| 摘要: |
| 过去基于学习用户和物品的表征向量的推荐系统算法在大规模数据中取得了较好的结果。相比早期经典的基于矩阵分解（matrix factorization，MF）的推荐算法，近几年流行的基于深度学习的方法，在稀疏的数据集中具有更好的泛化能力。但许多方法只考虑了二维的评分矩阵信息，或者简单的对各种属性做嵌入表征，而忽略了各种属性之间的内部关系。异构信息网络（heterogeneous information network，HIN）相比同构网络能够存储更加丰富的语义特征。近几年结合异构信息网络与深度学习的推荐系统，通过元路径挖掘关键语义信息的方法成为研究热点。为了更好地挖掘各种辅助信息与用户喜好的关联性，本文结合张量分解、异构信息网络与深度学习方法，提出了新的模型hin-dcf。首先，基于数据集构建特定场景的异构信息网络；对于某一元路径，根据异构图中的路径信息生成其关联性矩阵。其次，合并不同元路径的关联性矩阵后，得到包含用户、物品、元路径三个维度的张量。接着，通过经典的张量分解算法，将用户、物品、元路径映射到相同维度的隐语义向量空间中。并且将分解得到的隐语义向量作为深度神经网络的输入层的初始化。考虑到不同用户对不同元路径的关联性偏好不同，融入注意力机制，学习不同用户、物品，与不同元路径的偏好权重。在实验部分，该模型在精确度上有效提升，并且更好地应对了数据稀疏的问题。最后提出了未来可能的研究方向。 |
| 关键词: 异构信息网络 CP张量分解 深度学习 推荐系统 注意力机制 可解释推荐 |
| DOI：10.19363/J.cnki.cn10-1380/tn.2021.09.06 |
| 投稿时间：2021-04-30修订日期：2021-08-05 |
| 基金项目:本课题得国家自然科学基金（No.61876193）资助。 |
|
| Deep learning recommender system based on tensor decomposition of meta-paths of heterogeneous information networks |
|
XU Ronghai, WANG Changdong
|
| (Department of Computer Science and Engineering, Sun Yat-Sen University, Guangzhou 510000, China) |
| Abstract: |
| The existing representation learning based recommendation methods have achieved satisfactory results in large-scale data. Compare to the early classical matrix decomposition methods (matrix factorization, MF), the popular deep learning based methods have better generalization capability in sparse data in recent years. However, many methods only consider the two-dimensional scoring matrix information or simply do embedding vectors for various attributes, while ignoring the internal relationships between various attributes. Heterogeneous information networks (HIN) can store richer semantic features than homogeneous networks. In recent years, by combining heterogeneous information networks with deep learning for recommender systems, it becomes a hot research topic to mine key semantic information through meta-paths. Aiming at mining the relevance between the various auxiliary information and users' preferences, by combining tensor decomposition methods, path information of heterogeneous information networks and deep learning methods, this paper propose a new model called hin-dcf. Firstly, a heterogeneous information network of a specific scene is constructed based on dataset. And a correlation matrix for certain meta-path is generated according to the path information of heterogeneous network. Secondly, after merging the correlation matrices of different meta-paths, a tensor is obtained which contains three dimensions of user, item and meta-path. Then, by the classical tensor decomposition methods, the user, item, and meta-path are mapped into the same dimensional hidden semantic vector space. The hidden semantic vectors obtained by tensor decomposition are used as the initialization of the embedding layer of the deep neural network. Considering the different relevance preferences of different users for different meta-paths, an attention mechanism is incorporated to learn the preference weights of different users, item pairs, for different meta-paths. Experiments show that the hin-dcf model has achieved improvements in terms of both accuracy and convergence speed, and better copes with the sparse data problem. Finally, the possible research directions in the future are proposed. |
| Key words: heterogeneous information networks CP tensor decomposition deep learning recommendation system attention mechanism explainable recommendation |
