<p>One-shot action recognition is a challenging problem in computer vision requiring only one example to recognize an action, with practical applications in fields like surveillance, robotics, sports analysis, and augmented reality. Many methods have been proposed for one-shot action recognition, but they often suffer from low accuracy, high computational costs, or slow processing speed. Therefore, there is a need for new approaches that can improve the accuracy and efficiency of one-shot action recognition. This paper introduces a new data representation, the three-dimensional compressed action matrix (CAM), designed to enhance the separability of human actions from skeletal data. In our framework, the motion path of each keypoint is transformed into the compact CAM structure, which encodes spatio-temporal information in a highly discriminative format. This powerful representation is then processed by a remarkably lightweight convolutional neural network (CNN) featuring only two convolutional layers and a multi-similarity loss function. The effectiveness of this representation-centric approach is demonstrated on the NTU RGB + D 120 dataset. When trained on just 20 classes, our model achieved 56.2% accuracy, surpassing many prior methods that were trained on 100 classes. More significantly, when trained on 100 classes, our model reached 71.25% accuracy using only 1.9 million parameters, outperforming all previous state-of-the-art methods while being vastly more efficient. This work proves that by engineering a more effective data representation like CAM, it is possible for simpler, more efficient models to achieve superior performance in one-shot action recognition. CAM can enhance the precision and efficiency of one-shot action recognition in diverse fields like surveillance, robotics, sports analysis, and augmented reality.</p>