Seven researchers from Alibaba’s Taobao & Tmall Group published a paper on arXiv on August 26, 2026 describing a ranking layer that decides, per shopper and per request, how much a click prediction should count against a purchase prediction.
In a 41-day online A/B test against the conventional formula, the system lifted gross merchandise value by 0.36%, clicks by 0.36% and purchases by 0.12%.
Sub-1% sounds tiny. At marketplace volume it is not.
The problem: ranking scores items, business counts users
The paper calls it the granularity gap. A ranking model assigns a number to every candidate product inside one request. The commercial objective accumulates across a person over days — cumulative purchases, or cumulative GMV over an n-day window.
You cannot train directly on the user-level number, because every item shown to that shopper would inherit the same label despite contributing very differently.
The industry workaround is multi-objective fusion: upstream models predict click, cart-add, purchase and transaction value, then a fusion module blends them into one score. Practitioners tune those blend weights through repeated online A/B tests. The paper names two limits — a small set of globally shared parameters means almost no personalisation, and online experiments are slow and expensive.
Predicting a metric is not the same as moving it
This is the sharpest idea in the paper, and it should feel familiar to anyone who has fought about attribution.
Predicted GMV — click-through rate times conversion rate times expected order value — is a great predictor. But purchase intent leaks into every funnel stage, so a signal that correlates with high-value shoppers may simply be identifying them, not changing behaviour. The authors cite Tyler VanderWeele’s 2013 surrogate-measures work as the reference for why a good predictor need not be a good intervention target.
So they redefine the target as a relative causal effect: how much does four-day GMV improve under a given relative increase in the proxy metric?
How DCEO is built
It is an actor-critic setup, and only the actor ever ships.
- Actor: takes user and request features, outputs non-negative weights summing to one across 17 upstream predicted scores.
- Critic: estimates four-day GMV from user features plus the calibrated proxy metric.
- Causal effect loss: the actor maximises the gap between the critic’s prediction now and its prediction under a hypothetical 5% proxy increase.
- Stabiliser: a conditional normalised ranking loss with a Bradley-Terry pairwise term, combined at a coefficient of 0.3.
- Serving: the actor’s output enters the existing fusion formula as one extra logarithmic term. Nothing else is touched.
That last point is why other teams might actually copy it. Analyst Eric Seufert, summarising the paper on LinkedIn, said the approach is attractive because “it fuses the point models in an existing prediction stack rather than displacing it with an end-to-end rebuild”.
The weights are the real story
Offline, the relative causal effect hit 0.053 — a 2.41x improvement over predictive-association optimisation. Impression-to-click probability took the largest mean weight at 0.404. Click-to-purchase-above-1,000 took 0.205. Direct purchase probability after an impression got just 0.032.
And predicting GMV directly? Impression-to-GMV alone scored 0.027, the weakest of five score-set configurations tested.
Change the objective and the mix changes completely. Optimise for four-day click count and 0.999 of the weight goes to click prediction. Optimise for GMV and it spreads across transaction-value bands.
What D2C brands and retail media buyers should take from this
Your catalogue is ordered by a function whose objective was chosen inside the platform and never published. Same products, same query, different weighting depending on that choice.
Three practical moves:
- Stop optimising creative and feeds purely for click-through if the platform’s horizon is multi-day value — high-ticket conversion signals carry real weight.
- Treat onsite sponsored placement economics as dependent on an organic ranking function that is itself being retuned per shopper.
- In your own measurement, separate “predicts revenue” from “causes revenue”. They are not the same column.
The authors are unusually candid about limits: the offline estimate rests on observational logs, unobserved confounding may bias it, and the A/B test validates the end-to-end proxy, not the 0.053 number.
Source: PPC Land
