Representation Learning For Large-scale Attributed Networks

There are many complex network systems in real life such as social networks,biomolecular networks,and the Internet.The research on these complex systems has provided many innovative applications for society.Network representation(also known as network embedding)is a method that represents network nodes as low-dimensional continuous space vectors while preserving the network structure and its inherent properties.It is an effective method that promotes significant progress in downstream network data mining tasks.In recent years,it has received great attention from academia and industry.The industry has incorporated network representation learning technology into the basic architecture of the next-generation network analysis platform.This thesis proposes the following research contents for the efficient representation learning problem in large-scale attributed networks:(1)The thesis proposes a random projection-based efficient attributed network dynamic representation update algorithm for large-scale attributed networks dynamic representation learning.The algorithm is divided into two parts: static representation learning and dynamic representation update.First,the static representation learning part combines the high-order structural proximity and high-order attribute proximity of nodes in the network representation.Second,we design a pre-projection mechanism to efficiently model the proximities based on the associative law of matrix products.Finally,the dynamic representation update part realizes the efficient iterative update of the representation of the large-scale attribute network through incremental matrix calculation.(2)Aiming at the efficient representation learning problem of large-scale attributed multiplex heterogeneous networks,the thesis proposes an integrative architecture with scalable graph transformation and sparse random projection to automatically preserve both attribute semantics and multiplex relations in the learned embedding.First,the model decouples the network into multiple homogeneous and bipartite sub-networks to differentiate each relationship between nodes in the network.Second,the architecture performs graph transformation on these sub-networks to automatically obtain the meta-path proximities of different lengths between nodes.Finally,the meta-path proximities and node attributes are fed into a random projection-based architecture to efficiently learning the network representation.(3)The thesis proposes a motif-preserving dynamic attributed network representation learning framework to model the high-order dynamic characteristics of the network.First,the framework uses a motif-preserving graph neural network that integrates the adjacency features to capture the proximities of pair-wise nodes and motif features of the network to capture the local high-order structural proximities at different levels and scales.Second,the framework uses a temporal shift module that simulates the one-dimensional convolution operation on time dimension to effectively and efficiently capture the time evolution relation between different time snapshots.Finally,under the joint action of the motif-preserving graph neural network and the temporal shift model,the historical information of the network is merged into the dynamic representation of the attributed network.This thesis also verifies the effectiveness and efficiency of the three methods on multiple large-scale real-world datasets.