Graph representative learning

• Main Author: Li, Chuyun
• Other Authors: Lihui CHEN
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/141308
• Institution: Nanyang Technological University

Information networks are commonly used in multiple applications since large amount of data exists in information networks. These data and relationship with different types are of great significant to many application such as community recommendation, users' preference prediction, etc. Generally, the analysis and process of large-scale information networks has attracted widespread attention in the industry. Graph is an abstraction for representing relational data from multiple domains, and graph embedding is a good way for analysis of information networks. A good way of graph embeddings may be obtained by representing node in the graph as a vector, and these representations should contain useful structural information (features) of the graph. In this project, we aim to get feature vectors to represent graphs/sub-graphs in information networks so that machine learning approach can be applied for a classification task. We study two different state-of-the-art representation models: Deep-walk and ASPEM. Both of two algorithms are designed to learn representation of nodes in an information network. Deep-walk is an approach that generates a series of random walks, and the deep-walk based embedding uses Skip-gram to process these generated random walks because of the similarity between the distribution of the random walk in the network and that of sequences in vocabulary; ASPEM is more complicated embedding approach compared with Deep-walk because it proposed a concept of aspect. By dividing multiple aspects, ASPEM can reduce some impacts which effect the outcome of representative learning process and the final classification performance. We also make a comparative analysis of the performance of two algorithms on the node classification task for representative learning on four large-scale real-word data sets. The empirical results show that ASPEM outperforms Deep-walk embedding if multiple types of nodes or edges exists in the network.