FARE: Deep Reinforcement Learning For Fair Exposure Constrained Uncertainty Aware Financial Content Personalization

Arundeep Chinta; Lucas Vinh Tran; Jay Katukuri

cs.LGSep 25, 2026

FARE: Deep Reinforcement Learning For Fair Exposure Constrained Uncertainty Aware Financial Content Personalization

Authors: Arundeep Chinta, Lucas Vinh Tran, Jay Katukuri

Organizations: JPMorganChase Palo Alto, CA, USA · JPMorganChase London, UK

Abstract

Content personalization systems in financial services must ensure fair exposure across diverse offerings-a requirement driven by contractual obligations and the need to prevent "rich-get-richer" dynamics where content with high click-through rate (CTR) dominates while other relevant products receive minimal visibility. Share of Voice (SOV) constraints, which guarantee each content category a target fraction of top-position exposure, address this by promoting product diversity and balanced user discovery. While re-ranking layers atop CTR models are common in practice, we propose two key novelties: (1) framing SOV-constrained ranking as a deep reinforcement learning problem analogous to constrained trade execution in algorithmic finance, and (2) explicitly incorporating CTR prediction uncertainty into the agent's state space and policy design-enabling larger ranking adjustments for high-uncertainty predictions where deviation from CTR-optimal ordering is less costly. We introduce FARE (Fair Ranking Executor), a modular uncertainty-aware execution layer that translates any black-box CTR model's predictions into SOV-fair rankings without retraining the underlying model. Our uncertainty-weighted proportional control policy (FARE-PC) and learned neural policies (FARE-ES, FARE-PPO) demonstrate that uncertainty-aware approaches can substantially reduce SOV deviation from fairness targets while minimizing engagement loss, with gradient-free evolution strategies outperforming policy gradient methods on synthetic data and the ordering reversing on KuaiRand-Pure.

Figures & tables

Figure 1: Parallel architectures in (a) algorithmic trading and (b) content personalization, illustrating the signal

\rightarrow

execution decoupling pattern that FARE adapts for SOV-constrained re-ranking.

Method	SOV Error $\downarrow$	Reduction $\uparrow$	PWCL% $\downarrow$	Pos. Disp.	Kendall $\tau$	Top-1 Chg
CTR-Only (= TS)	0.247 $\pm$ 0.002	–	0.00%	0.00	0.00	0.0%
Max-Deficit	0.000 $\pm$ 0.000	100.0%	11.74 $\pm$ 0.29%	–	–	–
Quota-CTR	0.073 $\pm$ 0.008	70.4%	9.89 $\pm$ 0.11%	1.39 $\pm$ 0.01	0.42 $\pm$ 0.00	23.4 $\pm$ 1.3%
FARE-PC	0.024 $\pm$ 0.000	90.3%	2.64%	1.58	0.49	79.3%
FARE-ES	0.034 $\pm$ 0.004	86.2%	2.79 $\pm$ 0.2%	1.58 $\pm$ 0.001	0.49 $\pm$ 0.001	79.4 $\pm$ 0.1%
FARE-PPO	0.155 $\pm$ 0.080	37.2%	1.50 $\pm$ 0.1%	1.56 $\pm$ 0.02	0.48 $\pm$ 0.01	78.5 $\pm$ 0.7%

Table 1: Synthetic test set results (mean

\pm

std over 3 seeds). Target SOV: 20% per tile. Upper block: non-FARE baselines. Lower block: FARE variants. FARE-PC’s variation across seeds is negligible on PWCL and the ranking-stability columns and is omitted; its SOV Error std rounds to 0.000.

