VALOR: Value-Aware Revenue Uplift Modeling with Treatment-Gated Representation for B2B Sales

Vamshi Guduguntla; Kavin Soni; Debanshu Das

cs.LGApr 2, 2026

VALOR: Value-Aware Revenue Uplift Modeling with Treatment-Gated Representation for B2B Sales

Authors: Vamshi Guduguntla, Kavin Soni, Debanshu Das

Organizations: Google, California, USA

Abstract

B2B sales organizations must identify "persuadable" accounts within zero-inflated revenue distributions to optimize expensive human resource allocation. Standard uplift frameworks struggle with treatment signal collapse in high-dimensional spaces and a misalignment between regression calibration and the ranking of high-value "whales." We introduce VALOR (Value Aware Learning of Optimized (B2B) Revenue), a unified framework featuring a Treatment-Gated Sparse-Revenue Network that uses bilinear interaction to prevent causal signal collapse. The framework is optimized via a novel Cost-Sensitive Focal-ZILN objective that combines a focal mechanism for distributional robustness with a value-weighted ranking loss that scales penalties based on financial magnitude. To provide interpretability for high-touch sales programs, we further derive Robust ZILN-GBDT, a tree based variant utilizing a custom splitting criterion for uplift heterogeneity. Extensive evaluations confirm VALOR's dominance, achieving a 20% improvement in rankability over state-of-the-art methods on public benchmarks and delivering a validated 2.7x increase in incremental revenue per account in a rigorous 4-month production A/B test.

Figures & tables

Figure 1: The VALOR training pipeline processes features ( $X$ ) and treatment ( $T$ ) through three stages: (1) A Treatment-Gated Interaction Module uses a sigmoid gate to dynamically re-weight features, mitigating the “vanishing treatment” signal; (2) Sparse-Revenue Mixture Heads decouple the decision ( $\pi$ ) and value ( $\mu,\sigma$ ) processes to model zero-inflated revenue; and (3) A Hybrid Objective combines Focal-ZILN and Value-Weighted Ranking losses to ensure distributional robustness and maximize top-decile revenue capture.

	Synthetic Dataset					Production Dataset
Models	AUUC	Qini	Lift@30	KRCC	Inf(ms)	AUUC	Qini	Lift@30	KRCC	Inf(ms)
Baselines
T-Learner	0.1698	0.1643	13.93	0.0453	0.0214	0.8854	0.8556	36.54	0.1654	0.0321
TARNet	0.2566	0.2530	15.72	0.0762	0.0189	0.9125	0.8864	36.92	0.1689	0.0324
Causal Forest	0.1413	0.1381	11.67	0.0279	0.0017	0.6552	0.6258	30.13	0.1202	0.0026
DragonNet	0.2571	0.2489	15.77	0.0897	0.0187	1.0520	1.0210	41.25	0.1950	0.0328

Table 1: Overall performance comparison of the proposed VALOR framework against state-of-the-art baselines (including RERUM) on Synthetic and Production datasets. The best results in each group (Baselines vs. Proposed) are underlined . VALOR variants consistently outperform their respective backbones. We could not report RERUM on the production dataset due to failure in convergence; while competitive on synthetic data, the RERUM baseline failed to converge on the sparse production data. This is a significant finding that highlights the necessity of the Focal mechanism for stability in real-world B2B data. Note: Standard deviations are omitted for space but were consistently

\leq 0.02

across all metrics, confirming statistical significance.

	Synthetic Dataset				Production Dataset
Method Variants	AUUC	Qini	Lift@30	KRCC	AUUC	Qini	Lift@30	KRCC
VALOR (TARNet Backbone)
+ZILN+Focal+GTI+WR	0.2772	0.2714	16.39	0.0751	1.2145	1.1923	45.83	0.2588
+ZILN+Focal+GTI	0.2530	0.2507	16.05	0.0725	1.1852	1.1578	44.10	0.2612
+ZILN+Focal	0.2556	0.2527	16.04	0.0724	1.1644	1.1402	43.82	0.2455
Baseline (TarNet)	0.2566	0.2530	15.72	0.0762	0.9125	0.8864	36.92	0.1689

Table 2: Comprehensive Ablation Study dissecting the VALOR framework. Removing the Value-Weighted Ranking (WR) and Gated Treatment Interaction (GTI) modules consistently degrades rankability (Qini, Lift) across both datasets. Note: Standard deviations are omitted for space but were consistently

\leq 0.02

across all metrics, confirming statistical significance.

