简体中文 · English
TMallGS: Unified feature and sequence modeling for generative e-commerce search¶
Core mechanism reproduction. Trains field-adaptive interactions and progressive loss. Private Tmall fields and production scale are not reproduced.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv 2607.13398 |
| Institution | Taobao & Tmall Group of Alibaba |
| First public release | 2026-07-15 (arXiv v1) |
| Author code | No official author code in the paper when checked on 2026-08-09. |
| Local adapter or method key | tmallgs |
| Local reproduction code | src/auto_research/reproductions/tmallgs/ |
Original paper summary¶
Background and main changes¶
Search mixes heterogeneous query, short- and long-term behavior, user and item fields. Direct concatenation can amplify field noise. TMallGS learns field-specific Q/K/V and controls interactions with distribution-calibrated saliency and noise-adaptive gates. It injects content through FiLM, corrects final scores with context bias, and assigns progressive supervision by intermediate-layer errors.
flowchart LR
A["Query / UIH / UQH / item fields"] --> B["Field-wise QKV"]
B --> C["Saliency calibration"]
C --> D["Noise-adaptive gate"]
E["Content features"] --> F["FiLM"]
D --> F
F --> G["Context bias and ranking head"]
G -. "Layerwise error supervision" .-> BKey figure from the paper¶
Figure 1: Core architecture, modules and connections. From the paper; copyright belongs to the authors. Click for the source.
Core equations¶
Offline and online results¶
Offline PV-AUC +0.79%, IMP-GAUC +0.34%. A 30-day Tmall Search A/B (\(p<0.05\)) reports UCTCVR +1.38%, GMV +1.52% and roughly 6 ms additional latency.
Local reproduction¶
MovieLens genres and sequences become heterogeneous fields. Field-wise QKV, noise gates, explicit content FiLM, context bias and error-aware progressive loss all train.
Comparison: DIN/RankMixer-style late-fusion Transformer versus the TMallGS field-adaptive gated Transformer. Seed 42 NDCG@10 rises from 0.00068 to 0.00278, +310.42% relative.
See metrics/movielens-100k-seed42.json. Absolute hits are sparse. The result shows a runnable combination of training components, not statistical stability of Tmall's 0.05B production model.