摘要
传统的推荐系统中,用户的兴趣被认为是稳定不变的,而事实上,用户的兴趣会因为各种因素产生变化。为了更加有利地跟踪用户兴趣偏好变化进行内容推荐,提出了一种基于生成对抗网络的推荐算法L-GAN(long short-term memory via generative adversarial networks),利用长期和短期的兴趣偏好,通过生成对抗的训练策略来训练推荐模型,使推荐模型产生的推荐列表更加准确。在对抗训练过程中,将数据分为多个行为周期,按照时间顺序依次输入每个行为周期内的用户-项目评价矩阵,生成器模型产生推荐列表,而判断器模型则区分输入的推荐列表是否与真实历史记录的特征相似。最终,通过在两个公开的数据集上与多个推荐模型进行对比实验,结果表明在不同稀疏度的数据集上,L-GAN算法在推荐精度方面有较明显的提高,更善于挖掘数据的隐层特征。
In the traditional recommendation system,users’interests are considered to be stable and unchanged,but in fact,they will change due to various factors.In order to track the changes of users’interests and preferences more advantageously for content recommendation,we propose a recommendation algorithm L-GAN based on generating antagonistic network,which takes advantage of long-term and short-term interest preference to train the recommendation model by generating antagonistic training strategies,so that the recommendation list generated by the recommendation model can be more accurate.In the course of confrontation training,the data is divided into several action cycles,and the user-project evaluation matrix is input in time sequence in each action cycle.The generator model generates the recommendation list,while the judge model distinguishes whether the input recommendation list is similar to the characteristics of the real history record.Finally,by comparing the two open datasets with several recommendation models,the experimental results show that the L-GAN algorithm improves the recommendation accuracy significantly in different sparse datasets and is better at mining hidden features of data.
作者
康嘉钰
苏凡军
KANG Jia-yu;SU Fan-jun(School of Optical-electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)
出处
《计算机技术与发展》
2020年第6期35-39,共5页
Computer Technology and Development
基金
国家自然科学基金(61703278)。
关键词
推荐算法
生成对抗网络
循环神经网络
孪生网络
对比损失函数
recommendation algorithm
generative adversarial networks
recurrent neural network
siamese network
contrastive loss function
