← 研踪 PaperTrailarXiv:2611.20794 · v2
Regularizing Treatment Leakage for LLM-based Uplift Estimation
首次提交 2026-11-16 17:42
版本更新 2026-11-17 14:13
雷达观测 2026-11-18 19:43
摘要
Large language models (LLMs) provide semantic features for uplift estimation, but treatment-revealing text can induce prediction shortcuts that do not transfer to deployment. We propose \emph{LLM-based Uplift Leakage Regularization} (\textsc{LULR}) to address this risk. With the LLM encoder frozen, \textsc{LULR} penalizes dependence between the causal head's trainable shared representation and treatment while retaining separate outcome heads. The penalty uses a variational approximation to the contrastive log-ratio upper bound (CLUB) of mutual information, targeting treatment dependence rather than generic parameter shrinkage. Controlled proxy experiments provide exploratory evidence of improved ranking scores under injected treatment information relative to no regularization and the tested weight-decay setting. Structured-only benchmarks show modest, dataset-dependent effects. On an industrial pricing randomized controlled trial, mean offline is for \textsc{LULR}, compared with for the same-backbone baseline. A separate one-week online A/B test of the complete deployed pricing policy records higher sales volume and higher gross merchandise value than the control arm. These short-term policy-level gains do not isolate the regularizer's effect.
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v2 · 2026-11-17 14:13 · 最新观测版本
Regularizing Treatment Leakage for LLM-based Uplift Estimation
Large language models (LLMs) provide semantic features for uplift estimation, but treatment-revealing text can induce prediction shortcuts that do not transfer to deployment. We propose \emph{LLM-based Uplift Leakage Regularization} (\textsc{LULR}) to address this risk. With the LLM encoder frozen, \textsc{LULR} penalizes dependence between the causal head's trainable shared representation and treatment while retaining separate outcome heads. The penalty uses a variational approximation to the contrastive log-ratio upper bound (CLUB) of mutual information, targeting treatment dependence rather than generic parameter shrinkage. Controlled proxy experiments provide exploratory evidence of improved ranking scores under injected treatment information relative to no regularization and the tested weight-decay setting. Structured-only benchmarks show modest, dataset-dependent effects. On an industrial pricing randomized controlled trial, mean offline is for \textsc{LULR}, compared with for the same-backbone baseline. A separate one-week online A/B test of the complete deployed pricing policy records higher sales volume and higher gross merchandise value than the control arm. These short-term policy-level gains do not isolate the regularizer's effect.
查看此版本解读 →查看此版本原文 ↗v1 · 2026-11-16 17:42
Regularizing Treatment Leakage for LLM-based Uplift Estimation
Uplift modeling estimates the conditional average treatment effect (CATE) and supports industrial decisions such as promotional pricing. When large language models (LLMs) encode product text, treatment-revealing content can introduce shortcuts that do not transfer to deployment. We propose \emph{LLM-based Uplift Leakage Regularization} (\textsc{LULR}), which uses a CLUB-based penalty on the causal head's trainable shared representation while keeping the LLM encoder frozen. Controlled proxy experiments show a positive ranking-score interaction and higher scores than the tested weight-decay configuration; treatment-dependent evaluation limits their interpretation to exploratory model behavior. Structured-only comparisons provide reference performance, without establishing a general advantage over tree-based estimators. On a million-record industrial pricing RCT, \textsc{LULR} improves offline in seeds, although the difference is not statistically significant and estimated policy values are near-tied. A separate one-week online A/B experiment records higher sales volume and higher GMV than the control arm. These aggregate results provide short-term business evidence for the complete deployed pricing policy; they do not isolate the regularizer's contribution.
查看此版本解读 →查看此版本原文 ↗