基于艾宾浩斯遗忘曲线和注意力机制的推荐算法  

Ebbinghaus forgetting curve and attention mechanism based recommendation algorithm

作　　者：金楠 王瑞琴[1,2] 陆悦聪 JIN Nan;WANG Ruiqin;LU Yuecong(School of Information Engineering,Huzhou University,Huzhou 313000,China;Zhejiang Province Key Laboratory of Smart Management&Application of Modern Agricultural Resources,Huzhou 313000,China)

机构地区：[1]湖州师范学院信息工程学院,浙江湖州313000 [2]浙江省现代农业资源智慧管理与应用研究重点实验室,浙江湖州313000

出　　处：《电信科学》2022年第10期89-97,共9页Telecommunications Science

基　　金：浙江省自然科学基金资助项目(No.LY20F020006)。

摘　　要：传统基于注意力机制的推荐算法只利用位置嵌入对用户行为序列进行建模,忽略了具体的时间戳信息,导致推荐性能不佳和模型训练过拟合等问题。提出基于时间注意力的多任务矩阵分解推荐模型,利用注意力机制提取邻域信息对用户和物品进行嵌入编码,借助艾宾浩斯遗忘曲线描述用户兴趣随时间的变化特性,在模型训练过程中引入经验回放的强化学习策略模拟人类的记忆复习过程。真实数据集上的实验结果表明,该模型比现有推荐模型具有更好的推荐性能。Traditional attention-based recommendation algorithms only use position embeddings to model user behavior sequences,however,ignore specific timestamp information,resulting in poor recommendation performance and overfitting of model training.The multi-task matrix factorization recommendation model based on time attention was proposed,which used the attention mechanism to extract the neighborhood information for the user and item embedding,and used the Ebbinghaus forgetting curve to describe the changing characteristics of user interests over time.The model training process introduced a reinforcement learning strategy of experience replay to simulate the human memory review process.Experimental results on real datasets show that the proposed model has better recommendation performance than existing recommendation models.

关 键 词：艾宾浩斯遗忘曲线 注意力机制 强化学习 经验回放 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]