Method	SOV Error $\downarrow$	Reduction $\uparrow$	PWCL% $\downarrow$	Pos. Disp.	Kendall $\tau$	Top-1 Chg
CTR-Only	0.122 $\pm$ 0.005	–	0.0%	0.00	0.00	0.0%
FARE-PC	0.010 $\pm$ 0.001	91.8%	0.19 $\pm$ 0.01%	1.59 $\pm$ 0.003	0.497 $\pm$ 0.001	79.8 $\pm$ 0.2%
FARE-ES	0.099 $\pm$ 0.009	18.9%	0.06 $\pm$ 0.01%	1.59 $\pm$ 0.004	0.494 $\pm$ 0.001	79.5 $\pm$ 0.2%
FARE-PPO	0.089 $\pm$ 0.032	27.0%	0.10 $\pm$ 0.08%	1.59 $\pm$ 0.006	0.495 $\pm$ 0.003	79.6 $\pm$ 0.3%

Table 3: KuaiRand-Pure evaluation metrics (mean

\pm

std). Target SOV: 20% per tile.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A1: Synthetic CTR estimates $(\mu,\sigma^{2})$ and resulting SOV imbalance. Top row: (a) Position heatmap. (b) Position 1 SOV distribution. (c) Base $\mu$ and $\sigma^{2}$ per tile. Bottom row: (d–f) User-level variation in $\mu$ , $\sigma^{2}$ , and mean-variance space across 10,000 users.

Figure A2: KuaiRand-Pure MLP predictions. Top row: (a) Position heatmap shows non-monotonic ordering (T1 $\approx$ T3 $>$ T2 $\approx$ T4 $>$ T5). (b) Position 1 SOV reveals a milder fairness gap than synthetic data. (c) Per-tile median $\mu$ and $\sigma^{2}$ . Bottom row: (d) $\mu$ distribution across 8,614 users (Day 1). (e) $\sigma^{2}$ distribution. (f) Mean-variance space.

	T1	T2	T3	T4	T5	SOV Error	Reduction
CTR-Only	23.4	19.2	22.7	19.6	15.1	0.122	–
FARE-PC	20.3	19.9	20.2	19.9	19.7	0.010	91.8%
FARE-ES	22.1	18.9	22.7	19.6	16.7	0.099	18.9%
FARE-PPO	21.8	19.2	21.5	19.9	17.6	0.089	27.0%
Target	20.0	20.0	20.0	20.0	20.0	0.000	–

Appendix

Table A1: KuaiRand-Pure SOV distribution on test set (%), one representative seed. Target SOV: 20% per tile; T1–T5 denote the five category tiles.

Method	SOV Error $\downarrow$	PWCL% $\downarrow$	Pos. Disp.	Kendall $\tau$	Top-1 Chg
CTR-Only	0.547 $\pm$ 0.003	0.0%	0.00	0.00	0.0%
FARE-PC	0.116 $\pm$ 0.001	5.31 $\pm$ 0.02%	1.62 $\pm$ 0.001	0.508 $\pm$ 0.000	80.8 $\pm$ 0.1%
FARE-ES	0.120 $\pm$ 0.003	4.92 $\pm$ 0.18%	1.61 $\pm$ 0.003	0.505 $\pm$ 0.001	80.6 $\pm$ 0.1%
FARE-PPO	0.367 $\pm$ 0.144	2.24 $\pm$ 1.59%	1.57 $\pm$ 0.023	0.490 $\pm$ 0.009	79.0 $\pm$ 0.9%

Appendix

Table A2: Non-uniform SOV targets (synthetic): test set evaluation metrics (mean

\pm

std). Targets:

\tau=[10,15,20,25,30]\%

	T1	T2	T3	T4	T5	SOV Error	Reduction
CTR-Only	28.0	23.6	20.7	16.1	11.5	0.547	–
FARE-PC	14.6	16.2	19.0	22.8	27.4	0.116	78.8%
FARE-ES	16.0	14.1	19.2	24.0	26.7	0.120	78.1%
FARE-PPO	25.6	16.2	19.1	20.7	18.5	0.367	32.9%
Target	10.0	15.0	20.0	25.0	30.0	0.000	–

Appendix

Table A3: Non-uniform SOV targets (synthetic): final SOV distribution on test set (%), one representative seed. Targets:

