Cancer Classification Based On Adversarial Transfer Learning For Gene Expression Profile

Deep learning methods have gained extensive research and application in recent years,due to its powerful feature learning capabilities.Compared with traditional machine learning methods,deep learning methods can automatically learn the higher-order features of the data,saving the work of manually designing features.However,in the scenario of small training sample,the overfitting problem of the deep learning methods are quite serious,which makes it difficult for deep learning methods to obtain satisfactory results.For example,in the classification task of cancer gene expression profile data,the acquisition cost of gene expression profiling data is very high,resulting in very limited training samples.In such circumstance,it is difficult to obtain satisfactory cancer classification results using deep learning directly on a small number of gene expression profile data.Therefore,this article takes multiple related gene expression profile data sets as the research object,and uses transfer learning to mine the knowledge of multiple gene expression profile data,which effectively improves the performance of classification on cancer gene expression profile.This article mainly completed the following two works:(1)There are natural relationships between most cancer gene expression profile data sets,such as acute leukemia and chronic leukemia.So in order to mine the knowledge of the cancer gene expression profile data set to help classification of another cancer gene expression profile,this article proposes a novel transfer learning model called adversarial transfer learning for gene expression profile(ATL-GEP).By introducing generative adversarial networks,the ATL-GEP model can automatically remove the differences between the source domain representation and the target domain representation and thereby learns the shared representation of the two domains;in the target domain task,the ATL-GEP model can further learn the target Local representation of the domain,and at the same time the target domain task learns the weight parameters of shared representation and local representation to determine the degree of representation transfer.The shared representation and local representation are added up by this weight parameters.After several rounds of adversarial learning and representation fusion learning,the ATL-GEP can finally achieve the purpose of knowledge transfer and improve classification performance.Relevant experiments on fourteen cancer gene expression profiles shows that The ATL-GEP model improves the accuracy of two to eight percentage points on most gene expression profiledata sets compared to DNP model which achieves state-of-arts results on cancer gene expression profile.(2)The shared representation learned from the source domain has poor generalization ability if the source domain is also consist by small sample data set,which limits the performance improvement brought by knowledge transfer.Therefore,based on the ATL-GEP model,this paper proposes an adversarial multi-task learning model for Gene Expression Profile(AMTL-GEP).By introducing multiple tasks,the AMTL-GEP model can simultaneously learn the shared representation of multiple domains.The AMTL-GEP model can further mine the internal associations of multiple related tasks and learn local representations on each task.The AMTL-GEP model also adds up representations and shared representations by weight parameters which is learned by task of each domain.At the same time,the performance of multiple related tasks is improved at the same time.The related experiments on 14 cancer gene expression profiles have verified that The AMTLGEP model improves the accuracy of one to eight percentage points compared to the DNP model on most gene expression profile data sets.And in the multi-task experiments of14 tasks,the accuracy of the AMTL-GEP model in each task has been further improved compared to the ATL-GEP model.