Abstract: Deep learning,as a representative of machine learning methods,has been widely applied and achieved many successes.However,problems such as dataset distribution shift and long-tailed distribution can significantly degrade the performance of traditional deep learning methods,and these two issues often exist in real-world datasets.Although domain generalization and long-tailed learning research have provided good solutions to these two problems separately,the effect of a single domain generalization or long-tailed learning method is not satisfactory in the complex scenario of combining distribution shift and long-tailed distribution(LT-DS).To address the LT-DS problem,a unified approach can be taken from a causal perspective to solve both issues si-multaneously.For distribution shift,causal intervention and decomposition can be achieved through Fourier transform,and cross-domain invariant causal feature representations can be obtained through decorrelation weighting.For long-tailed distribution,a causal effect classifier can be constructed through debiasing training to eliminate momentum-induced biases,and further eliminate the impact of long-tailed distribution through Balanced Softmax and logit adjustment.Experimental results show that this method outperforms the best existing methods by an average of 8% and 5% on the AWA2-LTS dataset and ImageNet-LTS dataset,respectively,demonstrating competitive results on the LT-DS problem.

Key words: Deep learning, Domain generalization, Long-tailed learning, Causal inference