<p>Automated identification of yoga postures is gaining high demand since it is a critical tool in fitness training, rehabilitation, and real-time feedback systems. Existing approaches to human posture recognition using skeletal data typically treat all joints equally, without accounting for each joint’s reliability. However, in real-world scenarios, many joints may be partially or entirely occluded, leading to unreliable inputs for pattern recognition. This work proposes a lightweight and reliable model for yoga posture recognition, VisYogaNet, that uses a visibility-aware attention mechanism to attend to reliable joints and relevant geometric features selectively. Thereby improving pose recognition performance under challenging visibility conditions while reducing complexity. Experimental results demonstrate that the proposed model excels over machine learning techniques, including random forest, convolutional neural networks (CNNs), and vision transformer. In addition, the proposed model excels advanced CNN architectures, including MobileNet V2, MobileNet V3, Inception V3, ShuffleNet, and SqueezeNet, regarding accuracy and computational efficiency. Also, VisYogaNet outperformed state-of-the-art models, achieving a test accuracy of 98.24%. In the cross-validation analysis, the VisYogaNet model demonstrated stable and high performance across all folds, with an average accuracy of 0.9947 ± 0.0095, indicating reliable learning behavior with minimal deviation.</p>