Given the omnipresence of graph-structured data, graph machine learning has copious applications in multifarious fields such as social media, e-commerce platform, cyber-physical system, or chemical synthesis. Nonetheless, data driven models for graph data also face their unique challenges including over-smoothing caused by message passing-based graph neural networks, structural data sparsity brought by power-law distributions, lack of labelled data due to costly annotations, and noisy signals caused by spurious correlations. In order to address these challenges, other than developing more advanced and complicated machine learning models, graph data augmentation allows researchers to improve graph machine learning from the perspective of data.Works in this thesis develop advanced graph data augmentation techniques for various graph machine learning tasks: node classification, link prediction, and anomaly detection. Differs from traditional ad-hoc data augmentation techniques that integrated augmentation into the learning process of representations and decisions, this thesis introduces learn-to-augment approaches which leverage machine learning models for data augmentation. Therefore, this thesis designs a holistic learning process from data augmentation to representation learning to decision making. Such learn-to-augment approaches are able to achieve superior downstream task performance as well as alleviate the above-mentioned challenges in graph machine learning. Furthermore, by enhancing machine learning from data perspective, graph data augmentation solutions can be used with different graph machine learning models and would not significantly increase the model's complexity.