Research And Implementation Of Multi-Relational Network Representation Learning Algorithm

With the emergence of various types of information networks,how to mine and analyze network data becomes an urgent problem to be solved.Due to the complexity of real network information,more and more researches are not limited to the feature analysis of a single structure,but more focused on the mining of multi-dimensional and multi-relational network and the calculation and analysis of multi-source data.Among them,multi-relational network means that the nodes in the network contain a variety of different relationships,which preserves more complete topological characteristics of network nodes and network dependence characteristics of different networks.Network representation learning can effectively analyze and mine network data,and support subsequent network processing and analysis tasks through low-dimensional embedding of learning nodes.Therefore,how to use network representation learning to analyze multi-relational network data and quickly obtain comprehensive information representation is a significant task.Therefore,in the paper,the study of representation learning of multirelational network is carried out.The main contents are as follows:1.Network representation learning algorithm based on disentangled learning DAME is proposed.Firstly,the multi-relational network is modeled as a multiplex network,and generative adversarial learning is used to preserve the consistent information and complementary information in the multiplex network,and disentangled learning is introduced to capture the dependence between node attributes and topologies of each layer.Experiments on different data sets show the effectiveness of the model DAME on two downstream tasks.2.Parallel multi-relational network representation learning algorithm PDAME is proposed.The graph sampling technique is used to sample the large-scale multiplex network and decompose the large graph.Based on the parallel computing framework,module decoupling of DAME algorithm is carried out,and the parallelization strategy of each module is formulated to accelerate the running speed of the model.Experiments demonstrate that the parallel algorithm has some performance advantages over the previous algorithm.3.Network representation learning component is designed and implemented.The component encapsulate network presentation learning algorithms and is integrated into the big data analysis platform BDAP to provide users with more analysis functions.Finally,the usability of the system component is proved by testing the function of the component module.