A Research On The Approaches Of Feature Selection Via Deep Learning And Its Application

Biomedical field is an important application of artificial intelligence,which is expected to find potential laws from data through machine automation or intelligent analysis,so as to assist medical staff in diagnosis and decision-making.Recently,it has been recognized that there is a close relationship between microbiome and development of human diseases.The goal of microbiome-wide association study is to find microbial markers related to diseases,so as to achieve the purpose of precise medical treatment.However,there are many difficulties in the application of deep learning in data mining microbiome because of large amount,high dimension and high noise of microbiome data,as well as different complex data types,including tabular,graph and time series data type.Aiming at the problem of microbial marker identification,this thesis systematically studies the application of deep learning in different types of microbiome data mining with feature selection,the main research of the thesis includes:(1)When it comes to deep learning,it is believed to be deep neural networks,but deep neural networks are not the whole of deep learning.Although deep neural networks are the mainstream of deep learning,however,any model which could learn high-level representation with deep architecture should be considered as deep learning.The thesis introduced a deep learning model named Deep Forest,which was composed of decision trees.After that,it proposed an ensemble feature learning method and conducted analysis on microbiome data.Compared with deep neural networks,Deep Forest could adaptively learn the layer-wise structure according to the data complexity and the training process was more fast with comparable prediction accuracy.Moreover,the feature selection method based on the deep forest was interpretability and robustness.(2)Considering the complexity of data and the dependence between variables,the thesis proposed a feature selection method based on graph embedding deep feedforward neural network.Compared with the traditional graph embedding method,the proposed method could learn complex and abstract representation with good interpretability,which would be used for feature learning and feature selection.Firstly,the thesis described the background of graph embedding learning,then it summarize the graph embedding learning methods into three categories from the perspective of coder-decoder framework.After that,it compared the performance of each method on link prediction and node classification tasks.Finally,the thesis introduced feature selection based on graph embedding and explained the results in detail.(3)The development of human diseases is gradual change through time,it is of practical significance to predict the further changes from time series data.Due to the heterogeneity and incompleteness of medical data,traditional models such as Hidden Markov Model,which are ineffective.This thesis studied the prediction of infant's response to food allergy via Long Short Term Memory(LSTM)network,ordered Long Short Term Memory network and Gated Recurrent Unit network.Compared with Hidden Markov Model,Support Vector Machines,Random Forest and Lasso,Long Short Term Memory network has high prediction accuracy.It also evaluated whether feature dimensional reduction is helpful to the training.The thesis considered the representation of potential features via autoencoder before its training.In addition,the minimum redundancy maximum relevance and deep forest feature selection methods were also investigated.The organization structure of the thesis was as follows:the first chapter was the introduction part,which introduced the research progress of deep learning and feature selection in the context of big data and summarized the current research progress of interpretable work via deep learning.For the convenience of follow-up elaboration,the second chapter mainly introduced the theory and common of deep learning involved in the full text.The third chapter proposed an ensemble feature selection method and makes a detailed experimental analysis.The fourth chapter introduced a graph representation learning and feature selection method.At the same time,the feasibility and effectiveness of the method were verified.The fifth chapter introduced an approach to combine Recurrent Neural Network and feature selection and conducts a detail analysis;The last chapter was the conclusion of thesis.