Graph Representation Learning Based On Graph Partition Sampling Algorithm

Machine learning on networks is an universal and important task.In the field,sparsity of graph structure has to be resolved.The normal method is to encode graph structure to be used easily by machine learning algorithm.Traditional method relied on manual desgined algorithm to encode graph structure and extracted graph feature information further.However,this kind of method costs a lot and lacks flexibility.Recent years,using a method maps or encodes graph structure to low dimension space autoimatically becomes popular.This kind of method mainly bases on deep learning,matrix fraction or nonlinear dimensionality reduction.Method based on deep leaning can introduce attribute vector of nodes to learning more effective representation vectors.However,in this kind of scene,larger scale of graph or higher dimension of feature vector will cause the accessing feature matrix to be the training bottleneck.This paper aims to design a method based on graph partition,to leverage nueral network and attribute vectors without sacrificing the traning performance.Belows are the main contribution:· This paper reviews graph representation learning algorithms in detail.During the methods based on deep learning,little attention is paid to the effect of graph sparsity.This article optimizes the state of art algorithm using graph partition method combining the feature of power-law distribution and sampling method.· This paper bases on graph partition method,combines coarse-grained and finegrained sampling method to control the size of feature matrix in GPU.This paper desgins a buffer pool in GPU to cache splited feature matrix blocks to reduce the overhead of accessing feature matrix.During the train process,this paper provide hierarchical nagetive sampling to reduce the time complesxity.· This paper leverage the graph computing method to optimize the graph representation learning and uising experiments to prove the performance and correctness.The result shows that this paper achieves 4 to 7 speedup compare to the baseline algorithm on different scales of the real world datasets using mean aggregator and achieves 1.5 to 1.7 speedup on other aggregators.This paper also provide competitive accuracy during the downstream machine leaerning tasks.