Research On Network Representation Learning For Dynamic Heterogeneous Information Network

Machine learning on graphs is an important and ubiquitous task,and its application has gradually spread throughout our daily lives.The main challenge in this research domain is to find a way to represent or code the graph structure so that machine learning models can easily exploit it.However,traditional graph based representation methods typically rely on adjacency matrices,adjacency tables,or design features,which makes it not suitable for machine learning tasks on today's large-scale networks.Network representation learning aims to represent nodes in the network as low-dimensional dense vectors and solve subsequent machine learning tasks with these low-dimensional vectors.In recent years,this method has become a new research hotspot,since it can effectively reduce the complexity of various network analysis tasks.However,most existing network representation learning methods are only aimed at static networks,and cannot perform dynamic representation learning for evolving networks.Regarding the above problem,this thesis studies the problem of dynamic heterogeneous information network representation learning and uses random walk-based method to perform representation learning of nodes.Random walk-based methods first construct node sequences via random walk and then throw these node sequences into SKIP-GRAM model to learn the low-dimensional representation of nodes.However,existing random walk-based methods almost do not take advantage of network topology,semantics,text and time information together to improve the effect of network representation learning.In view of this,this thesis proposes a new representation learning method for heterogeneous information networks based on dynamic random walk.This thesis constructs a dynamic heterogeneous information network fir-st,and then proposes an automatic method to mine and extend meta-paths.Next,an edge-based dynamic random walk method combined with timestamps are proposed to construct node sequences,where two random walk control strategies are adopted for adjusting length and quantity of node sequences so as to improve the quality of node sequences in real time.Finally,the node sequences are thrown into a heterogeneous SKIP-GRAM model,and the low-dimensional vector of nodes is learnt.We use two different real datasets to validate our proposed method.Extensive experimental results show that the dynamic heterogeneous information network representation learning method proposed in this thesis outperforms state-of-art methods in various mining tasks,such as node classification,clustering and visualization.This thesis first introduces the reseairch status of network representation leairning and analyzes the problem of existing methods.Then,our proposed dynamic representation learning method for heterogenous information network is elaborated,followed by its design and implementation.Next,the effectiveness and accuracy of our method are verified by experiments.Finally,we conclude this thesis and give the future work.