AdaF²M²: multi-forward feature learning and state adaptation¶

简体中文

Fidelity: core mechanism. Runs three feature-mask forwards, shared aggregation, and an adaptive user/item-state adapter.

Paper information¶

Field	Details
Paper	arXiv 2501.15816
Organization	ByteDance / Douyin
First public release	2025-01-27 (arXiv v1)
Authors' code	The paper did not supply official or author code when checked on 2026-08-09.
Adapter / method	`adaf2m2`
Local reproduction code	`src/auto_research/reproductions/adaf2m2/`

Original paper¶

Motivation and method¶

A single all-feature forward can over-rely on strong features, and different user or item states need different features. AdaF²M² uses repeated forwards under distinct masks to learn complementary representations and fuses them dynamically with a state-aware adapter.

flowchart LR
  X["Full features"] --> M1["Full mask"]
  X --> M2["Sparse mask A"]
  X --> M3["Sparse mask B"]
  M1 --> A["State-aware adapter"]
  M2 --> A
  M3 --> A
  S["User activity and item popularity"] --> A
  A --> Y["Ranking"]

Key figure from the paper¶

Figure 1: Core architecture, modules, and their connections. From the original paper; copyright remains with the authors.

Core equations¶

\[ h_k=f(x\odot m_k),\qquad h=\sum_k \alpha_k(s)h_k,\quad \alpha(s)=\operatorname{softmax}(W_s s). \]

Paper results¶

Private offline scenarios show gains from masked multi-forward and the adapter. Online Douyin cumulative active days rose +1.37% and app time +1.89%; the method was deployed for retrieval, ranking, and cold start.

Local reproduction¶

Local comparison: NDCG@10 is +2.42% relative to a single-forward full-feature baseline.

Compared with that baseline, NDCG@10 moves 0.07514→0.07696 (+2.42%), and head share moves 0.21762→0.15976. See the metrics.

auto-research reproduce --paper adaf2m2 --dataset-dir data --seed 42

Reproduction boundaries¶

States use only history length and item popularity. There are no private Douyin features, full-pipeline joint training, or online resource-cost measurements.