposition through a nested optimization problem. The lower level optimizes projector parameters for fixed representations,

$z_i = f_{\mathrm{LLM}}(x_i)$5

and the upper level refines the semantic representations themselves,

$z_i = f_{\mathrm{LLM}}(x_i)$6

The full bi-level program is therefore

$z_i = f_{\mathrm{LLM}}(x_i)$7

The fairness-aware loss is itself adaptive. For each item group $z_i = f_{\mathrm{LLM}}(x_i)$8, BiFair defines a group loss

$z_i = f_{\mathrm{LLM}}(x_i)$9

combines these losses with a learned simplex weight vector $z_u = \frac{1}{|I_u|}\sum_{i\in I_u} z_i$0, and uses the entropy of the softmaxed group-loss vector as a fairness signal. The update direction is

$z_u = \frac{1}{|I_u|}\sum_{i\in I_u} z_i$1

and the weight search is solved with Frank–Wolfe. This mechanism is intended to emphasize high-loss groups while maintaining non-adverse alignment with every group’s descent direction.

The empirical evaluation uses three Amazon Review domains—Movies, Games, and Books—with popularity fairness and genre fairness as groupings. The reported results show that BiFair achieves the lowest CV and the highest or near-highest MIN on all datasets for popularity fairness, and the lowest CV and highest MIN on most genre-fairness settings, while maintaining nearly identical or slightly better accuracy relative to AlphaRec and outperforming prior fairness baselines. Averaged over datasets, it reduces CV by about $z_u = \frac{1}{|I_u|}\sum_{i\in I_u} z_i$2 for popularity fairness and $z_u = \frac{1}{|I_u|}\sum_{i\in I_u} z_i$3 for genre fairness relative to the strongest baseline, while increasing MIN by around $z_u = \frac{1}{|I_u|}\sum_{i\in I_u} z_i$4 and $z_u = \frac{1}{|I_u|}\sum_{i\in I_u} z_i$5, respectively (Zhang et al., 6 Jul 2025).

3. Bi-Level Fairness in stochastic bandits as a BiFair formulation

In stochastic multi-armed bandits, the related notion is Bi-Level Fairness. Arms are partitioned into groups $z_u = \frac{1}{|I_u|}\sum_{i\in I_u} z_i$6, and the framework distinguishes fairness in group exposure from fairness in arm exposure within each group. Group exposure at time $z_u = \frac{1}{|I_u|}\sum_{i\in I_u} z_i$7 is

$z_u = \frac{1}{|I_u|}\sum_{i\in I_u} z_i$8

where $z_u = \frac{1}{|I_u|}\sum_{i\in I_u} z_i$9 is the pull count of arm $g_{\boldsymbol{\theta}}$0.

The first level is $g_{\boldsymbol{\theta}}$1-Group Exposure Fairness. For a fairness vector $g_{\boldsymbol{\theta}}$2, with $g_{\boldsymbol{\theta}}$3 and $g_{\boldsymbol{\theta}}$4, a policy satisfies group exposure fairness if

$g_{\boldsymbol{\theta}}$5

This is an anytime guarantee: each group must receive at least its minimum share at every horizon $g_{\boldsymbol{\theta}}$6.

The second level is Meritocratic Fairness within each group. Each arm $g_{\boldsymbol{\theta}}$7 has unknown mean reward $g_{\boldsymbol{\theta}}$8, and a merit function $g_{\boldsymbol{\theta}}$9 assigns positive merit $e_i = g_{\boldsymbol{\theta}}(z_i)$0 subject to Lipschitz continuity and boundedness. Conditional on selecting group $e_i = g_{\boldsymbol{\theta}}(z_i)$1, the within-group policy $e_i = g_{\boldsymbol{\theta}}(z_i)$2 is meritocratic if

$e_i = g_{\boldsymbol{\theta}}(z_i)$3

The fair-optimal within-group policy is

$e_i = g_{\boldsymbol{\theta}}(z_i)$4

A policy is $e_i = g_{\boldsymbol{\theta}}(z_i)$5-Bi-Level Fair if it satisfies anytime group exposure fairness and, for each group $e_i = g_{\boldsymbol{\theta}}(z_i)$6, the within-group policy converges to $e_i = g_{\boldsymbol{\theta}}(z_i)$7. The regret benchmark is the best policy that satisfies both constraints. A central technical result is the decomposition

$e_i = g_{\boldsymbol{\theta}}(z_i)$8

which separates regret due to excess pulls of suboptimal groups from regret due to imperfect within-group meritocracy.

