DCEO: Direct Causal Effect Optimization for Long-Term User Value Modeling in E-commerce Search

Junzhao Zhang; Tao Zhang; Liren Yu; Feiyi Dong; Zhixuan Zhang; Dan Ou; Haihong Tang

cs.LGAug 26, 2026

DCEO: Direct Causal Effect Optimization for Long-Term User Value Modeling in E-commerce Search

Authors: Junzhao Zhang, Tao Zhang, Liren Yu, Feiyi Dong, Zhixuan Zhang, Dan Ou, Haihong Tang

Organizations: Taobao & Tmall Group of Alibaba Hangzhou, Zhejiang, China · Taobao & Tmall Group of Alibaba Beijing, China

Abstract

Industrial e-commerce search systems ultimately aim to optimize the user-level long-term objective, such as n-day cumulative purchases or gross merchandise value (GMV) per user. However, such objectives are defined at the user level, whereas search ranking is based on item-level scores within each request. Existing methods typically bridge this granularity gap through manually designed multi-objective fusion, where predictions of multiple item-level objectives, such as clicks, carts, purchases, and transaction value, are combined into a ranking score that serves as a proxy for the ultimate objective. Such hand-crafted fusion schemes rely on a small set of manually tuned weights, limiting fine-grained personalization and leading to suboptimal alignment with the ultimate objective. In this paper, we propose DCEO (Direct Causal Effect Optimization), a data-driven framework for learning item-level proxy scores that are better aligned with the ultimate objective. We first aggregate the item-level proxy scores into a user-level proxy metric and quantify its alignment with the ultimate objective using a relative causal effect. We then develop an actor-critic framework, where the critic estimates the ultimate objective for a given user-level proxy metric, and the actor dynamically generates context-dependent fusion weights over multiple objectives to construct the item-level proxy scores and is trained to directly optimize the relative causal effect. Extensive offline experiments and analyses demonstrate the effectiveness and interpretability of DCEO. In addition, DCEO has been deployed in a large-scale industrial e-commerce search system, outperforming the conventional GMV proxy by 0.36% in GMV in a 41-day online A/B test.

Figures & tables

Panel of Figure 1 — Figure 1. Overview of DCEO. During offline training, the actor $f_{\theta}$ generates request-specific weights $\mathbf{w}_{ur}$ from the user and request features $\mathbf{z}_{u}$ and $\mathbf{z}_{ur}$ . The weights combine the selected upstream predicted scores $\mathbf{v}_{uri}$ into the item-level proxy score $p_{uri}$ , which is aggregated into the user-level proxy metric $P_{u}$ . The critic $h_{\phi}$ models the ultimate objective $Y_{u}$ from $\mathbf{z}_{u}$ and $P_{u}$ . The actor and critic are optimized using $\mathcal{L}_{\mathrm{actor}}$ and $\mathcal{L}_{\mathrm{critic}}$ , respectively. During online serving, only $f_{\theta}$ is deployed, and $p_{uri}$ is added to the multi-objective fusion score $s_{uri}$ to improve the item ranking. Overview of DCEO. During offline training, the actor $f_\theta$ generates request-specific weights $\mathbfw_{ur}$ from the user and request features $\mathbfz_u$ and $\mathbfz_{ur}$ . The weights combine the selected upstream predicted scores $\mathbfv_{uri}$ into the item-level proxy score $p_{uri}$ , which is aggregated into the user-level proxy metric $P_u$ . The critic $h_\phi$ models the ultimate objective $Y_u$ from $\mathbfz_u$ and $P_u$ . The actor and critic are optimized using $\mathcalL_{\mathrm{actor}}$ and $\mathcalL_{\mathrm{critic}}$ , respectively. During online serving, only $f_\theta$ is deployed, and $p_{uri}$ is added to the multi-objective fusion score $s_{uri}$ to improve the item ranking.

Score name	Meaning
impr2click	Probability of a click after an impression.
impr2cart	Probability of a cart addition after an impression.
impr2pay	Probability of a purchase after an impression.
impr2gmv	GMV generated after an impression.
impr2pay10	Probability of a purchase with transaction value above 10 after an impression.
impr2pay30	Probability of a purchase with transaction value above 30 after an impression.

Table 1. Selected upstream predicted scores used by DCEO. These predicted scores are estimated by upstream models.

Statistic	impr2click	impr2pay	click2pay10	click2pay30	click2pay100	click2pay1000
Mean	0.404	0.032	0.154	0.079	0.105	0.205
Standard deviation	0.115	0.096	0.057	0.046	0.049	0.067

Table 2. Mean and standard deviation of representative actor weights in the final 4-day GMV model. The standard deviation measures personalization strength and is computed over impressions.

Actor loss	RCE
Predictive association optimization: $-h_{\operatorname{sg}(\phi)}(\mathbf{z}_{u},P_{u}(\theta))$	0.022
Causal effect optimization: $\mathcal{L}_{\mathrm{CE}}$	0.031
Causal effect optimization: $\mathcal{L}_{\mathrm{actor}}$	0.053

Table 3. Comparison of predictive association and causal effect optimization. Higher RCE is better.

