摘要
传统的矩阵分解算法为用户和项目分别独立设置了偏置项,而没有深入挖掘特定用户对于特定项目的隐性偏好;同时,传统的排序预测推荐算法将用户所有打分过的项目都统一地设置为该用户的正例项目(无论用户给出了好评还是差评),这导致训练完成的系统在实际应用中很可能会为用户继续推荐其厌恶的项目。因此提出了一种基于深度网络的推荐系统偏置项改良方案,该改良方案考虑了用户为特定项目所作的评分背后所蕴含的好恶态度,并学习出一个用户-项目联合偏置项加入到推荐过程中以提升推荐性能。在三个公开数据集上进行的对比实验结果表明,该改良方案可以有效地提升推荐的性能表现。
Traditional matrix factorization algorithm sets bias for users and items independently,without digging into the latent preferences of specific users for specific items.As in traditional ranking prediction recommendation algorithm,all the rated items of a user are uniformly set as the user′s positive items(regardless of whether the user gives a good or a bad review).As a result,the trained system is likely to continue to recommend projects that users hate in practical applications.Therefore,an improved bias improvement scheme of recommendation system based on deep neural network is proposed,which takes into account the liking and disliking behind the ratings made by users for specific items,and a joint bias is learned for the recommendation process.The results of comparative experiments on three open datasets show that the improved scheme can effectively improve the recommended performance.
作者
张天蔚
Zhang Tianwei(Shandong Computer Science Center(National Super Computer Center in Jinan),Jinan 250014,China)
出处
《信息技术与网络安全》
2021年第8期42-46,共5页
Information Technology and Network Security
关键词
推荐算法
协同过滤
深度神经网络
隐式反馈
recommendation algorithm
collaborative filtering
deep neural network
implicit feedback
