Abstract:

Users' preference mining is one of the key issues in the research field of recommendation system, and it plays a very important role in improving the recommendation performance. Users' preference mining generative adversarial networks (UPM-GAN) is proposed to better analyze the implicit users' preference in the recommendation procedure from two aspects. On one hand, user-rating matrix is processed by the state-of-the-art triplet loss algorithm. It means better positive samples are obtained by the hard-negative mining procedure of the triplet loss algorithm, which will build a strong foundation for more accurately portraying users’ preference. On the other hand, SVD++ algorithm is utilized in turn to create the generation model of the UPM-GAN. The SVD++ algorithm can mine implicit users' preference by adding bias information and latent parameters. It helps improve the rating prediction accuracy of recommendation system. Finally, the state-of-the-art GAN framework is utilized to train the proposed recommendation system and experimental simulation is completed on two mainstream datasets: MovieLens-100K and MovieLens-1M. Experimental results demonstrate that the proposed UPM-GAN is superior to other baselines among all evaluation indices including Precision@K, mean average precision (MAP). Moreover, it has the advantages of faster convergence speed and stable training process. The proposed recommendation system based on UPM-GAN has very large practical value.

Key words: recommendation systems, generative adversarial networks (GAN), users' preference mining, singular value decomposition (SVD++), triplet loss, hard negative samples

摘要：

用户偏好挖掘是推荐系统研究中的关键问题，它对于改善推荐质量具有非常重要的作用。提出用户偏好挖掘生成对抗网络（UPM-GAN），从两个角度深入分析用户隐含偏好：基于三元组损失算法对用户评分矩阵进行处理，挖掘难分负样本，以更好地确立正样本，为准确刻画用户偏好奠定基础；基于奇异值分解（SVD++）算法构建UPM-GAN的生成模型，利用SVD++算法中的偏置信息及隐式参数描述用户隐含偏好，  以提高评分预测精度。最后使用最新生成对抗网络（GAN）框架完成推荐系统训练，在MovieLens-100K、MovieLens-1M这两个主流数据集上展开实验仿真。实验表明UPM-GAN的Precision@K、均值平均精度（MAP）等多项指标均优于对比基线，且它还具有收敛速度快、训练过程平稳等优点。基于UPM-GAN的推荐系统具有一定实用价值。

关键词: 推荐系统, 生成对抗网络（GAN）, 用户偏好挖掘, 奇异值分解（SVD++）, 三元组损失, 难分负样本