A Unsupervised Domain Adaptation Framework For Semi-Supervised Melody Extraction Using Confidence Matrix Replace and Nearest Neighbour Supervision

Melody extraction from polyphonic music is a crucial task in the field of music information retrieval (MIR), yet it faces hurdles like data annotation scarcity and domain shift. Existing unsupervised domain adaptation framework for semi-supervised melody extraction methods often adopt a "passive adaptation" paradigm, attempting to fit uncertainties in unlabeled data, which is frequently constrained by pseudo-label noise and large domain discrepancies. To overcome these limitations, we propose an innovative "active repair" paradigm, introducing a novel unsupervised domain adaptation framework for semi-supervised melody extraction that incorporates Confidence Matrix Replace (CMR) and Nearest Neighbor Supervision (NNS). Firstly, Our CMR method actively diagnoses the model’s blind spots by identifying low-confidence regions in unlabeled target data and replacing them with high-confidence counterparts from their augmented versions. Secondly, our NNS method leverages the target domain’s internal structure by employing a nearest neighbor contrastive learning strategy, which semantically aligns the feature spaces of the source and target domains. Extensive experiments on six cross-domain melody extraction tasks demonstrate the superiority of our approach, especially in tasks with larger domain shifts (classical to jazz, jazz to classical).