PREDICTIVE MODELS FOR DANCE COMPETITION JUDGING

Abstract [English]

Judging dance competitions is dependent on the subjective judgment of highly multimodal dances, which is very likely to be subjective and unequal. The paper will present a predictive modeling architecture of dance competition judging based on the multimodal performance analytics and decision support algorithms built on machine learning. The representation of dance performances with synchronized visual movement, audio rhythm, and spatial trajectory properties is supervised learning ECs, which are recorded in criterion-wise and aggregate scores of judges. Various predictive methods, such as gradient boosted decision trees, recurrent and convolutional temporal modeling, and transformer-based models are reviewed within either regression- or ranking-based formulation. The experimental findings indicate that learning based models significantly enhance the accuracy of score prediction and ranking alignment than a linear baseline with deep temporal models performing optimally. Moreover, inter-judge consistency test proves that model-assisted judging can decrease the variations of scores and increase the stability of the evaluation. The suggested framework puts AI into a status of a human-in-the-loop analytical instrument that facilitates justice, openness, and uniformity of competitive dance judging settings.