貢獻

  • 幾位研究者提出了GRAPHEDM(Graph Encoder Decoder Model)模型，將semi-supervised learning(e.g. GraphSage,GCN,GAT)與unsupervised learning of graph representations(e.g. DeepWalk, node2vec)歸納為一個統一的方法。
  • 提供了最新的GRL開源庫於:https://github.com/google/gcnn-survey-paper
  • 從相同的角度檢視GRL不同方面的工作，提出了一個general taxonomy來關注不同工作之間的相似性和差異。

GRL

GRL(Graph Representation Learning) methods aim at learning low-dimensional continuous vector representations for graph-structured data, also called embeddings.

GRL大體上可以被分為兩種：

  • unsupervised GRL: Learning low-dimensional Euclidean representations that preserve the structure of an input graph.
  • supervisedd GRL: Learning low-dimensional Euclidean representations for a specific downstream prediction task such as node or graph classification.(for特殊下游預測任務如節點或圖分類)

GRAPHEDM

GRAPHEDM將現有工作分為四主要類別： * shallow embedding methods * auto-encoding methods * graph regularization methods * graph neural networks(GNNs)

Gneralized network embedding problem

Network embedding: the task that aims at learning a mapping function from a discrete graph to a continuous domain. 研究由離散圖向連續域的映射函數

Formally, given a graph with weighted adjacency matrix(带權鄰接矩陣) , the goal is to learn low dimensional vector representations (embeddings) for nodes in the graph ,such that important graph properties(e.g. local or global structure) are preserved in the embedding space. 將圖轉換為低緯度的向量表達且保留原本圖的一些性質。

A Taxonomy of Graph Embedding Models

The GRAPHEDM framework

Input: the GRAPHEDM framework takes as input an undirected weighted graph(有權無向圖)) , with adjacency matrix , and optional node features

Model: the GRAPHEDM framework can be decomposed as follows:

  • Graph encoder network , parameterized by , which combines the graph structure with node features(or not) to produce node embedding matrix as:

  • Graph decoder network , parameterized by , which use the node embedding to compute similarity or dissimilarity scores for all node pairs, producing a matrix as:

  • Classification network , where is the label space. This network is used in (semi-) supervised settings and parameterized by . The output is a distribution over the labels

Output: the GRAPHEDM model can return a reconstructed graph similarity or dissimilarity matrix (often used to train unsupervised embedding algorithms), as well as a output labels for supervised applications. The label output space varies depending on the supervised application.

TODO

參考

  • [1] Chami, I., Abu-El-Haija, S., Perozzi, B., Ré, C., & Murphy, K. (2020). Machine learning on graphs: A model and comprehensive taxonomy. arXiv preprint arXiv:2005.03675.
  • https://blog.csdn.net/diviner_s/article/details/106978910