基于生成对抗网络的协同过滤推荐

Collaborative Filtering Recommendation Based on Generative Adversarial Network

针对个性化推荐中存在的数据稀疏问题,文章提出基于生成对抗网络(GAN)和知识图谱的协同过滤推荐算法。该算法利用知识图谱将用户简单的序列行为提取为语义信息并构建用户行为路径;针对稀疏路径,提出基于序列对抗网络生成伪行为路径的方式填充稀疏数据,提高推荐性能。在阿里真实数据集UserBehavior上的实验结果表明,基于GAN的行为路径协同过滤推荐算法,较原来的行为路径协同过滤算法,在可推荐人数上最多可提高约104%,在覆盖率上最多可提升一个数量级。同时,结合该算法的混合推荐算法在收藏、加购和购买3个维度上正确率分别提高7.9%、2.6%和2.1%。

A collaborative filtering recommendation algorithm based on generative adversarial network(GAN)and knowledge graph is proposed to solve the problem of data sparsity in personalized recommendation.The algorithm utilizes knowledge graph to extract semantic information from user’s simple sequential behavior to construct user behavior paths.For sparse paths,a method of generating pseudo behavior routes based on sequence adversarial network is raised to increase the amount of user behavior routes to improve the recommendation performance.A series of experiments conducted on a real data set UserBehavior show that the proposed collaborative filtering recommendation algorithm is well behaved.The recommended number of users can be enhanced by about 104%,and the coverage can be improved by up to an order of magnitude.Additionally,the precision of the combined recommendation method incorporating the proposed algorithm into collaborative filtering is enhanced by 7.9%,2.6%,and 2.1%in three dimensions,respectively.