A method for mining dance movement relationships and generating graph structures based on temporal and spatial feature encoding

Abstract

To address the computational issues of multi-source data of dance movements in heterogeneous alignment, spatio-temporal dependency modeling, relationship discrimination, and graph structure generation, a method for action relationship mining and graph structure generation based on spatio-temporal feature encoding is proposed. The method synchronously aligns, corrects anomalies, and standardizes the input of skeleton coordinates, joint angles, speed, and rhythm, and constructs a unified spatio-temporal feature representation; subsequently, it uses the spatio-temporal encoding network to extract action embedding features, introduces similarity constraints and relationship classification mechanisms to calculate the association strength between action nodes, and completes graph structure generation through edge weight fusion, redundant edge pruning, and consistency optimization. Experimental results show that the relationship recognition accuracy of the proposed method on the test set reaches 91.3%, the F1 value is 0.895, the graph integrity reaches 0.92, and the average inference delay is 118 ms, demonstrating good relationship modeling ability and engineering implementation efficiency.