Explicit encoding of temporal dynamics for interpretable melody similarity modeling: insight into machine learning...

Temporal melody similarity is a fundamental problem in music modelling, and current methods are mostly based on recurrent or attention-based architectures that implicitly learn sequential structures.

This work adopts an explicit feature-representation approach using temporal-style descriptors, including first-order deltas, relative ratios, polynomial interactions, and rolling statistical features. These features are represented in a model-agnostic manner to enable controlled comparison, interpretability, and reproducible evaluation across architectures. Seven model families were evaluated, including linear baselines, ensemble methods, recurrent networks, attention-based models, and dense fusion architectures, using a large-scale melody similarity dataset containing over 11,000 samples. Cross-validation and bootstrap-based uncertainty estimation were applied for performance evaluation.

Results show that the engineered temporal feature space exhibits significant nonlinearity, with the highest performance achieved by Deep GRU (R ² = 0.863, MAE = 0.058), followed by XGBoost among non-recurrent models (R ² = 0.824, MAE = 0.091). Baseline models showed lower performance, demonstrating the nonlinear characteristics of the task. Feature attribution analysis identified temporal descriptors and embedding-based variables as the most influential features.

The results demonstrate that TFR provides an interpretable and reproducible foundation for evaluating future sequence-learning architectures, particularly for data-constrained or deployment-oriented applications.