User Fairness Counter Mechanisms

• User Fairness Counter is a mechanism that quantifies user-level fairness via evaluation metrics, service ledgers, or control variables.
• They are applied in systems like recommender evaluations, LLM serving, and online scheduling to dynamically monitor and enforce fairness.
• These counters guide decision-making by integrating fairness measurements into runtime scheduling, algorithm updates, and privacy-preserving audits.

“User Fairness Counter” denotes a family of quantitative mechanisms that track, summarize, or control fairness at the user or client level. In recent literature, the term does not refer to a single canonical construct. It can denote an evaluation metric, such as Pairwise User unFairness (PUF) for recommender systems; a runtime scheduling ledger, such as the Virtual Token Counter (VTC), the User Fairness Counter (UFC), or the Weighted Service Counter (WSC) in LLM serving; a per-user fairness ratio or deprivation index in online scheduling; or a monitoring pipeline that counts outcomes by protected group under privacy constraints (Rampisela et al., 23 Jan 2026, Sheng et al., 2023, Wei et al., 19 Aug 2025, Khan et al., 2024, Dwibedy et al., 2020, He et al., 2 Feb 2026). This suggests that the expression is best understood functionally: a user fairness counter is a variable, score, or protocol that makes user-side fairness observable and actionable under a specified fairness notion.

1. Conceptual scope and core forms

Across the cited work, fairness is instantiated in several technically distinct ways. In recommender systems, individual user fairness is framed as treating similar users similarly in recommendation effectiveness, rather than merely in list appearance (Rampisela et al., 23 Jan 2026). In multi-tenant LLM serving, fairness is framed as comparable service received by backlogged clients, measured in weighted processed tokens and sometimes latency-adjusted quality of service (Sheng et al., 2023, Wei et al., 19 Aug 2025). In online scheduling, fairness is defined relative to each user’s own objective value, with per-user fairness ratios, deprivation indices, and global aggregates (Dwibedy et al., 2020). In fairness monitoring, the counter becomes a group-statistical audit instrument that aggregates outcomes such as selection rates or mean outcomes across protected groups (He et al., 2 Feb 2026).

Setting	Counter quantity	Function
Recommender evaluation	PUF	Similarity-weighted effectiveness disparity
LLM serving	VTC / UFC / WSC	Per-client cumulative service ledger
Online scheduling	$RF$, $UDI$, $RDI$	Objective-relative user fairness and deprivation
Fairness monitoring	$\mu(G)$, $\delta$	Group outcome counting and disparity auditing
Post-training control	$\mathcal{L}_{\text{fair}}^{(t)}(u)$	User-level fairness surrogate across control levels

A common structural feature is that the counter is not merely descriptive. It often participates directly in decision-making. VTC schedules the client with the smallest virtual counter, UFC enters a holistic fairness score used for batch admission, and adaptive fairness-control methods update dual variables or level-conditioned representations based on measured fairness status (Sheng et al., 2023, Wei et al., 19 Aug 2025, Lu et al., 17 Feb 2026, Chen et al., 28 Jan 2026).

2. Measurement in recommendation and ranking

The clearest metric-style user fairness counter is PUF, introduced for recommender systems under an attribute-free individual fairness setting. The paper starts from the condition

$D(M(L_u), M(L_{u'})) \leq d(u,u'),$

then operationalizes unfairness as

$$\text{PUF} = \frac{2}{m(m-1)} \sum_{u\in U}\sum_{u'\in U\setminus\{u\}} sim(u,u') \times |S(u) - S(u')|.$$

Here $sim(u,u') \in [0,1]$ is user similarity and $S(u)$ is a per-user effectiveness score such as P@$UDI$0 or NDCG@$UDI$1, scaled to $UDI$2. Lower is better. The crucial design point is that unfairness is high when similar users receive very different effectiveness, and low when either users are dissimilar or their effectiveness is close. The paper argues that standard deviation, Gini index, and envy-based measures ignore user similarity, whereas UF ignores recommendation relevance; PUF is presented as the first measure to jointly capture both dimensions (Rampisela et al., 23 Jan 2026).

