Resolve whether artist-centroid subtraction cleanly separates identity from musical feel

Determine whether the failure of artist-centroid subtraction to separate tracks by highly cohesive artists reflects a genuine property of those artists’ catalogs or a limitation of the linear centroid-based separation method.

Background

The residual method assumes that an artist’s mean track embedding adequately represents artist identity and that subtracting this mean isolates track-specific experiential information. Several artists retained predominantly same-artist nearest neighbors after residualization. The paper explicitly states that the current analysis cannot distinguish whether this pattern is substantive or methodological.

References

Artists with unusually cohesive catalogs (Frank Ocean, Taylor Swift, Kendrick Lamar in the seed-track results) did not separate cleanly under this method. That may reflect a genuine property of those artists' output, or a limitation of centroid subtraction as a separation technique: the current analysis does not distinguish between the two.

— Project Qualia: Recovering Experiential Music Structure from Session Co-occurrence Data  (2609.10862 - Mohammed et al., 9 Sep 2026) in Section 5.1, “What This Result Does and Doesn’t Establish”