Panel of Figure 2 — Figure 2: End-to-end system architecture for the deployed VALOR framework. Program 2 utilizes a hybrid quarterly pacing strategy (VALOR prioritization followed by legacy propensity fallback). Treated accounts pass through Opportunity and Conversion gates before all accounts enter a mandatory 90-day cooling-off period.

Metric	Incumbent	Challenger	Abs. Lift	Confidence Interval
	Incumbent	VALOR	( $\Delta$ )	(Range)
Opp. Rate	9.3%	17.6%	+8.3%	$[7.1\%,9.5\%]$ (95% CI)
Incr. Revenue	$445	$1185	+$740	$[\$ 429,$1051]$ (95% CI)

Table 3: Online A/B Experiment Results. The VALOR model demonstrates statistically significant gains in both Opportunity Rate and Per-Account Revenue compared to the Incumbent model.

Uplift Modeling

See all 12 papers

UpliftBench: Revealing Outcome-Regime and Objective Mismatch in Uplift Evaluation
Aug 2, 2026Binshuang LiConditional Average Treatment Effect Estimation Benchmark Design
Independent Researcher USA
Orthogonal Uplift Learning with Permutation-Invariant Representations for Combinatorial Treatments
Feb 23, 2026Xinyan Su, Jiacan Gao, Mingyuan Ma +6Uplift Modeling Causal Effect Estimation
Didi Chuxing, Beijing, China · School of Statistics, East China Normal University, Shanghai, China · School of Mathematics and Statistics, Beijing Jiaotong University, Beijing, China +2
A Large-Scale Empirical Comparison of Meta-Learners and Causal Forests for Heterogeneous Treatment Effect Estimation in Marketing Uplift Modeling
Apr 7, 2026Aman SinghUplift Modeling Conditional Average Treatment Effect Estimation

Aug 2, 2026J/K move · Enter open · S save

UpliftBench: Revealing Outcome-Regime and Objective Mismatch in Uplift Evaluation

Binshuang Li

Independent Researcher USA

Conditional Average Treatment Effect Estimation Benchmark Design Policy Evaluation Uplift Modeling Benchmark Validity Individualized Treatment Effect Estimation

Open paper AskPDF arXiv

Uplift modeling (conditional-average-treatment-effect estimation) drives personalized targeting, yet published uplift benchmarks frequently disagree on which estimator performs best; we show the disagreement is substantially about metrics, not models. UpliftBench evaluates 12 uplift estimators under an outer-test-isolated, multi-objective protocol across seven dataset families; its two findings are identified where a reference objective exists -- F1 on the standard continuous benchmark (IHDP), F2 in a within-sample case study on Jobs. On that benchmark, Qini shows no detectable alignment with effect accuracy -- across all 100 IHDP realizations its mean rank correlation with effect accuracy is +0.07, 95% CI [-0.03, +0.16] -- while AUUC is consistently more aligned (paired prefix-mean-AUUC-over-Qini gap +0.49 [+0.40, +0.59]; the shipped cumulative-gain AUUC aligns better still, +0.73). On Jobs, ranking metrics are structurally insufficient for a sign-threshold policy because they discard the score level; empirically, within the released split-rotation analysis direct policy-risk selection yields lower benchmark regret than random model selection while Qini, AUUC, and uplift-at-

k

do not (14-15% regret). Calibrating the decision threshold removes 81% of the Qini-selection regret. Both findings are bounded, not universal: F1 is not detected on either validation family (the ACIC and Revenue-Synthetic gaps are both indistinguishable from zero), and F2 vanishes under a budgeted-value objective where rank suffices. UpliftBench releases versioned loaders, fixed protocols, result artifacts, and a reproducible living leaderboard; the public repository accompanies the paper.

Abstract

Uplift Modeling

Explore similar work

UpliftBench: Revealing Outcome-Regime and Objective Mismatch in Uplift Evaluation

Orthogonal Uplift Learning with Permutation-Invariant Representations for Combinatorial Treatments

A Large-Scale Empirical Comparison of Meta-Learners and Causal Forests for Heterogeneous Treatment Effect Estimation in Marketing Uplift Modeling