Related recommendation work uses different user-side counters. In two-sided recommendation with item-fairness constraints, user fairness is defined by normalized user utility

$UDI$3

so the counter is the worst-off user’s utility relative to that user’s own ideal recommendation. This formulation is used to study the “price of item fairness on user fairness” and to show that diverse preferences can create “free” item and user fairness, whereas users with misestimated preferences can be especially disadvantaged by item-fairness constraints (Greenwood et al., 2024).

Group-conditioned ranking work shifts the counter from individual users to user groups. In two-sided market ranking, group-specific expected utility

$UDI$4

and social-welfare aggregation

$UDI$5

serve as the user-side fairness quantities. Empirically, the paper uses ratios such as $UDI$6 as a direct disparity signal. Counterpart Fairness moves in a different direction: it restricts fairness evaluation to matched cross-group counterparts and computes metrics such as counterpart demographic parity gap on those matched sets, thereby turning the fairness counter into a matched-comparison statistic rather than a whole-population average (Wang et al., 2020, Wang et al., 2023).

3. Runtime service ledgers in LLM serving and communication systems

In LLM serving, the user fairness counter becomes a live scheduling state. VTC defines per-client service with token-based cost functions, most prominently

$UDI$7

maintains a virtual counter $UDI$8 for each client, and always admits requests from the client with the smallest counter. Input-token cost is charged when a request is admitted to the running batch, output-token cost is charged during each decode step, and a “counter lift” prevents idle clients from carrying stale credit into later overload periods. The central guarantee is that for any two continuously backlogged clients,

$UDI$9

with a matching lower-bound argument showing the $RDI$0 worst-case factor is tight under work conservation and no preemption (Sheng et al., 2023).

Equinox generalizes the same idea toward latency-aware QoS. Its UFC is updated using a weighted-token quantity in the numerator and a latency compensation term in the denominator involving $RDI$1, $RDI$2, and $RDI$3 in the reported implementation. Output tokens are weighted four times more than input tokens, reflecting the prefill/decode asymmetry. UFC is then combined with the Resource Fairness Counter through

$RDI$4

and the scheduler picks the client with minimum HF. FairServe adopts a related but application-aware design: its WSC maintains a per-user service counter $RDI$5, initialized to $RDI$6, and increments it by weighted input, system prompt, and output tokens normalized by the expected token budget of the corresponding application stage, with $RDI$7, $RDI$8, and $RDI$9. In both systems, the counter is not a post hoc fairness score but the priority variable driving online admission and scheduling (Wei et al., 19 Aug 2025, Khan et al., 2024).

Communication-network work replaces service volume by service recurrence. Channel cycle time (CCT) defines a user’s cycle as the duration between two refresh moments of that user such that all other users have successfully accessed the channel at least once in between. The resulting quantity is a per-user “time-until-fair-service-recurrence” counter. It is explicitly designed to capture short-term fairness, delays, and jitter, and the paper shows analytically that CSMA/CA exhibits superior short-term fairness over slotted Aloha under this metric (Shen et al., 2023).

4. Counters as control variables and training-time regularizers

A second major usage treats the counter as a control signal inside optimization. In fairness-controlled multi-user beamforming, fairness is measured by Jain’s index over user rates,

$\mu(G)$0

with a target lower bound $\mu(G)$1. The training loss is

$\mu(G)$2

and the adaptive multiplier is updated by

$\mu(G)$3

When fairness falls below target, $\mu(G)$4 increases; when fairness exceeds target, $\mu(G)$5 decreases. The counterbalance is therefore implemented by a monitored fairness signal, a threshold, and a corrective dual update (Lu et al., 17 Feb 2026).

Cofair applies the same control logic to recommendation, but with post-training flexibility. After a single training run, it can generate user embeddings $\mu(G)$6 for fairness levels $\mu(G)$7 by combining a shared representation with fairness-conditioned adapters. Its explicit user-level fairness surrogate is

$\mu(G)$8

and progressive user-wise fairness is encouraged by

$\mu(G)$9

The paper states that the adversarial objective upper bounds demographic parity and that the regularization term enforces progressive fairness at user level (Chen et al., 28 Jan 2026).