\tau=[10,15,20,25,30]\%

		FARE-PC			FARE-ES			FARE-PPO
Scheme	Base	Err.	Red.	PWCL	Err.	Red.	PWCL	Err.	Red.	PWCL
Top-1	.247	.024 $\pm$ .000	90.3%	2.64%	.034 $\pm$ .004	86.2%	2.79%	.155 $\pm$ .080	37.2%	1.50%
Top-3	.137	.040 $\pm$ .001	70.8%	2.27%	.036 $\pm$ .005	73.7%	2.39%	.114 $\pm$ .023	16.8%	0.63%
DCG	.061	.032 $\pm$ .000	47.5%	1.46%	.067 $\pm$ .005	$-$ 9.8%	$-$ 0.32%	.063 $\pm$ .001	$-$ 3.3%	$-$ 0.08%

Appendix

Table A4: Multi-position SOV: performance across weighting schemes (uniform target

\tau{=}20\%

). Mean

\pm

std over seeds.

		FARE-PC			FARE-ES			FARE-PPO
Scheme	Base	Err.	Red.	PWCL	Err.	Red.	PWCL	Err.	Red.	PWCL
Top-1	.122	.010 $\pm$ .001	91.8%	0.19%	.099 $\pm$ .009	18.9%	0.06%	.089 $\pm$ .032	27.0%	0.10%
Top-3	.063	.017 $\pm$ .001	72.6%	0.16%	.055 $\pm$ .012	12.7%	0.04%	.047 $\pm$ .016	25.4%	0.09%
DCG	.030	.015 $\pm$ .001	49.6%	0.11%	.029 $\pm$ .001	3.3%	0.01%	.025 $\pm$ .009	16.7%	0.05%

Appendix

Table A5: KuaiRand-Pure multi-pos SOV: perf. across wtg. schemes (unif. target

\tau{=}20\%

). Mean

\pm

std.

	FARE-PC			FARE-ES			FARE-PPO
$\boldsymbol{\lambda_{\textbf{div}}}$	SOV $\downarrow$	PWCL	Div $\uparrow$	SOV $\downarrow$	PWCL	Div $\uparrow$	SOV $\downarrow$	PWCL	Div $\uparrow$
0.0	.024	2.64%	2.383	.034	2.71%	2.398	.172	1.34%	2.339
0.5	.010	2.83%	2.402	.048	2.85%	2.409	.099	2.41%	2.414
1.0	.043	3.03%	2.405	.049	2.83%	2.409	.122	2.28%	2.416
2.0	.175	3.87%	2.427	.063	2.86%	2.415	.120	1.86%	2.399
5.0	.632	6.69%	2.471	.088	3.10%	2.427	.119	2.19%	2.414

Appendix

Table A6: Multi-objective

\lambda_{\text{div}}

sweep: SOV Error, PWCL, and Category Diversity across methods. CTR-Only baseline: SOV Err = 0.247, Diversity = 2.300.

		FARE-PC			FARE-ES			FARE-PPO
K	Base SOV Err.	Err. $\downarrow$	Red.	PWCL	Err. $\downarrow$	Red.	PWCL	Err. $\downarrow$	Red.	PWCL
3	0.252	.014 $\pm$ .000	94.4%	2.45%	.030 $\pm$ .012	88.1%	2.41%	.169 $\pm$ .086	32.9%	0.96%
5	0.247	.024 $\pm$ .000	90.3%	2.64%	.034 $\pm$ .004	86.2%	2.79%	.155 $\pm$ .080	37.2%	1.50%
7	0.249	.039 $\pm$ .001	84.3%	2.73%	.055 $\pm$ .003	77.9%	2.78%	.158 $\pm$ .047	36.5%	1.82%
10	0.253	.058 $\pm$ .001	77.1%	2.87%	.066 $\pm$ .026	73.9%	3.49%	.251 $\pm$ .012	0.8%	$-$ 0.11%

Appendix

Table A7: Performance as

K

(number of tiles) increases. Uniform targets (

\tau=1/K

). Mean

\pm

std over seeds.

