RedMall：面向大规模电商的生成式推荐
原标题：RedMall: Generative Recommendation for Large-Scale E-commerce
AI 摘要
该论文提出 RedMall，一种面向大规模电商的生成式推荐系统。它通过残差量化构建分层语义 ID，引入轻量 SID 编码器建模层间依赖，并结合稀疏 MoE、双门控注意力与多目标预测优化会话级推荐；同时用曝光感知损失和 GRPO 加 LambdaRank 辅助损失对齐用户偏好与列表排序。在四个公开数据集上，RedMall 的 Recall 和 NDCG 指标均优于基线；在线 A/B 测试中，广告场景的广告主价值提升 2.24%、CTR 提升 0.59%，推荐场景的 GMV 提升 1.08%、日活跃买家提升 0.95%。
正文节选
RedMall: Generative Recommendation for Large-Scale E-commerce Abstract Generative recommendation has shown considerable potential in industrial recommendation systems. However, the performance of existing generative retrieval methods is limited by several challenges. 1) Semantic IDs (SIDs) are treated as one-dimensional sequences, overlooking the hierarchical dependencies among different layers. 2) Next-token prediction (NTP) used in existing methods is unable to explicitly optimize session-wise