FuXi-α: Feature-Interaction Enhanced Transformer¶

Local core-mechanism evidence. Paper outcomes and local results are separate.

Paper information¶

Field	Value
Paper	arXiv 2502.03036
Affiliations	Huawei / USTC
First public release	2025-02-05 (arXiv v1)
Authors' code	Authors' code.
Adapter / method	`fuxi-alpha`
Local code	`src/auto_research/reproductions/fuxi_alpha/`

Original paper¶

Background and contribution¶

Scaling a recommendation Transformer can bottleneck on shallow feature interactions. Use adaptive temporal, semantic, and popularity channels with staged FFN fusion.

flowchart LR
  I[Sequence and dense features] --> T[Temporal channel] --> F[Multi-stage FFN]
  I --> S[Semantic channel] --> F
  I --> P[Feature interactions] --> F --> Y[Prediction]

Key figure from the paper¶

Original Figure 3: FuXi-α. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ H_c=\operatorname{softmax}(Q_cK_c^\top/\sqrt d)V_c,\qquad H^{l+1}=H^l+\operatorname{FFN}_l([H^l,H_1,H_2,H_3]). \]

Paper results¶

The paper reports scaling gains on MovieLens-1M, KuaiRand, and industrial data. A seven-day 30%-traffic trial reports +4.67% songs played and +5.10% listening time.

Local reproduction¶

Compared with one-channel attention, NDCG@10 0.07514→0.07194 (−4.25%), Hit@10 0.14048→0.14286. Keep the ranking regression. Metrics.

auto-research reproduce --paper fuxi-alpha --dataset-dir data --seed 42

Reproduction limits¶

Small numeric implementation, not billion-scale training, private Huawei Music features, or distributed execution.