面向会话推荐的注意力图神经网络

Attention Mechanism Enhanced Graph Neural Network for Session-based Recommendation

面向会话的推荐方式起源于无法获得用户历史数据的应用场景,它是通过匿名会话来预测用户的行为.现有面向会话的推荐方法,虽然可以准确获得项目嵌入和考虑项目的复杂转换,但不能从多维度提取会话序列中隐藏的用户的长期兴趣和短期偏好,造成推荐性能低.该文引入注意力机制,提出一种多头注意力机制和软注意力机制有机结合的新机制,并据此提出面向会话推荐的注意力图神经网络.该注意力机制通过给不同的输入数据赋予不同权重,实现对当前推荐任务更为关键的信息的聚焦,以此从不同角度提取用户的兴趣和偏好.该模型在电商数据集上进行实验,与已有的基准模型相比,该文所提模型在各项评论指标上均有显著提升.在Dgeca数据集上,P@20可达61.77%,充分表明了所提方法的有效性.

The Session-based recommendation method originates from the application scenarios where historical data of users cannot be obtained.User behavior can be predicted through anonymous sessions.Although the existing session-based recommendation methods have accurately obtained item embedding and considered the complex transformation of items,the users′long-term interests and current preferences hidden in the session sequence cannot be extracted from multiple dimensions,resulting in low recommendation performance.The attention mechanism is introduced,a new mechanism which organically combines the multi-head attention mechanism and soft attention mechanism is proposed,and an attention mechanism enhanced graph neural network for session-based recommendation is put forward in this paper.By giving different weights to different input data,the attention mechanism can focus on more critical information for the current recommendation task,in order to extract users′interests and preferences from multiple dimensions.The model is tested on e-commerce datasets,compared with the existing benchmark model,the proposed model has significantly improved on various evaluation metrics.The evaluation metrics P@20 of Dgeca can reach 61.77%,which fully demonstrates the effectiveness of the proposed method.