In this work we present LFHAR (Latent Features for Human Action Recognition), a novel architecture that utilizes multiple spatio-temporal latent representations to improve action feature extraction. The approach applies graph-based transformations to individual skeletal frames in temporal sequences, then arranges the derived graph features into spatio-temporal matrices. The method produces substantial performance improvements, achieving accuracy increases of 2.7% and 2.1% on the NTU-RGB+D 60 and NTU-RGB+D 120 datasets, respectively, confirming its efficacy in improving skeleton-based action recognition.