Heterogeneous information fusion in network embedding for data mining applications

Abstract

We live in a world full of networks, instinctively and inevitably. Human beings by nature are social animals. In modern life, we can extend our social relationships to others in any place of the world with the help of world wide web, which can essentially connect all entities attached it. Besides, human body is comprised of biological molecules, and biological molecules interact to realize biological functionalities. Analyzing the networks has profound meanings, e.g., studying the categories of entities can better understand the networks, studying the groups of entities can partition the large networks into smaller sub-networks, and studying the relationships among entities can obtain insights into how the networks evolve. We can see that these network applications either can benefit from utilizing the relationship information, e.g., the category of a particular entity is likely to depend on whom it usually interacts with, or directly target at relationships, e.g., the network evolvement. In big data era, we are fortunate to have access to much of the network information including both entities and relationships. All these information provides huge opportunities for machine learning and data mining models whose success largely depends on the data fed to them. However, the network representation, e.g., adjacency matrix, is usually high-dimensional and sparse, which is not suitable for the main steam but tuple-based machine learning and data mining models, such as support vector machines (SVMs) and artificial neural networks. Moreover, many networks usually have heterogeneous information, i.e., multiple types of entities and interactions. In this thesis, we thus study heterogeneous information fusion in network embedding, which is to learn low-dimensional and dense node representations from multiple heterogeneous network information for data mining applications.