The algorithm BF-UCB implements this structure with two nested layers. At the outer layer, it maintains an under-exposed group set

$e_i = g_{\boldsymbol{\theta}}(z_i)$9

prioritizing the group with the largest deficit whenever the anytime fairness constraint is at risk. When all groups satisfy their quotas, it runs a UCB-style optimistic group-selection rule. At the inner layer, conditional on a chosen group, it applies an Exposure subroutine that estimates arm means, constructs confidence regions, computes an optimistic vector $e_u = g_{\boldsymbol{\theta}}(z_u)$0, and samples arms according to the induced merit-proportional distribution. The paper proves anytime group exposure fairness, asymptotic meritocratic fairness within each group, and an overall regret bound of $e_u = g_{\boldsymbol{\theta}}(z_u)$1, with experiments showing sublinear regret and better group and individual exposure guarantees than single-level baselines (Pokhriyal et al., 2024).

4. Related bi-level and biologically informed fairness formulations

Several adjacent lines of work do not use the name BiFair but share its methodological structure. A prominent example is "Meta Balanced Network for Fair Face Recognition," which studies skin-tone-related bias in deep face recognition. That work introduces the Identity Shades benchmark, skin-tone-aware training datasets, and a meta-learning algorithm that learns adaptive margins in large-margin losses. Fairness is evaluated through per-bin verification accuracy together with the Standard Deviation of accuracies across skin-tone bins and the Skewed Error Ratio,

$e_u = g_{\boldsymbol{\theta}}(z_u)$2

Its optimization is explicitly bilevel: model parameters are updated on biased training data, while group-dependent margins are updated by minimizing a meta skewness loss on a clean and unbiased meta set via second-order differentiation. This suggests a biologically grounded analogue of BiFair in which fairness is steered by an outer meta-objective rather than enforced by a single in-processing constraint (Wang et al., 2022).

A more explicit biological-fairness formulation appears in "Component-Based Fairness in Face Attribute Classification with Bayesian Network-informed Meta Learning." That paper defines face component fairness by treating morphological attributes such as big lips, arched eyebrows, big nose, double chin, and no beard as sensitive attributes. Its main fairness metric is mean True Positive Rate Disparity,

$e_u = g_{\boldsymbol{\theta}}(z_u)$3

with a companion Disparate Impact Gap metric. The proposed BNMR method uses a Bayesian Network calibrator to model dependencies among component attributes, labels, and predictions, then performs meta-learning-based sample reweighting. The empirical finding that improving component-level fairness also improves gender fairness supports a broader interpretation in which “BiFair” can refer not only to bi-level optimization but also to biologically informed fairness objectives (Liu et al., 3 May 2025).

In classification under corrupted supervision, "Bias-Tolerant Fair Classification" provides another relevant pattern. Its B-FARL objective is designed to tolerate both label bias and selection bias, with a decomposition into a clean-risk term, a fairness regularization term based on subgroup risk equality, and a bias regularization term. The method then uses a meta-learning framework to optimize group-specific coefficients in this loss. This line of work is not named BiFair, but it exemplifies the same idea that fairness correction may need to operate on more than one latent source of distortion rather than on a single observed fairness constraint (Zhang et al., 2021).

For biometric verification, "Fairness in Biometrics: a figure of merit to assess biometric verification systems" contributes a threshold-aware evaluation formalism rather than a training method. Its Fairness Discrepancy Rate is

$e_u = g_{\boldsymbol{\theta}}(z_u)$4

where $e_u = g_{\boldsymbol{\theta}}(z_u)$5 and $e_u = g_{\boldsymbol{\theta}}(z_u)$6 are the maximum discrepancies in false match rate and false non-match rate across demographic groups at a single global threshold $e_u = g_{\boldsymbol{\theta}}(z_u)$7. This work is methodologically orthogonal to BiFair, but it clarifies that fairness in biometric systems is often operationalized as single-threshold group error-rate parity rather than representation balancing or nested optimization (Pereira et al., 2020).