Figure A3: FARE-PC

\alpha

sensitivity under uniform SOV targets (20%). (a) SOV Error decreases monotonically with

\alpha

. (b) PWCL cost increases with diminishing marginal cost. (c) Per-tile SOV converges to target. Dashed line: CTR-Only baseline error (0.247).

Figure A4: FARE-PC SOV Error vs CTR-Only across uncertainty scale factors

s\in\{0.25\times\text{--}4\times\}

Learning to Rank

Learning to rank develops models that order items by relevance or utility for search, recommendations, and other decisions.

See all 273 papers

Discrepancy-Rounded Fair Bandits with Static and Time-Varying Exposure Floors
Jul 24, 2026Ibne Farabi Shihab, Joyanta Jyoti Mondal, Anuj SharmaMulti-Armed Bandits Fair Resource Allocation
Department of Computer Science, Iowa State University, USA · Department of Computer and Information Sciences, University of Delaware, USA · Department of Civil, Construction & Environmental Engineering, Iowa State University, USA
Representation Curriculum: Stagewise Training for Robust Ranking and Allocation
Jun 3, 2026Ehsan Ebrahimzadeh, Sina Baharlouei, Abraham BagherjeiranCold-Start Recommendation Representation Learning
eBay Search Ranking and Monetization San Jose, California, USA
Exposure-Based Reinforcement Learning to Rank
Jul 21, 2026Harrie Oosterhuis, Rolf Jagerman, Zhen Qin +1Reinforcement Learning Learning to Rank
University of Amsterdam · Google DeepMind

Jul 24, 2026J/K move · Enter open · S save

Discrepancy-Rounded Fair Bandits with Static and Time-Varying Exposure Floors

Ibne Farabi Shihab, Joyanta Jyoti Mondal, Anuj Sharma

Department of Computer Science, Iowa State University, USA · Department of Computer and Information Sciences, University of Delaware, USA · Department of Civil, Construction & Environmental Engineering, Iowa State University, USA

Multi-Armed Bandits Fair Resource Allocation Group Fairness

Open paper AskPDF arXiv

Minimum-exposure constraints arise in recommendation, content curation, and regulated allocation when each provider, arm, or group must receive guaranteed exposure inside a period rather than only in aggregate. We study stochastic bandits with exact exposure floors and show that the right object is a rounding problem: a fractional fair schedule is realized as integral pulls, and the exposure error is exactly a discrepancy vector. The main contribution is a blockwise model with time-varying floors. BDQ-UCB satisfies every block floor deterministically and has fair regret governed by the nonmandatory budget

R

, not the horizon

T

, with high-probability regret

O(\sqrt{KR\log(KT)})

. A MOSS residual variant attains

O(\sqrt{KR})

, and a matching lower bound gives the minimax rate

Θ(\sqrt{KR})

, even with positive mandatory exposure; a kl-UCB

^{++}

residual rule adds instance-dependent optimality. The formulation becomes essential for overlapping group floors: per-arm rounding can violate a group constraint by

Ω(s)

in the group size, whereas Beck--Fiala null-space rounding meets every group floor within the block budget with violation below the arm degree

t

, and composes with UCB at the same

R

-parametrized regret. For learned group plans, we close disjoint systems at

\widetildeΘ(\sqrt{KT})

, give a dual-ledger decomposition explaining why naive index rules fail under overlap, and prove a plan-sampling rule that is pathwise feasible under an initial cover-slack condition and attains a conditional

\widetilde O(\sqrt{KT})

guarantee, leaving the condition-free overlap rate open. Experiments on synthetic floors, MovieLens-100k genre exposure, and deployment stress tests show exact feasibility without penalty tuning and regret competitive with tuned Lagrangian baselines.

Abstract

Learning to Rank

Explore similar work

Discrepancy-Rounded Fair Bandits with Static and Time-Varying Exposure Floors

Representation Curriculum: Stagewise Training for Robust Ranking and Allocation

Exposure-Based Reinforcement Learning to Rank