AMS: Adaptive Modality Scheduling for Industrial Multi-Scenario CTR Prediction

作者：Kailiang Hao, Chaohua Yang, Wei Zeng, Jianan Su, Kaixin Shen, Jingtong Wu, Dugang Liu, Xing Tang, Jingyang Bin, Xiuqiang He, Zhong Ming · 年份：2026 · DOI：10.1145/3770855.3818319 · 研究领域：Computer science、Artificial intelligence、Machine learning、Data mining、Real-time computing

With the increasing diversity of recommendation scenarios and the widespread adoption of multi-modal information, jointly modeling multi-scenario and multi-modal data has become a critical trend in industrial click-through rate (CTR) prediction. However, existing methods primarily focus on personalizing user-level modality preferences, overlooking the varying importance of modalities across different scenarios. Moreover, directly extending these methods to multi-scenario settings often leads to suboptimal performance and parameter inflation. To address these challenges, we propose a lightweight Adaptive Modality Scheduling (AMS) framework for industrial multi-scenario CTR prediction, which introduces an adaptive modality mask generation network that performs sample-level on-demand modality scheduling to explicitly model scenario-specific users' modality preferences. Specifically, AMS first performs projection-based alignment of multi-modal features, and then uses a hypernetwork-based mask-generation module, conditioned on scenario and modality information, to model sample-level preferences and hierarchically generate discretized modality activation masks. These masks schedule the modality combinations that best match each sample's preferences, and the activated modalities are further integrated by a lightweight attention-based modality fusion module before being fed into a downstream multi-scenario recommendation (MSR) backbone network for CTR prediction. In addition, the mod...