A Study on the Representation Learning Method of Temporal Social Networks


Student thesis: Doctoral Thesis

View graph of relations


Related Research Unit(s)


Awarding Institution
  • Wenbin Hu (External person) (External Supervisor)
  • Xiaohua JIA (Supervisor)
Award date27 Jun 2022


Compared with static social networks, temporal social networks (TSN) record the whole network's growth process and evolution trend. The time-dimensional information maintained by TSN provides richer semantics for network analysis. Network representation learning (NRL), also known as network embedding, is one of the critical methods for social network analysis. NRL is committed to learning the low dimensional and dense vector representation of network nodes to preserve the topology and internal properties of the network. By transforming graph structure data into European spatial data, NRL provides excellent convenience for introducing machine learning methods into network analysis and improving the efficiency of network analysis. At present, most NRL methods are devoted to static social networks, while the research on temporal social network representation learning is still in its infancy. While providing more time-dimensional information, TSN also poses a series of challenges: 1) The structure of TSN is highly nonlinear. It is a great challenge to preserve the highly nonlinear structure of TSN. 2) Side information such as node attributes is another critical information source for network embedding. Taking into account the dynamic topology and side attribute information in the representation learning process is the second challenge. 3) Evolution mechanism is an essential property of TSN. Analyzing the historical evolution process of TSN and learning the representation of abstract evolution mechanisms is also a significant challenge. 4) There may be network noise such as expired links and false links in the growth process of TSN. The fourth challenge is to eliminate the impact of network noise and preserve the actual network structure.

In this thesis, we systematically investigate the problem of temporal social network representation learning by utilizing some useful technologies, such as auto-encoder, graph neural networks, motifs, and link prediction. The novel contributions are summarized as follows:

1) Aiming at the challenge of nonlinear structure preservation, a high-order nonlinear information preserving representation learning method is proposed in this thesis. Firstly, we define three kinds of temporal proximities of nodes in TSN based on a time exponential decay model. Then, we propose a temporal random walk algorithm to embed the temporal proximity into the walking path. Finally, we propose a novel deep guided auto-encoder to learn node representations for networks. This auto-encoder has multiple layers of nonlinear functions to capture the highly nonlinear network structure. By training the proposed auto-encoder with the walking path set, our method can preserve the temporal proximity and highly nonlinear structure of TSN.

2) Aiming at the challenge of side information preservation, an attributed temporal network embedding method is proposed in this thesis. The core of this method is to dynamically, and orderly aggregate node attributes to generate node representation vectors based on the topology of TSN. We propose a biased sampling algorithm for neighbor nodes sampling by exploring the historical interaction information between nodes. This algorithm is based on the point process to model the relationship strength between node vi and its neighborhood. After sampling neighbor nodes, this method proposes an inductive representation learning algorithm, which trains an inductive graph convolution neural network model for network representation learning.

3) Aiming at the challenge of conceptual evolution mechanism representation learning, a triad-based network evolution mechanism representation learning method is proposed in this thesis. This method takes a triad transition matrix (TTM) as a carrier to represent the abstract evolution mechanism. Firstly, we analyze the feasibility of using the TTM as the carrier by counting the number distribution of different triads in a mechanism controllable artificial network. Then, we propose a CNN-LSTM model to learn the evolution law of TTM. Finally, we propose a link prediction algorithm based on TTM to verify the ability of this method. The results in real social networks and evolution mechanism controllable artificial networks indicate that the proposed method can well preserve the abstract evolution mechanism of TSN.

4) Aiming at the challenge of noise-resilient network representation learning, a noise-resilient high-order temporal similarity preserving network representation learning method is proposed in this thesis. Firstly, considering the time-dimensional information of TSN, we modify five node similarity indexes and define a comprehensive similarity index. Based on the comprehensive index, we propose a high-order similarity construction algorithm to construct a high-order temporal similarity matrix S. An accurate high-order similarity can serve as a metric of the authenticity of the observed network structure. Then, we correct the first-order temporal similarity based on the constructed S and propose a correction matrix construction algorithm to construct a correction matrix C. Finally, we propose an embedded model for network representation learning. The model considers both C and S. By optimizing the model, this method can effectively save the real network structure of TSN.