推荐算法(四）HGCL

“Heterogeneous Graph Contrastive Learning for Recommendation” (Chen 等, 2023, p. 1) (pdf) 用于推荐的异构图对比学习

背景

“i) how to effectively transfer the side knowledge across different views; ii) how to perform heterogeneous relational contrastive learning with personalized augmentation.” (Chen 等, 2023, p. 2) (pdf) i）如何有效地跨不同视图传递辅助知识； ii）如何通过个性化增强进行异构关系对比学习。

模型

“To represent social relationships among users, graph G_uu = {Vu, Euu } is defined to include user-wise social connections with the edge set Euu . To incorporate item-wise relations, we define the item graph Gii = {Vi, Eii } to connect dependent items with external knowledge (e.g., item category). For these defined graphs, we define three adjacent matrices Aui ∈ Rm×n, Auu ∈ Rm×m and Aii ∈ Rn×n,” (Chen 等, 2023, p. 3) (pdf) 为了表示用户之间的社交关系，图 Guu = {Vu, Euu } 被定义为包括与边集 Euu 的用户方面的社交关系。为了合并逐项关系，我们定义了项图 Gii = {Vi, Eii } 以将依赖项与外部知识（例如，项类别）连接起来。对于这些定义的图，我们定义三个相邻矩阵 Aui ∈ Rm×n、Auu ∈ Rm×m 和 Aii ∈ Rn×n，

初始化

“initialized by xavier initializer [8],” (Chen 等, 2023, p. 3) (pdf) 🔤由 xavier 初始化程序 [8] 初始化，🔤

“The initial embeddings are fed into different graph encoders for user-item domain, user-user domain, and item-item domain.” (Chen 等, 2023, p. 3) (pdf) 初始嵌入被输入到用户-项目域、用户-用户域和项目-项目域的不同图编码器中。

先Xavier初始化用户项目的embeddingE0，然后通过E.q.1来嵌入不同的domain的embedding，Euu是用户-用户图，Eii是项目-项目图

“embeddings E0 uu, E0 ii not only share common semantic with initial embeddings E0 u, E0 i for user-item interactions, but also gain the flexibility to characterize the user-user and item-item relations.” (Chen 等, 2023, p. 3) (pdf) 嵌入 E0 uu、E0 ii 不仅与用于用户-项目交互的初始嵌入 E0 u、E0 i 共享共同语义，而且还获得了表征用户-用户和项目-项目关系的灵活性。

消息传递

UI图的消息传递如E.q.2

“our relation-aware message passing paradigm is configured without transformation and non-linear activation.” (Chen 等, 2023, p. 3) (pdf) 我们的关系感知消息传递范例的配置无需转换和非线性激活。

UU图和II图的消息传递机制和E.q.2类似

异构信息聚合

$f$ 是均值池化

E.q.4是UI图上用户和项目的最终的信息表示

UU和II图的异构信息和E.q.3，E.q.4类似

跨视图的Meta Network

“some users are more likely to be influenced by the recommendations from their social friends, while others often adopt items based on their own preference.” (Chen 等, 2023, p. 4) (pdf) 一些用户更容易受到社交朋友推荐的影响，而另一些用户则经常根据自己的喜好来选择项目。

元知识提取

“Specifically, the distilled meta knowledge for the user-user relation view and the item-item relation view is obtained as follows:” (Chen 等, 2023, p. 4) (pdf) 具体地，获得用户-用户关系视图和项目-项目关系视图的蒸馏元知识如下：

个性化的跨视图知识转移

“Both two parameter tensors contain m matrices for each of the m users. The customized transformations are generated according to the unique characteristics of the corresponding users and items to realize the personalized knowledge transfer.” (Chen 等, 2023, p. 4) (pdf) 两个参数张量都包含 m 个用户中每个用户的 m 个矩阵。根据相应用户和物品的独特特征生成定制转换，实现个性化知识转移。

$\sigma (\cdot )$表示PReLU

最终得到用户表示：

$\alpha$ 是超参

item的表示也是类似的生成

损失

“With this design, the embeddings of the auxiliary views serve as effective regularization to influence the user-item interaction modeling with the self-supervised signals.” (Chen 等, 2023, p. 5) (pdf) 通过这种设计，辅助视图的嵌入可以作为有效的正则化来影响具有自监督信号的用户-项目交互建模。

“The meta network is trained to filter noisy features in the auxiliary views to match the user-item interaction view.” (Chen 等, 2023, p. 5) (pdf) 元网络经过训练，可以过滤辅助视图中的噪声特征，以匹配用户-项目交互视图。为什么？

“. E𝑀 𝑢𝑢, E𝑢 for users, and E𝑀 𝑖𝑖 , E𝑖 for items.” (pdf)

对比学习损失

用户间的对比学习损失：

$e_{uu}$ 哪来的，应该是公式写错了。。

$e_{uu}^M$ 是$E_{uu}^M$ 中的向量，$e_u$ 是$E_u$ 中的向量，s是余弦相似度

item的对比学习损失类似

所以，总对比学习损失：

PBR损失

总损失

实验

“In our evaluation, we perform case studies on Ciao dataset to visualize the learned personalized contrastive transformation matrix (R16×16) to reflect the diverse influence between the auxiliary view (e.g., social relationships) and the user-item interaction view. In Figure 4, we sample four users who are more (e.g., u1481, u3033)/less (e.g., u233, u255) likely to be influenced by social relationships when adopting items. The corresponding personalized contrastive transformation matrices of different users are visualized to capture diverse knowledge transfer between the social view and interaction view. We can observe that larger values in the learned contrastive transformation matrix indicate larger social influence for this user. With the integration of meta network and contrastive learning, the adaptive contrastive data augmentation can be realized based on the personalized characteristics of users.” (Chen 等, 2023, p. 8) (pdf) 在我们的评估中，我们对 Ciao 数据集进行案例研究，以可视化学习到的个性化对比变换矩阵（R16×16），以反映辅助视图（例如社交关系）和用户-项目交互视图之间的不同影响。在图 4 中，我们对四个用户进行了抽样，他们在采用项目时更容易（例如，u1481、u3033）/更少（例如，u233、u255）受到社交关系的影响。不同用户相应的个性化对比变换矩阵被可视化，以捕获社交视图和交互视图之间的多样化知识转移。我们可以观察到，学习到的对比变换矩阵中的值越大，表明该用户的社会影响力越大。通过元网络和对比学习的结合，可以实现基于用户个性化特征的自适应对比数据增强。