<p>Demonstrating robustness to variations in viewpoint, human body scale, and movement speed, alongside real-time and online performance, 3D skeleton-based action recognition has emerged as a prominent topic in computer vision. Over the years, this task has been approached using two primary methods: traditional handcrafted features and deep-learning-based features. However, existing surveys on action recognition have predominantly emphasized methods relying on RGB data, depth maps, video sequences, and skeleton-based inputs. The limited reviews specifically addressing skeleton data primarily focus on its representation or the performance of certain classic techniques on specific datasets. Moreover, despite the widespread application of deep learning techniques in this domain, no prior research has comprehensively reviewed the field from the perspective of deep learning architectures. To address these gaps, this work systematically underscores the importance of action recognition and the critical role of learning methodologies. Furthermore, it provides an extensive overview of Self-Supervised, Semi-Supervised, Unsupervised, and Reinforcement learning techniques as key approaches to action recognition. All papers published in the last decade range from 2015 to 2025, and most importantly, the majority are from top journal papers, such as TPAMI, CVPR, AAAI, etc. Secondly, we discuss the most commonly used and significant 3D skeleton datasets, such as NTU-RGB+D 60/120, UCF50, Kinetics, UAV-Human dataset, and ANUBIS datasets. This review thoroughly examines deep learning techniques employed for action recognition using 3D skeleton data. Additionally, it highlights the challenges in this field and outlines promising directions for future research on skeleton data-based action recognition. In conclusion, this work is a valuable resource for new researchers seeking to explore and advance this study area.</p>