基于用户长短期兴趣的自注意力序列推荐

SELF-ATTENTION SEQUENTIAL RECOMMENDATION BASED ON USERS LONG AND SHORT TERM INTERESTS

用户的行为序列既包含了用户的短期兴趣,也包含了用户的长期偏好。针对此类问题,提出一个基于用户长短期兴趣的自注意力模型。使用循环神经网络来处理变长的用户序列,得到用户的长期兴趣表示;用自注意力网络计算序列中的每个项目对预测用户下一次交互项目的重要性程度,得到用户的短期兴趣表示;将循环神经网络的输出作为查询输入到自注意力网络中,得到综合长短期兴趣的用户表示,并用这个表示来预测用户的下一次交互。该算法在三个真实世界的数据集上评估了提出的模型,其中命中率指标提高7%~30%。

The user s interaction sequence contains both the user s short-term interests and the long-term preferences.In view of this,this paper proposes a self-attention model based on users long and short term interests.The recurrent neural network was used to process variable-length user sequences,which obtained the users long-term interest representation.The self-attention network was used to calculate the importance of each item in the sequence to predict the user s next interaction item,which obtained the user s short-term interest expressed.The output of the recurrent neural network was taken as query input to the self-attention network to obtain a user s representation of comprehensive long and short term interests,and the representation was used to predict user s next interaction.The proposed model was evaluated on three real-world data sets,in which the hit rate index increased by 7%-30%.