In recent years, an increasing number of studies have highlighted the critical role of high-quality training data in developing robust and accurate ML models [184, 155, 270, 266]. However, preparing high-quality training data is a laborious and time-consuming process. It requires users to thoroughly examine a large number of samples, identify potential quality issues, and then identify effective strategies to mitigate these issues. In this context, visualization emerges as a promising method that offers significant assistance in these tasks. Consequently, many VIS4AI methods have been developed to streamline the process of preparing high-quality training data. Since each data sample includes an instance, an annotation, and a feature vector, current methods are classified into three categories: instance diagnosis, annotation diagnosis, and feature engineering [152, 270]. Instance diagnosis aims to identify and address issues at the individual data instance level such as missing values, out-of-distribution (OoD) instances, and inexact instances whose content is not easy to recognize. Annotation diagnosis focuses on annotation-level issues, including inaccurate annotations, insufficient annotations, and inexact annotations. Feature engineering seeks to improve the performance of ML models by adding critical features or removing redundant features. Within each category, three settings are considered, including inaccurate, insufficient, and inexact. Figure 3.1 shows the relationships between the data, three settings, and different visualization techniques for data preparation. Furthermore, Fig. 3.2 uses concrete examples to illustrate the intersections between the three settings and three categories of the associated VIS4AI methods. n practical applications, tasks within these categories are not performed separately. Instead, they are conducted iteratively and collectively throughout the data preparation process. This ensures a continuous improvement in data quality and thereby leads to better model performance.