简体中文 · English

DCEO: Direct causal-effect optimization for long-term user value¶

Core mechanism reproduction. Context-conditioned objective weights, user-level proxy aggregation and relative causal interventions run. No production critic or online actor serving.

Paper information¶

Field	Details
Paper	arXiv 2608.25635
Institution	Taobao & Tmall Group, Alibaba
First public release	2026-08-26 (arXiv v1)
Author code	No public author code found on 2026-08-28.
Local adapter or method key	`dceo`
Local reproduction code	`src/auto_research/reproductions/dceo/`

Original paper summary¶

Background and main changes¶

Fixed weights for click, conversion and GMV cannot adapt to user state or directly represent an action's long-term causal value. The actor outputs context-dependent simplex weights and a user-level critic estimates relative intervention effects. Only the lightweight actor runs online.

flowchart LR
  X[User and query context] --> A[Objective-weight actor]
  S[Click, conversion and GMV scores] --> F[Conditional fusion]
  A --> F
  F --> R[Ranking action]
  R --> C[User-level long-term critic]
  C --> A

Key figure from the paper¶

Figure 1: Training workflow and key optimizations. From the paper; copyright belongs to the authors. Click for the source.

Core equations¶

\[ w(x)=\operatorname{softmax}(g_\theta(x)),\quad s(x,i)=\sum_k w_k(x)s_k(x,i),\quad \Delta V=V(u\mid do(a))-V(u\mid do(a_0)). \]

Offline and online results¶

A 41-day Taobao Search production A/B reports GMV +0.36%, the quantitative production evidence for inclusion in the industrial implementation queue.

Local reproduction¶

The local Evolve search includes causal_gain × temperature. Rank 20 candidates on validation by NDCG@10 + 0.15 × Hit@10; evaluate only the winner once on held-out test. This avoids calling a fixed 0.75 gain an overall improvement when Hits rise and NDCG falls.

Comparison: Fixed log fusion of the same proxy scores versus the validation-selected context actor (causal_gain=0.75, temperature=0.75). MovieLens-1M Hit@10 0.1469 → 0.1500 (+2.13%), NDCG@10 0.07127 → 0.07095 (-0.45%), head share 0.2038 → 0.1988 (-2.45%). Search mitigates the fixed configuration's NDCG loss but does not beat the baseline on all metrics.

See metrics.

auto-research reproduce --paper dceo --dataset-dir data --seed 42

Scope¶

Proxy outcomes are not actual long-term GMV. The local negative-correlation diagnostic explicitly prevents treating a Hit gain as long-term value improvement.