DANet: Discount-Aware Conversion Prediction¶

Core mechanism evidence. Paper and local evaluations are separate.

Paper information¶

Field	Value
Paper	arXiv 2607.12578
Affiliations	Alibaba Group / Tmall
First public release	2026-07-14 (arXiv v1)
Authors' code	Author code.
Adapter / method	`danet`
Local code	`src/auto_research/reproductions/danet/`

Original paper¶

Background and contribution¶

Interest alone does not explain price sensitivity. IPN models general interest. DPN FFT-filters a 400-day discount series; DCM-User extracts preferences from similar purchases, DCM-Context gates promotion regimes, and positive purchases supervise discount regression.

flowchart LR
  I[Interest] --> P[IPN]
  D[Discount series] --> F[FFT and IFFT] --> C[User and context DCM] --> Y[Prediction and auxiliary regression]
  P --> Y

Key figure from the paper¶

Original Figure 2: DANet. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\[ Z=FFT(S^d),\quad S_l^d=IFFT(f_t(Z)),\quad S_h^d=S^d-S_l^d, \]

\[ \mathcal L=\mathcal L_{CE}+\alpha\mathbf 1[y=1]\operatorname{MSE}(y_d,\hat y_d). \]

Paper results¶

Tmall online pCVR +3.63%, GMV +2.23%; discount-item PVR 7.56→11.14 (+47.32%).

Local reproduction¶

Compare IPN with TFTM/DCM using periodic popularity/rating as a discount proxy: Hit@10 +0.00%, NDCG@10 −1.46%, fresh Hit@10 +50.00%. Metrics.

auto-research reproduce --paper danet --seed 42

Reproduction limits¶

No real prices, 400-day Tmall discounts, or promotion features. NumPy implementation is distinct from authors' TensorFlow; ranking is not CVR.