融合自注意力与渐进抽取机制的电商多场景多任务预测方法研究

Research on a Multi-Scenario, Multi-Task Prediction Method for E-Commerce Based on Self-Attention and Progressive Extraction Mechanisms

摘要: 随着电商平台由单场景、单目标逐步迈向多场景、多任务的统一建模范式，如何在充分共享跨场景共性知识的同时有效规避任务间冲突与负迁移，已成为制约模型可用性与训练稳定性的核心挑战。针对该问题，提出一种面向电商多场景多任务预测的分层专家增强框架：以Star场景融合底座为基础，在嵌入层之后引入自注意力模块，以显式刻画高阶特征依赖；同时采用渐进分层抽取替代传统单层MoE的共享机制，使共享–特化在不同语义层级实现自适应迁移与有效解耦，从结构上缓解跷跷板现象并抑制负迁移。此外，模型保留场景/任务专家，并通过可学习的平衡混合机制增强对场景不均衡与分布漂移的鲁棒性。在两个公开数据集上开展了广泛实验，结果表明该方法在整体性能与稳定性方面均显著优于对比模型。

Abstract: As e-commerce platforms gradually evolve from single-scenario and single-objective frameworks toward a unified multi-scenario and multi-task modeling paradigm, how to fully share cross-scenario common knowledge while effectively avoiding inter-task conflicts and negative transfer has emerged as a core challenge restricting model usability and training stability. To address this issue, we propose a hierarchical expert-enhanced framework for multi-scenario multi-task prediction in e-commerce. Building upon a Star-topology scenario fusion foundation, a self-attention module is introduced after the embedding layer to explicitly capture high-order feature dependencies. Simultaneously, a progressive hierarchical extraction mechanism replaces the traditional single-layer MoE sharing approach. This enables the adaptive transfer and effective decoupling of shared and task-specific representations across different semantic levels, fundamentally mitigating the seesaw phenomenon and suppressing negative transfer from a structural perspective. Furthermore, the model retains scenario and task-specific experts, utilizing a learnable balanced mixing mechanism to enhance robustness against scenario imbalance and distribution shift. Extensive experiments conducted on two public datasets demonstrate that the proposed method significantly outperforms baseline models in terms of overall performance and stability.