Training-time user fairness control also appears in recommender debiasing by representation alignment. In-UCDS defines user-oriented fairness as parity of expected recommendation performance across advantaged and disadvantaged users,

$\delta$0

measured in practice by

$\delta$1

Its fairness loss pulls each disadvantaged user embedding toward the mean embedding of a constrained dominant set of similar advantaged users, and the complete training objective is $\delta$2. Here the counter is both the measured gap $\delta$3 and the in-processing fairness regularizer that counters biased learning dynamics (Han et al., 2023).

5. Auditing, monitoring, and governance

In fairness auditing, the counter is often a privacy-preserving monitoring pipeline rather than a scheduler variable. For post-deployment fairness monitoring, one paper studies a secure two-party computation workflow in which a user’s sensitive attribute $\delta$4 is secret-shared as

$\delta$5

and reconstructed only inside MPC. Group statistics are then computed as

$\delta$6

In this setting, the fairness counter is a legally and cryptographically constrained audit mechanism that aggregates outcomes by protected group without exposing raw attributes (He et al., 2 Feb 2026).

The same paper emphasizes that feasibility depends on user acceptance of the monitoring protocol. Acceptance was higher for “Anonymize and encrypt personal data, then distribute the encrypted data so no single party has full access to sensitive personal data” than for “Anonymize and encrypt the personal data,” and users preferred research centers or auditing agencies/regulators over companies developing automated hiring systems. The implication is that a fairness counter that depends on user-contributed sensitive data is only deployable when its institutional design, privacy mechanism, storage arrangement, and fairness objective are acceptable to the people whose data sustain it (He et al., 2 Feb 2026).

A neighboring but distinct temporal use appears in BankFair. There the monitored quantity is provider-side rather than user-side fairness: the remaining required minimum exposure

$\delta$7

acts as a fairness deficit ledger, and the Talmud rule distributes this remaining obligation across future periods according to predicted user traffic. Although this is not a user fairness counter, it illustrates a broader counter architecture in which fairness is represented as debt, carryover, and repayment over time rather than instantaneous disparity (Ye et al., 2024).

6. Limitations, distinctions, and recurrent design choices

A central limitation is that a fairness counter is only as meaningful as its underlying fairness notion. PUF depends critically on how user similarity is computed and on the quality of per-user effectiveness estimates; poor similarity functions or noisy relevance judgments can make the score misleading (Rampisela et al., 23 Jan 2026). Representation-based control methods such as Cofair optimize a user-level adversarial surrogate rather than a direct per-user ranking-fairness outcome, so the “counter” is a proxy for fairness in representation space, not a universal fairness observable (Chen et al., 28 Jan 2026). In-UCDS depends on activity-based advantaged/disadvantaged partitioning and on the existence of similar advantaged users from whom disadvantaged users can learn (Han et al., 2023).

A second distinction is between diagnosis and remedy. PUF is explicitly an evaluation measure, not a training algorithm that directly fixes unfairness (Rampisela et al., 23 Jan 2026). The MPC study analyzes stated acceptance of fairness-monitoring designs, not actual donation behavior in a deployed system (He et al., 2 Feb 2026). VTC, UFC, WSC, and adaptive dual-variable schemes do intervene online, but they optimize the specified service or fairness objective rather than fairness in every broader ethical sense (Sheng et al., 2023, Wei et al., 19 Aug 2025, Lu et al., 17 Feb 2026).

A third limitation is computational and operational. Pairwise user fairness metrics are naturally $\delta$8 in the number of users, although PUF is reported to remain much faster than several existing pairwise fairness measures (Rampisela et al., 23 Jan 2026). LLM-serving counters rely on unknown output lengths, dynamic batching, and capacity variation; prediction quality directly affects fairness quality in systems such as Equinox (Wei et al., 19 Aug 2025). Temporal fairness ledgers such as BankFair require traffic forecasting, and mispredicted future capacity can distort the scheduled fairness obligation (Ye et al., 2024).

This recurring pattern suggests three dominant forms of user fairness counter. The first is the evaluator, which outputs a scalar unfairness or fairness score from observed outcomes. The second is the service ledger, which accumulates user-side service and drives runtime scheduling. The third is the control variable, which updates penalties, adapters, or regularizers to move the system along a fairness–accuracy or fairness–throughput frontier. The literature does not collapse these forms into one definition; instead, it treats the counter as the mechanism by which a chosen fairness notion becomes measurable, enforceable, or both.