A machine-learning method designed to steer e-commerce search toward users’ longer-term value recorded an offline relative causal-effect estimate, or RCE, of 0.053 and reported a 0.36% relative change in GMV in an online comparison with a conventional GMV proxy. The offline number is a model-based local effect estimate from observational logs, so it should not be read as a direct causal result.
The method, called Direct Causal Effect Optimization, or DCEO, is intended to learn an item-level proxy score aligned with a user-level long-term objective. It generates that score from context-dependent weights on selected upstream predictions and adds it to the search system’s existing multi-objective fusion score.
Building a score for what happens later
The study targets a search-ranking problem: learning an item-level score aligned with a user-level long-term objective. During training, DCEO uses a critic and calibrated aggregation; online, it deploys only the actor, which produces the proxy score. That score is added to the existing fusion formula without changing its original components.
For the offline work, each sample contained all impressions for one user on a reference day, along with user-, request- and impression-level features. The experiments used 14 consecutive training days and the following day for evaluation, with cumulative GMV over four days as the default long-term objective.
The reported offline RCE is a model-based local effect estimate from a critic trained on observational logs. The online test lasted 41 days and compared a conventional GMV-proxy control with the DCEO treatment under the same integration mechanism and boost strength.
The strongest results came from combining objectives
The final configuration recorded an offline RCE of 0.053. In the paper’s comparison of training objectives, RCE was 0.022 with predictive association, 0.031 with LCE, and 0.053 after adding a conditional ranking loss. The paper reports the final figure as a 2.41-fold improvement over predictive association. No confidence interval, p-value or sample size was reported for these offline comparisons.
An ablation of the training losses pointed in the same direction. LCE alone produced an RCE of 0.031, while the conditional ranking loss alone produced 0.048. Their best combined result was 0.053 at alpha = 0.3; alpha = 0.1 performed similarly, while alpha = 1.0 produced 0.047.
The choice of upstream predictions also mattered in the reported tests. The full set of 17 predicted scores had an RCE of 0.053, compared with 0.027 for a GMV-only set. Basic, value-aware and conversion-funnel sets recorded 0.039, 0.041 and 0.040, respectively.
The proxy’s composition changed with the objective horizon. The mean weight on an impression-to-click score was 0.331 for one-day GMV and 0.404 for four-day GMV, a pattern the authors interpret as adaptation to the objective being optimized.
In the reported online comparison, DCEO showed a 0.36% relative change in GMV versus the conventional proxy. The corresponding reported changes were 0.36% in click count and 0.12% in purchase count. The analysis did not provide absolute baselines for these measures.
Where the evidence stops
The offline RCE is not a direct measurement of what happened after a live intervention. It is a model-based local effect estimate produced by a critic trained on observational logs, so it depends on which confounding factors were observed, whether the data covered the relevant cases and how accurately the critic modeled them.
The offline calibration also held impressions fixed. That makes proxy rates easier to compare, but it leaves out effects that might operate through ranking-related changes in user activity or in the number of impressions a user receives. The offline RCE therefore is not an estimate of the total effect of deploying the system.
The online figures are difficult to weigh without more reporting. The supplied analysis says the paper does not give confidence intervals, p-values, group sizes or absolute baselines, leaving the robustness of the small reported differences unclear.
The document is an arXiv preprint dated 26 August 2026. No funding statement is reported in the supplied text or metadata; the front matter lists an Alibaba affiliation.
Paper data and sources
Original title: DCEO: Direct Causal Effect Optimization for Long-Term User Value Modeling in E-commerce Search
Authors: Junzhao Zhang, Tao Zhang, Liren Yu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text