Relation2vec: Contextualized network embedding for heterogeneous information networks

From biology to sociology, network theory has been used as a tool for modeling the relationships among entities in complex systems. However, with the emergence of big data, applying the conventional network analysis algorithms, which are often combinatorial, to the contemporary networks is challenging due to the sheer size of the present day network data. On the other hand, deep neural networks (DNNs) have been exceptionally successful at learning from big data and achieved superhuman performance in the fields of computer vision and natural language processing. Hence, there is an opportunity to harness the power of DNNs for network analysis. Since DNNs require their inputs to be in the form of real-valued vectors, we need to use network embedding techniques to generate low-dimensional vectorized representations of the network that accurately encode the rich information and features that are relevant to the network analysis task. In this thesis, we present relation2vec, a contextualized network embedding method for heterogeneous information networks (HINs), in which there are more than one type of nodes and edges. Our method leverages the idea of meta-path, which allows a high-level abstraction of paths and denotes complex relationships among the nodes, to provide context for the nodes in the network, and then utilizes DNN, specifically, bi-directional long short-term memory, to encode the contextual information for each node, along with other rich information and heterogeneity, in the embeddings for both nodes and paths in HINs. We evaluate the performance of relation2vec across DBLP and Yelp datasets via four network analysis tasks: visualization, clustering purity test, node type prediction and path type prediction. The experiment results show that rela- tion2vec outperforms other state-of-the-art HIN embedding methods by 17% to 33% for average cluster purity and on average 0.36 for DBLP and 0.16 for Yelp in f-score for path type prediction. The performance of node type prediction is also competitive with the state of the art.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2020

Author dc:creator

• Zeng, Ziheng

Contributors dc:contributor

• Vasudevan, Shobha

• heterogeneous information networks
• network embedding
• meta-path based
• graph representation learning

Statement dc:rights

• Copyright 2019 Ziheng Zeng

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/106491

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/106491