Network Representation Learning Methods Based On Deep Learning

In real life,complex systems can be modeled as network structures.They are used to mine valuable information.Due to the intricate connection relationships between nodes,traditional data mining methods cannot be applied well to non-Euclidean data.By preserving the structural and attributive information,network representation learning aims to learn lowdimensional vectors for nodes in a network.The representations of nodes can facilitate subsequent network analysis tasks.Therefore,related works have attracted widespread attention from the academic and business in recent years.The network structures may evolve over time.How to learn the structural information and evolution of the network at the same time are still questions worth exploring.In this paper,the representation learning methods based on static or dynamic network are investigated.Firstly,a network representation learning method based on extractive summary is proposed.This solves the problem that uses redundant sampling sequences in random walk methods.At the same time,it effectively combines the information of structural and attributive information for improving the quality of the representations.Secondly,a dynamic network representation learning method based on aligned network snapshots is proposed.The alignment transformation rule preserves the temporal continuity of network snapshots.And the influence of structural changes is considered for updating of the representation learning.The main contributions of this thesis are as follows:1.Existing methods based on random walk usually use redundant or low quality sampling sequences.Low-quality vector representations generated by these ways affect the performances of subsequent analysis tasks.A network representation learning method based on extractive summary is proposed.Firstly,the additional attribute information is encoded by the encoder,the hidden state vectors and a compressed semantic vector are obtained.Secondly,by using the control characteristics of the selective gate network,the hidden state vectors containing the attribute information are filtered.Finally,the filtered key information and a compressed semantic vector are input to a decoder with attention mechanism,and the high-level features of the nodes without repeated sampling sequences are output.The representation of nodes is denoted by these high-level features.Extensive experimental results on node classification demonstrate that this method can generate high-quality node representation vectors.2.Existing dynamic network representation learning methods usually ignore the continuity characteristics of time between adjacent network snapshots for extracting structural information independently.This results in that network embeddings do not preserve coherence.To solve above problems,an alignment transformation rule is proposed for dynamic network snapshots.This method obtains the transformation coefficients by matching the network snapshots of the previous and current time.Then the network alignment is achieved by using coefficients interpolating the current snapshot.At the same time,a dynamic network representation learning method based on aligned network snapshots is proposed.Firstly,the structural representations of nodes across different times are learned by the structural attention module.Secondly,representations of adjacent network snapshots contain the continuity characteristics of time by alignment transformation rules.Finally,the evolution of the network is captured by the time attention module.The objective function is designed from the spatial and temporal perspectives.The experiments of link prediction show effectiveness of the method in dynamic network.In summary,the network representation learning methods proposed in this paper can generate higher quality low-dimensional vectors,which can be used in static and dynamic networks,respectively.They are conducive to improving the accuracy of subsequent network analysis tasks,such as node classification and link prediction.