Research On Network Representation Learning Based On Stacking Ensemble And Graph Neural Network

In the real world,many relationships between data entities can be modelled as networks,such as social networks,paper citation networks,biological/chemical networks,traffic flow networks,and so on.Methods of network representation learning are an important way to process network data by representing the nodes in a network as lowdimensional dense vectors,and these vector representations are able to maintain the relevant structural information of the original network.The core mechanism of GNN is to learn by aggregating information from neighbours,and the aggregation is also a kind of "ensemble" of information from neighbours."This is the starting point of this paper,which inspires us to explore the integration capabilities of GNNs.Most existing network representation learning methods have weak generalization ability,and ensemble learning theory can effectively solve these problems.A strong learner formed by effectively integrating multiple learners can always achieve better performance than each learner,its generalization performance is better than each learner.At present,there are relatively few studies that apply ensemble learning to network representation learning.In addition,the design of stacking ensemble learning is more flexible.Therefore,this paper applies stacking ensemble learning to the study of network representation to obtain the more effective network representation by designing an effective ensemble model.Two types of stacking ensemble models are proposed based on the effectiveness of ensemble:(1)a network representation learning method for node feature ensemble;(2)a network representation learning method for network structure ensemble.Experimental results show that the performance of both ensemble methods have been effectively improved.experimental validation is performed on several real network datasets,that is,the experimental results indicates the network representation capability and ensemble ability of GNN.The main research of the thesis has the following two aspects:Firstly,the node-feature-oriented network representation learning ensemble method,referred to as SE-NRL,is proposed.Firstly,for the same network,the classical network representation learning methods Deep Walk,Node2 Vec and Line based on shallow neural networks are selected as primary learners;Secondly,the output results of the primary learners are combined to obtain new feature vectors;Finally,the graph neural network GAT which is simulated human graph attention mechanism is selected as the secondary learner to stack heterogeneous ensemble of the new feature vectors to obtain the final network representation.In addition,the loss function and five evaluation metrics were designed based on the unsupervised nature of the training process.Secondly,the network structure-oriented network representation learning ensemble method,referred to as SCE-NRL,is proposed.The clustering ensemble method combines multiple base clustering results into one cluster,which can effectively fuse the advantages of different algorithms.In essence,it is also a stacking ensemble method.For the same network,a few community detection results with high accuracy and high variability are firstly selected as base clustering members using a suitable selection strategy.Then,the network representation is obtained by integrating the clusters of the base cluster members using the graph convolutional network GCN,in which the loss function is designed.Finally,the network representation was used in a community detection task to evaluate the community detection results using modularity evaluation metrics.