Actor loss	RCE
$\mathcal{L}_{\mathrm{CE}}$	0.031
$\mathcal{L}_{\mathrm{CNR}}$	0.048
$\mathcal{L}_{\mathrm{CE}}+0.1\mathcal{L}_{\mathrm{CNR}}$	0.052
$\mathcal{L}_{\mathrm{CE}}+0.3\mathcal{L}_{\mathrm{CNR}}$	0.053
$\mathcal{L}_{\mathrm{CE}}+1.0\mathcal{L}_{\mathrm{CNR}}$	0.047

Table 4. Ablation of the actor losses. Higher RCE is better.

Predicted-score set	RCE
impr2gmv only	0.027
Basic impression-level set	0.039
Value-aware set	0.041
Conversion-funnel set	0.040
Full 17-score set	0.053

Table 5. Effect of the upstream predicted-score set. Higher RCE is better.

Ultimate objective	impr2click	click2pay	click2pay10	click2pay30	click2pay100	click2pay1000
4-day click count	0.999	0.000	0.000	0.000	0.000	0.000
4-day purchase count	0.531	0.101	0.256	0.000	0.000	0.000
4-day GMV	0.404	0.032	0.154	0.079	0.105	0.205
4-day purchase count $+\,0.1\times$ GMV	0.407	0.000	0.056	0.098	0.103	0.187

Table 6. Mean actor weights under different ultimate objectives. The table reports representative predicted scores defined in Table 1 ; the remaining predicted scores account for the unreported weight mass.

Ultimate objective	impr2click	click2pay	click2pay10	click2pay30	click2pay100	click2pay1000
1-day GMV	0.331	0.034	0.108	0.144	0.127	0.158
2-day GMV	0.379	0.000	0.000	0.140	0.129	0.189
3-day GMV	0.417	0.000	0.155	0.081	0.122	0.202
4-day GMV	0.404	0.032	0.154	0.079	0.105	0.205

Table 7. Mean actor weights for GMV objectives with different accumulation windows. The table reports representative predicted scores defined in Table 1 ; the remaining predicted scores account for the unreported weight mass.

Metric	Relative change
Click count	$+0.36\%$
Purchase count	$+0.12\%$
GMV	$\mathbf{+0.36\%}$

Table 8. Online relative changes of DCEO over the conventional GMV proxy in the 41-day A/B test.

E-Commerce Search

E-commerce search studies how to match shoppers’ queries, preferences, and product signals with useful items in online marketplaces.

See all 42 papers

Towards Sustainable Growth: A Multi-Value-Aware Retrieval Framework for E-Commerce Search
May 18, 2026Yifan Wang, Yixuan Wang, YiDan Liang +2Cold-Start Recommendation E-Commerce Search
Taobao & Tmall Group of Alibaba · HangZhou, China
From Prediction to Incrementality: Causal Optimization for Large-Scale Targeting and Recommendation
Aug 10, 2026Changshuai Wei, John Bencina, Phuc Nguyen +2Causal Effect Estimation Constrained Optimization
LinkedIn · Seattle, USA · Sunnyvale, USA
AIGP: An LLM-Based Framework for Long-Term Value Alignment in E-Commerce Pricing
Jun 25, 2026Chennan Ma, Yanning Zhang, Siqi Hong +3LLM Alignment Dynamic Pricing
Taobao & Tmall Group of Alibaba Hangzhou, Zhejiang, China

May 18, 2026J/K move · Enter open · S save

Towards Sustainable Growth: A Multi-Value-Aware Retrieval Framework for E-Commerce Search

Yifan Wang, Yixuan Wang, YiDan Liang, Qiang Liu+1

Taobao & Tmall Group of Alibaba · HangZhou, China

Cold-Start Recommendation E-Commerce Search Generative Retrieval Information Retrieval E-Commerce Multi-Objective Optimization

Open paper AskPDF arXiv

New item growth is critical for maintaining a healthy ecosystem in large-scale e-commerce platforms. However, existing systems tend to prioritize presenting users with already popular items, a phenomenon often referred to as the "Matthew effect". In the context of search retrieval, current cold-start models suffer from the misalignment between training objectives and online business metrics, and they lack effective mechanisms to measure an item's growth potential. In this paper, we propose a Multi-Value-Aware retrieval framework tailored for e-commerce search, designed to better align with the cascaded online values across different stages of the search system while balancing immediate conversion and long-term item growth. Our framework GrowthGR consists of two key components: an Item Long-term Transaction Value Prediction (ItemLTV) module and a Multi-Value-Aware Generative Retrieval (MultiGR) module. First, in the ItemLTV module, we employ counterfactual inference to quantify the long-term value increment attributable to a single user interaction. Second, in the MultiGR module, building upon a semantic-ID-based generative retrieval architecture, we leverage structured samples with the search cascade signals and adopt a Multi-Value-Aware Policy Optimization (MoPO) training paradigm to align with multi-stage online values, while explicitly balancing short-term transactional value and long-term growth potential estimated by ItemLTV. We successfully deployed GrowthGR on Taobao's production platform, achieving a substantial 5.3% lift in new item GMV while delivering a non-trivial 0.3% gain in overall search GMV. Extensive online analysis and A/B testing demonstrate its positive impact on the overall ecosystem value.

Abstract

E-Commerce Search

Explore similar work

Towards Sustainable Growth: A Multi-Value-Aware Retrieval Framework for E-Commerce Search

From Prediction to Incrementality: Causal Optimization for Large-Scale Targeting and Recommendation

AIGP: An LLM-Based Framework for Long-Term Value Alignment in E-Commerce Pricing