A domain generalization algorithm based on causal invariant representation
Aims:Aiming at the current domain generalization methods where the potential causal mechanism between data and labels was not cousidered and the potential semantic dependencies between features was ignored,a domain generalization method realized by decoupling causal invariant representation(DCIR)was proposed.Methods:Firstly,the moment was used to carry out causal intervention on the original data;and the enhanced data with different non-causal characteristics with the same causal characteristics was generated to simulate the variability of the non-causal characteristics.Furthermore,a decoupling proximity factorization algorithm(DFPA)was proposed,based on the empirical covariance matrix,by setting the sliding window and the cutting value,to obtain the invariant representation of features before and after the intervention and the potential semantic dependencies between features.Results:Extensive experiments were conducted on three benchmark datasets,namely Digits,PACS,and CIFAR10-C.The proposed DCIR method achieved average accuracy improvements of up to 13.6% ,26.29% ,and 18.9% ,respectively,compared to the existing baseline models.Conclusions:The domain generalization method implemented through decoupling causal invariant representation can effectively improve the generalization performance of the model.
国家科技期刊平台
NSTL
万方数据
