文章摘要
吴哲夫,程界斌,方路平.基于注意力机制的对抗性协同过滤推荐算法[J].高技术通讯(中文),2023,33(10):1020~1028
基于注意力机制的对抗性协同过滤推荐算法
Adversarial collaborative filtering recommendation algorithm based on attention mechanism
  
DOI:10. 3772/ j. issn. 1002-0470. 2023. 10. 002
中文关键词: 协同过滤; 注意力机制; 对抗性学习; 鲁棒性
英文关键词: collaborative filtering, attention mechanism, adversarial learning, robustness
基金项目:
作者单位
吴哲夫 (浙江工业大学信息工程学院杭州 310023) 
程界斌  
方路平  
摘要点击次数: 512
全文下载次数: 394
中文摘要:
      针对协同过滤推荐算法中用户所交互的物品对其决策的不同贡献度问题,提出了一种基于相关注意力的协同过滤推荐算法。该算法结合深度学习中的注意力机制为不同物品分配不同的权值来捕获与目标物品最相关的物品,探索不同物品的权重对模型预测的影响并以此提升推荐的准确度;在此基础上,为了解决推荐算法鲁棒性低的问题,进一步提出了注意力协同对抗性训练的推荐算法,通过对抗性学习的方法并使用快速梯度符号算法(FGSM)构建对抗样本输入模型进行对抗训练,缓解模型受扰动的影响从而提升算法鲁棒性。在Pinterest和MovieLens-1M这2个数据集上的实验结果表明,所提算法不仅有效提升了推荐算法的准确度,同时也增强了推荐系统的鲁棒性。
英文摘要:
      Aiming at the different contribution of items interacted by users to their decision-making in collaborative filtering recommendation algorithm, a collaborative filtering recommendation algorithm based on relevant attention is proposed. The algorithm combines the attention mechanism in deep learning to assign different weights to different items to capture the items most relevant to the target item, explore the influence of the weight of different items on the model prediction, and thereby improve the accuracy of the recommendation; further, in order to solve the problem of low robustness of the proposed model, a recommendation algorithm for attention collaborative adversarial training is proposed. Through the adversarial learning method and using the fast gradient sign method (FGSM) algorithm to build an adversarial sample input model for adversarial training, the model is alleviated from the impact of disturbance and the robustness of the model is improved. Experimental results on the two datasets of Pinterest and MovieLens-1M show that the proposed algorithm not only effectively improves the accuracy of the recommendation algorithm, but also enhances the robustness of the recommendation system.
查看全文   查看/发表评论  下载PDF阅读器
关闭

分享按钮