[论文] LightCRS: Lightweight Multi-Domain Conversational Recommendation Syste...

论文概要

研究领域: ML 作者: He Ma, Lan You, Qiyang Wu, Chen Liu 发布时间: 2026-10-22 arXiv: 2610.26948

中文摘要

对话式推荐系统最近采用视觉-语言模型来匹配用户对话意图与多模态商品信息。然而，将这些方法扩展到多个商品领域面临一个困境：为每个领域单独部署一个适配的VLM，成本随领域数量线性增长，且需要将每个对话路由到正确的模型；而将所有领域天真地合并到一个VLM中会因为不同类别的商品特征差异很大而导致准确率下降。我们提出LightCRS，一个轻量级的查询条件化统一多模态对话推荐框架。LightCRS使用冻结编码器离线预计算多模态商品表示。在线阶段，一个紧凑的530万参数模块通过查询条件化交叉注意力为每个对话融合这些证据，其参数在所有领域间共享，并通过候选间Transformer联合比较候选商品。为实现严格评估，我们构建了MM-R2A-25，一个包含25个领域的基准测试，将Reddit上的自然推荐求助对话与Amazon商品列表配对。LightCRS在最先进的平均HR@1、NDCG@5和MRR指标上均取得最优，HR@1比最强基线提升35.4%，同时将每次查询延迟降低了266倍。联合多领域训练本身比单独训练的单领域模型HR@1提升15.2%。

原文摘要

Conversational recommender systems have recently adopted vision-language models to match user dialogue intents with multimodal product information. However, extending these methods to many product domains faces a dilemma: deploying a separately adapted VLM per domain scales cost linearly with the number of domains and requires routing each dialogue to the right model, while naively merging all domains into a single VLM degrades accuracy because product characteristics vary widely across categories. We propose LightCRS, a lightweight, query-conditioned framework for unified multi-domain multimodal conversational recommendation. LightCRS precomputes multimodal product representations once offline with frozen encoders. Online, a compact 5.3M-parameter module fuses this evidence for each dialo...

#论文 #arXiv #ML #小凯