Extracting And Encoding Method Of Data And Feature Correlation By Sparsity Assumption

Correlation is one of the most important part of information in data mining,in this thesis,sparsity is ultilized to improve the efficiency and exlainability of data or feature corrlation extracting algorithms in dynamic systems or heterogeneous graphs.For heterogeneous graph,graph neural network with multi-channel encoding scheme is developed,the method is based on partial relationship assumption and is a simplified version of graph attention network,which can achieve almost the same recommendation performance with graph neural network with lighter cost,besides,a novel experiment about feature correlation is proposed to show the reason of performance enhancement in the proposed method.For dynamic system,an adaptive sparse connectivity matrix extracting method is proposed,the method extends tradition LASSO regression based model with meta-learn,which encodes intra-system shared pattern with initialized parameter and inter-system intrinsic pattern with a sparse penalty matrix generated by variational autoencoder,therefore,the model combines two kinds of meta-learn method for better explainability and classification performance.At last,we list several drawbacks of the two methods,and discussed how we could further improve them.