Recovering Expert Critic-Sourced Network Adjacency between Musical Artists from Acoustic Distributions: A Construct-Validity Approach

This paper investigates the construct validity of critical adjacency—the pairwise relationship between musical artists established when expert critics co-mention them in long-form prose. The authors seek to determine whether these expert-defined links are grounded in the intrinsic acoustic properties of the music or if they primarily reflect sociological factors such as subcultural context, narrative positioning, and canonical status.

To test this, the authors represent artists as empirical distributions over 80 low-level acoustic descriptors (Essentia). They model the proximity between artists using marginal Wasserstein (optimal-transport) distances, which capture the full shape of an artist's sound rather than just the mean. Using a stacked ensemble learner (Super Learner), they attempt to predict the presence of critic-sourced edges in an artist-disjoint, cold-start evaluation. They further analyze how model performance changes based on the level of critical consensus (the number of sources confirming an edge) and across different genre typologies.

• Sonic Recoverability: The model recovers critic-sourced edges with an out-of-sample AUC of 0.767, demonstrating that critical discourse is substantially grounded in the music itself.
• The Sonic Core vs. Sociological Remainder: Recoverability rises monotonically with critical consensus, reaching an AUC of 0.865 for multi-source attested edges. This suggests that high-consensus links represent a reproducible sonic core, while single-source edges likely reflect a sociological remainder driven by narrative and subcultural factors.
• Genre Dynamics: The model performs better on tightly bounded, scene-based genres (e.g., Electronic, Hip-Hop/Rap) than on broad industry umbrella terms (e.g., Pop, Classic Rock), aligning with sociological models of genre evolution.