DIEN¶

Core-mechanism reproduction. Interest GRU, next-behavior supervision, and target-dependent evolution.

Paper information¶

Field	Value
Paper	arXiv 1809.03672
Affiliations	Alibaba
First public release	2018-09-11 (arXiv v1)
Authors' code	Author code.
Adapter / method	`dien`
Local code	`src/auto_research/reproductions/dien/`

Original paper¶

Background and contribution¶

Unlike DIN's direct interest activation, DIEN extracts stepwise interest with auxiliary supervision, then lets the target item control state evolution.

flowchart LR
  H[Behavior sequence] --> G[Interest extractor GRU]
  G --> A[Auxiliary next-item loss]
  G --> R[Target-aware AUGRU]
  T[Candidate item] --> R --> P[CTR]

Key figure from the paper¶

Original Figure 1: DIEN architecture. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ L=L_{\rm ctr}+\alpha L_{\rm aux},\qquad \tilde u_t=a_tu_t,\quad h_t=(1-\tilde u_t)h_{t-1}+\tilde u_t\tilde h_t. \]

Paper results¶

The Taobao display-ad experiment reports CTR +20.7%, eCPM +17.1%, and PPC −3.0%.

Local reproduction¶

Compare DIN with GRU, auxiliary loss, and target-aware evolution under the same training protocol. Three-seed NDCG@10 −1.98% relative; no paper gain. See metrics.

auto-research reproduce --paper dien

Reproduction limits¶

Short sparse MovieLens histories do not replace Taobao ad features. No production fused AUGRU.