| 摘 要: 针对对话式推荐系统中忽视用户负反馈和对话历史信息的问题,创新地提出了一个对话式推荐模型。首先,该模型通过构建负反馈图捕获和编码用户拒绝的属性和项目信息,结合动态奖励函数,使系统能更精准地理解用户的实时偏好。其次,将序列模型融入智能体实现了对每步对话状态的编码,从而能基于全局对话状态做出更准确的推荐决策。为验证模型的有效性,在LastFM和LastFM*数据集上开展实验,相较于最优的基线模型,本文方法的推荐成功率分别提升了21%和13.7%,平均推荐轮次数也分别降低了1.56轮和2.35轮。实验结果表明,用户负反馈和对话历史的深度整合,为对话式推荐系统带来更高的准确性。 |
| 关键词: 推荐系统;对话式推荐;强化学习;图表示学习 |
|
中图分类号: TP315
文献标识码: A
|
| 基金项目: 国家重点研发计划项目(2022YFB3105401) |
|
| Research on Conversational Recommender Technology Based on User Feedback and Conversation History |
|
YANG Chang, YAO Yue, FANG Linfeng, ZHOU Renjie
|
(School of Computer Science, Hangzhou Dianzi University, Hangzhou 310012, China)
yangchang@hdu.edu.cn; yaoyue@hdu.edu.cn; terrywu59@gmail.com; rjzhou@hdu.edu.cn
|
| Abstract: This paper proposes an innovative conversational recommender model to address the issue of overlooking user negative feedback and dialogue history information in conversational recommender systems. Firstly, this model captures and encodes the attributes and item information that users refuse by constructing a negative feedback graph, and combines dynamic reward function to make the system understand the user' real-time preferences more accurately. Secondly, the sequence model is integrated into the intelligent agent to realize encoding of each dialogue state, so that more accurate recommendation decisions can be made based on the global dialogue state.To verify the effectiveness of the model, experiments are conducted on the LastFM and LastFM* datasets. Compared to the optimal baseline model, the recommendation success rate of the proposed method has increased by 21% and 13.7% , respectively, and the average number of recommendation rounds has decreased by 1. 56 and 2. 35 rounds, respectively. The experimental results indicate that the deep integration of user negative feedback and dialogue history brings higher accuracy to conversational recommendation systems. |
| Keywords: recommender system; conversational recommender; reinforcement learning;graph representation learning |
