Adaptively Non-negative Latent Factorization Of Tensors Based Dynamical Network Representation Model And Application

Recent years researches on network representation learning are more focus on simple network.However,with the development of computer technology,both the network relationship of natural things and the network relationship of industrial applications have changed from static to dynamic.With the cryptocurrency developed,it has built a vast and complex network of transactions.As Ethereum develops rapidly and the involved accounts increase drastically,its transaction network form a high-dimensional and incomplete(HDI)state.At the same time,this transaction network still contains the following characteristics: dynamic,directed,weighted,and multiple.This kind of network is called HDI dynamic and complex Ethereum transaction network in this paper.In order to perform NRL on HDI dynamic and complex Ethereum transaction network to extract the latent features from it,an adaptively non-negative latent factorization of tensor merging auxiliary matrix based network representation model has proposed in this paper.The adaptively non-negative latent factorization of tensor merging auxiliary matrix based network representation learning model is mainly based on the following four ideas:1)using tensor structure to describe the dynamic,direct and weight information of HDI dynamic and complex Ethereum transaction network;2)using matrix structure to model the multiplicity of edges of HDI dynamic and complex Ethereum transaction network;3)building a set of non-negative learning objective updating rules based on the principle of data density-oriented;4)implementing model hyper-parameter self-adaptive via using a particle swarm optimization(PSO)algorithm in the training process.network representation learning and potential transaction link prediction experiments are conducted on several real Ethereum transaction datasets in this paper.The experiment results show that the adaptively non-negative latent factorization of tensor merging auxiliary matrix based network representation learning model not only outperform the similar network representation models in computational efficiency in HDI dynamic complex network representation learning,but also outperform the similar models in prediction accuracy of potential transaction links.