BST¶

Core-mechanism reproduction. Candidate token, positional encoding, and behavior Transformer.

Paper information¶

Field	Value
Paper	arXiv 1905.06874
Affiliations	Alibaba
First public release	2019-05-15 (arXiv v1)
Authors' code	No author code linked by the paper as of 2026-08-09.
Adapter / method	`bst`
Local code	`src/auto_research/reproductions/bst/`

Original paper¶

Background and contribution¶

Use self-attention over behaviors and the candidate together, rather than fixed pooling or single-target attention.

flowchart LR
  H[Behavior and position] --> T[Transformer encoder]
  C[Candidate token] --> T
  T --> M[Sequence and candidate fusion] --> P[CTR]

Key figure from the paper¶

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

Core mechanism¶

\[ H'=\operatorname{softmax}\!\left(\frac{QK^\top}{\sqrt d}\right)V,\qquad \hat y=\sigma(\operatorname{MLP}([h_{\rm cand},\operatorname{pool}(H')])). \]

Paper results¶

The Taobao recommendation A/B reports 7.57% higher CTR than WDL.

Local reproduction¶

Compare DIN with candidate-in-encoder Transformer. Three-seed NDCG@10 was +20.29% relative. See structured metrics.

auto-research reproduce --paper bst

Reproduction limits¶

Public data replace Taobao logs. User profiles, context features, and production inference were not reproduced.