5. Structural fairness and bi-side constraints in graph mining

A structurally different but conceptually adjacent use of the “bi-fair” idea appears in fairness-aware maximal biclique enumeration on attributed bipartite graphs. The paper "Fairness-aware Maximal Biclique Enumeration on Bipartite Graphs" introduces single-side fair bicliques and bi-side fair bicliques, where fairness is defined by lower bounds and bounded count disparity across attribute groups on one side or both sides of a biclique.

For a single-side fair biclique $e_u = g_{\boldsymbol{\theta}}(z_u)$8, the required conditions are

$e_u = g_{\boldsymbol{\theta}}(z_u)$9

and

$s(e_u,e_i)$0

The bi-side model imposes the symmetric conditions on $s(e_u,e_i)$1 as well. The paper also defines proportion-based extensions requiring

$s(e_u,e_i)$2

To make enumeration tractable, the authors develop pruning techniques based on fair $s(e_u,e_i)$3-$s(e_u,e_i)$4 cores, colorful fair $s(e_u,e_i)$5-$s(e_u,e_i)$6 cores, and ego colorful $s(e_u,e_i)$7-cores on a 2-hop graph, followed by branch-and-bound algorithms FairBCEM, BFairBCEM, and faster combinatorial variants. The output-space explosion is substantial: on IMDB, with $s(e_u,e_i)$8, the reported number of maximal bicliques is $s(e_u,e_i)$9, while the number of single-side fair bicliques is $\hat y_{u,i} = s(e_u,e_i)$0; with $\hat y_{u,i} = s(e_u,e_i)$1, the number of maximal bicliques is $\hat y_{u,i} = s(e_u,e_i)$2, while the number of bi-side fair bicliques is $\hat y_{u,i} = s(e_u,e_i)$3. This is not BiFair in the recommender-system sense, but it is a clear example of bi-side fairness becoming a first-class structural constraint rather than a post hoc metric (Yin et al., 2023).

6. Distinctions, limitations, and open directions

A recurring misconception is that BiFair names a single universal fairness doctrine. The literature instead supports a narrower conclusion: the term designates, or is used to interpret, architectures in which fairness is decomposed across two loci of intervention or evaluation. In LLM-enhanced recommendation, the two loci are prior unfairness in $\hat y_{u,i} = s(e_u,e_i)$4 and training unfairness in $\hat y_{u,i} = s(e_u,e_i)$5; in stochastic bandits, they are group exposure and within-group meritocracy; in biometrics, related work separates fairness-aware evaluation, demographic fairness, and component-level or skin-tone-level bias; in graph mining, fairness can be one-side or bi-side and may be enforced through count or ratio constraints rather than through predictive loss (Zhang et al., 6 Jul 2025, Pokhriyal et al., 2024, Liu et al., 3 May 2025, Pereira et al., 2020, Yin et al., 2023).

The main technical limitations are domain specific. In recommender systems, the bi-level optimization adds overhead, depends on group labels such as popularity buckets or genres, keeps the LLM itself frozen, and may become heavy when optimizing representations $\hat y_{u,i} = s(e_u,e_i)$6 for very large item catalogs. In stochastic bandits, the group-level guarantees require a specified fairness vector $\hat y_{u,i} = s(e_u,e_i)$7, while within-group meritocracy is a limit property rather than an exact finite-time equality. In face-recognition and face-attribute settings, meta-learning approaches require sensitive or component labels and often a clean, balanced meta set; intersectional extensions remain largely open. In biometric verification, fairness metrics such as FDR remain evaluation criteria rather than direct mitigation methods (Zhang et al., 6 Jul 2025, Pokhriyal et al., 2024, Wang et al., 2022, Liu et al., 3 May 2025, Pereira et al., 2020).

A plausible implication is that BiFair is best understood as a modular research program rather than a single algorithmic family. The common ingredients are explicit decomposition of unfairness, coupled optimization or search across those components, and fairness metrics that are local to the task—group utility in recommendation, exposure and merit in bandits, error-rate parity in biometrics, or balanced representation in graph patterns. Current work suggests several extensions: broader group definitions, including intersectional groups; simultaneous user-side and item-side fairness in recommender systems; richer fairness objectives in bandits beyond exposure; and integration of biologically grounded fairness with demographic auditing in facial biometrics (Zhang et al., 6 Jul 2025, Pokhriyal et al., 2024, Liu et al., 3 May 2025).