摘要
基于自注意力机制的序列推荐算法在捕获用户交互序列的全局特征方面表现出了强大的能力,得到了广泛应用.然而交互序列当中只有一部分关键行为会对用户未来行为的演化起到决定性作用,其余冗余的噪声行为会干扰推荐结果的准确性.同时,单一尺度的自注意力机制难以从不同粒度上捕获用户行为.该文提出基于行为路径的多尺度自注意力机制序列推荐算法,在不同粒度上动态地捕获对最终推荐起到决定性作用的行为演化模式,屏蔽冗余的非关键行为,提高了推荐系统的用户体验.该模型在三个公开数据集上与同类型方法进行比较,实验结果显示,该文所提出的算法在不同的评估指标上较基线方法均有一定的提升,验证了模型的有效性.
Abstract
The sequence recommendation algorithm based on self-attention mechanism shows strong ability in captu-ring the global features of user interaction sequence.However,not all the behaviors in the interaction sequence will play a decisive role in the evolution of the user's future behavior,and the single-scale self-attention mechanism is dif-ficult to capture user behavior from different granularity.This paper proposes a multi-scale self-attention mechanism based on behavior path for sequence recommendation.It dynamically captures the behavior evolution mode that plays a decisive role in the final recommendation at different granularity,and removes redundant non-critical behaviors.The experimental results on three public datasets show that the proposed algorithm has a certain improvement over the baseline method in different evaluation metrics.
维普
万方数据