AdaScene: Adaptive Scenario Modeling for Multi‑Scene Recommendation in Meituan DSP

1 Introduction

Meituan Delivery DSP (Demand‑Side Platform) is responsible for recommending and delivering material on external media. Rapid business expansion brings a large number of heterogeneous display scenarios (e.g., splash screen, interstitial, feed, pop‑up). This diversity leads to two major problems: (1) negative transfer caused by the long‑tail distribution of traffic across channels, and (2) extreme data sparsity for each fine‑grained scenario, which hampers model convergence.

2 Adaptive Scenario Modeling

Multi‑scene modeling follows the classic Embedding + Mixture‑of‑Experts (MoE) paradigm, where the overall loss is the weighted sum of losses from all scenarios. To address the above challenges, we propose AdaScene, which consists of two complementary modules: a knowledge‑transfer module and a scene‑aggregation module.

2.1 Adaptive Scenario Knowledge Transfer

We design an Adaptive Knowledge Transfer Network (AKTN). The network contains:

Scene Feature Adaptation : a Squeeze‑and‑Excitation‑based Scene Adaptation Layer that learns per‑scenario weights for raw features, allowing the model to focus on the most informative features for each scenario.

Scene Knowledge Transfer : a GRU‑based Scene Transfer Layer that gates the flow of information between global experts and scenario‑specific experts, thereby controlling the degree of knowledge sharing and mitigating negative transfer.

In addition, we replace the deterministic expert selection with a sparse expert network that automatically chooses K experts for each scenario via a differentiable gating mechanism. Experiments (Table 1) show consistent gains on both head and tail channels, especially for low‑traffic scenarios.

2.2 Adaptive Scenario Aggregation

Because each fine‑grained scenario is extremely sparse, we cluster scenes to form a manageable number of scene groups. We adopt a two‑stage strategy:

Stage 1 : compute pairwise similarity between scenes (using gradient‑based metrics) and find a grouping that maximizes overall similarity.

Stage 2 : fine‑tune the model on the grouped scenes with a cross‑entropy loss.

We explore three similarity‑estimation methods:

Gradient Regulation : add a regularization term based on the distance between gradients of scene‑specific losses.

Lookahead Strategy : update shared parameters with the gradient of one scene and measure the gain on other scenes to capture asymmetric influence.

Meta Weights : learn a scene‑wise correlation matrix (meta‑weights) and optimize it with a MAML‑style bi‑level objective.

Extensive offline experiments (Tables 2‑5, Figures 5‑9) demonstrate that gradient‑based aggregation consistently outperforms rule‑based grouping, and that increasing the number of scene groups improves GAUC up to a saturation point.

3 Conclusion and Outlook

We have presented AdaScene, an adaptive scenario modeling framework that alleviates negative transfer and data sparsity in large‑scale recommendation systems. Future work includes exploring finer scene partitioning (e.g., media × position × time) and developing end‑to‑end traffic‑aware aggregation methods.