Multi-nodes Link Prediction Approach Based On 3D Convolutional Neural Network

The opportunistic network is one of self-organizing networks using the encounter chance brought by node mobility to communicate,and it does not need a complete link between source node and destination node.It is more conformed for the demand of self-organization in real environment because of the characteristics such as dynamic topology,limited resources,non-connected structure,etc,which has become a hot spot of academia in recent years.Link prediction is a difficult problem in the research of opportunistic network,and its goal is to estimate the possibility of the link among nodes according to the known link and the properties of these nodes.Effective link prediction method not only can find the potential relationship among nodes,but also can further analyze the law of message propagation,thereby providing support for the upper application of opportunistic network.At present,the research of link prediction mainly focuses on the single node-pair in the social network which the topological structure is relatively stable.Aiming at the multi-nodes situation in opportunistic network,this thesis proposed a link prediction method based on deep learning,using the history information of global topology evolution to infer the possible change trend of local link in the future: 1)in the stage of representation,chaos theory is applied to determining the length of slice-time,and time series analysis method is adopted to quantify the dynamic process.After looking for possible attribute of opportunistic network under different dimensions such as time,space and correlation through analyzing the characteristics of the different scenarios,an integrated representation method of opportunistic network is proposed by using multidimensional attribute;2)in the stage of modeling and predicting,pattern classification method is adopted to classify of multi-nodes future link combination.In terms of 3d convolution neural network,the structure features is extracted which can influence local link state from multidimensional attribute.The evolution trend of link in the future is inferred according to these features,so as to realize the multi-nodes link prediction.The thesis uses the ITC dataset of Dartmouth College.The experiments are conducted on the Sklearn and Keras,the common performance metrics are adopted in classification problem such as Accuracy,Precision and AUC(Area Under the Receiver Operating Characteristic Curve),to evaluate prediction performance.Theexperiments are divided into two parts: 1)model selection,using validation-set to finish the process of prediction model selection after setting the different scenarios such as slice-time length,frame size,and the number of nodes,so as to determine the optimal 3d convolution neural network(3D-CNN);2)performance comparison,using test-set to realize the comparison between 3D-CNN and other prediction models based on similarity metrics,so that the effectiveness of 3D-CNN is verified.The experimental results show that 3D-CNN method has better precision and stability compared with ones based on CN,AA,RA,Jaccard and Katz.