Characterisation of Data Augmentation Techniques Using Visualisation

Data Augmentation (DA) alleviates the data-space limitations by generating new instances for training machine learning (ML) models. However, in practice, the effectiveness of the DA techniques adopted in varying downstream tasks lacks analytical explainability, mainly due to its domain-specificity nature. This paper describes a new ideation of characterising DA techniques from a visualisation perspective by explaining feature-space characteristics of augmented and original data in the high-dimensional space. This is accomplished by assessing the feasibility of using principal component analysis (PCA) for capturing the distinctiveness inherited by DA techniques. In a problem setting of COVID-19 CT detection, the proximity analysis suggests that computing PCA using low-dimensional features can contribute significant analytical values for the characterisation of DA techniques. The comparison of the clustering performances using data instances from the image and feature domains further highlights the significance of abstract features for more effective visual analysis and interpretation. Furthermore, the non-trivial connection between feature inputs and image outputs demonstrated in the connectivity analysis also indicates the significance of visualisation in analysing ML decisions, particularly with high-level visual explanations.