Research on robustness evaluation system of dance movement quality based on causal inference and invariant learning

Aiming at the problems of poor generalization ability and unstable evaluation of current dance movement evaluation systems based on deep learning when facing unknown dancers, clothing, and environment, this paper proposes a Causal Invariant Dance Evaluation System. The system deeply integrates causal inference and invariant learning theory, aiming to strip surface-level correlations from multi-domain data and learn stable mappings between action quality and core causal characteristics. The core work of this paper is to construct a causal feature decoupling framework that separates the input action sequence into domain-invariant core causal features and domain-specific style features, and uses the principle of invariant risk minimization to ensure that the predictor based on core features is consistent across multiple training environments. To verify the effectiveness and robustness of the Causal Invariant Dance Evaluation System, this paper conducts a comprehensive experiment on a dance dataset comprising 125 dancers across 5 styles and 5 shooting environments. The results show that the system’s evaluation accuracy reaches 92.5% on the known test set, comparable to that of mainstream deep learning methods. In the more challenging cross-domain generalization test, the system is significantly better than the baseline model. Its average absolute error is reduced by 38.7% compared with the optimal baseline model, and the correlation coefficient remains above 89.1%. Ablation experiments further confirmed the necessity of each module, and removing the invariant learning module resulted in a cross-domain performance degradation of more than 15%. This study provides an effective solution for building a highly robust intelligent dance teaching system.