Network embedding and its applications

Hou, Chengbin ORCID: 0000-0001-6648-793X (2022). Network embedding and its applications. University of Birmingham. Ph.D.

Preview

Hou2022PhD.pdf
Text - Accepted Version
Available under License All rights reserved.
Download (2MB) | Preview

Abstract

Apart from the attached attributes of entities, the relationships among entities are also an important perspective that reveals the topological structure of entities in a complex system. A network (or graph) with nodes representing entities and links indicating relationships, has been widely used in sociology, biology, chemistry, medicine, the Internet, etc. However, traditional machine learning and data mining algorithms, designed for the entities with attributes (i.e., data points in a vector space), cannot effectively and/or efficiently utilize the topological information of a network formed by relationships among entities. To fill this gap, Network Embedding (NE) is proposed to embed a network into a low dimensional vector space while preserving some topologies and/or properties, so that the resulting embeddings can facilitate various downstream machine learning and data mining tasks.

Although there have been many successful NE methods, most of them are designed for embedding static plain networks. In fact, real-world networks often come with one or more additional properties such as node attributes and dynamic changes. The central research question of this thesis is "where and how can we apply NE for more realistic scenarios?". To this end, we propose three novel NE methods, each of which is for addressing the new challenges resulting from one type of more realistic networks. Besides, we also discuss the applications of NE with the focus to the drug-target interaction prediction problem.

To be more specific, first, we investigate how to embed the attributed network, which can better describe a real-world complex system by including node attributes to a network. Previous Attributed Network Embedding (ANE) methods cannot effectively embed attributed networks especially when networks become sparse, and/or are not scalable to large-scale networks. To deal with these challenges, we propose a scalable ANE method to effectively and robustly embed attributed networks with different sparsities. Second, we study how to embed the dynamic network, which is often the case in real-world scenarios as real-world complex systems often evolve over time. Most previous Dynamic Network Embedding (DNE) methods try to capture the topological changes at or around the most affected nodes and accordingly update node embeddings. Unfortunately, this kind of approximation, although can improve efficiency, cannot effectively preserve the global topology of a dynamic network at each timestep, due to not considering the inactive sub-networks that receive accumulated topological changes propagated via the high-order proximity. To tackle this challenge, we propose a DNE method for better global topology preservation. Third, comparing to static networks, dynamic networks have a unique character called the degree of changes, which can be used to describe a kind of dynamic character of an input dynamic network about its rate of streaming edges between consecutive snapshots. The degree of changes could be very different for different dynamic networks. However, it remains unknown if existing DNE methods can robustly obtain good effectiveness to different degrees of changes, in particular for corresponding dynamic networks generated from the same dataset by different slicing settings. To answer this open question, we test several state-of-the-art DNE methods, and then further propose a DNE method that can more robustly obtain good effectiveness to the dynamic networks with different degree of changes. Fourth, regarding a specific application of NE to a real-world problem, we propose a NE based Drug-Target Interaction (DTI) prediction method by additionally utilizing the two implicit networks which are extracted from a given DTI network but are ignored in previous DTI prediction methods. A case study indicates that the proposed method can predict novel DTIs.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):