SAPF: A Similarity-Based Pre-Evaluation Framework for Determining Effective Augmentation Ratios in Time-Series Data
1
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
In deep learning, data augmentation is key for enhancing model performance and improving generalization. However, its effectiveness depends not only on the applied augmentation method but also on the augmentation ratio, whichdetermines the scale of the generated data. Conventional approaches either arbitrarily set this ratio or use repetitive modeltraining under all conditions, resulting in high computational costs and limited practicality. To overcome these limitations, wepropose a Similarity-based Augmentation Performance Framework (SAPF) that predicts the optimal augmentation ratio withoutiterative training. SAPF extracts embedding vectors from a Bidirectional Gated Recurrent Unit(Bi-GRU) model trained on theoriginal dataset and quantifies the distributional difference between the original and augmented data using the WassersteinDistance (WD). By analyzing the WD growth pattern across augmentation ratios (×2 to ×100), SAPF identifies the optimal ratioas the saturation point where further augmentation yields negligible distributional change. Experimental results show that WDand classification accuracy increased rapidly at lower ratios (×2 to ×25) and then saturated, with a strong positive correlation( = 0.94–1.00) between WD and accuracy. Furthermore, SAPF maintained comparable performance while reducing thetraining time by more than 99.28% compared to conventional methods, thereby demonstrating its efficiency and practicality indesigning effective augmentation strategies.
Abstract
In deep learning, data augmentation is key for enhancing model performance and improving generalization. However, its effectiveness depends not only on the applied augmentation method but also on the augmentation ratio, whichdetermines the scale of the generated data. Conventional approaches either arbitrarily set this ratio or use repetitive modeltraining under all conditions, resulting in high computational costs and limited practicality. To overcome these limitations, wepropose a Similarity-based Augmentation Performance Framework (SAPF) that predicts the optimal augmentation ratio withoutiterative training. SAPF extracts embedding vectors from a Bidirectional Gated Recurrent Unit(Bi-GRU) model trained on theoriginal dataset and quantifies the distributional difference between the original and augmented data using the WassersteinDistance (WD). By analyzing the WD growth pattern across augmentation ratios (×2 to ×100), SAPF identifies the optimal ratioas the saturation point where further augmentation yields negligible distributional change. Experimental results show that WDand classification accuracy increased rapidly at lower ratios (×2 to ×25) and then saturated, with a strong positive correlation( = 0.94–1.00) between WD and accuracy. Furthermore, SAPF maintained comparable performance while reducing thetraining time by more than 99.28% compared to conventional methods, thereby demonstrating its efficiency and practicality indesigning effective augmentation strategies.