Deep learning models are developed under the assumption that the training dataset distribution closely mirrors that of the real world. However, this assumption is often violated, presenting significant challenges. These limitations complicate the relationship between input data x and labels y, as various influencing factors are not adequately represented. This study addresses the issue of distribution shifts, where the training dataset distribution differs from the real-world (x, y). Such shifts can occur in both the marginal distributions of the input P(X) and the label P(Y), complicating the model’s ability to infer the causal relationship between x and y. Adversarial examples, domain changes, and style variations changes in P(X) that can reduce model performance. To enhance robustness against these shifts, it is crucial for models to accurately learn the critical relationships within the data. This study proposes methods to analyze distortions and extract latent factors across various data types and problems. Using domain knowledge-based factor extraction methods, we aim to develop models that are robust against diverse distortions. The proposed multi-factor approach enables robust modeling against distribution changes, securing robustness against real-world distortions unaddressed by previous studies. By extracting and analyzing frequency components of periodic time-series signal data, the study ensures robustness in various distortion scenarios. Additionally, it introduces a novel approach to understanding the dispersion relationships of normal data to detect patterns across diverse environments. This comprehensive analysis of distribution shifts and robust framework for handling joint data distortions provides a foundation for developing deep learning models resilient to real-world challenges. Experimentally, we confirmed that our proposed multi-factor suppression method demonstrates robust modeling across various distribution distortions. Based on our proposed methodology, we achieved robust modeling against various real-world distortions. Our approach not only addresses different data and problem types effectively but also enhances robustness, interpretability, and adaptability by analyzing structural patterns through domain knowledge.
