The algorithmic nature of song-sequencing: statistical regularities in music albums

TL;DR

The study investigates whether professional music albums exhibit reliable sequencing regularities by testing three concepts—adjacent transitioning, absolute positioning, and overall trajectory—using a large Spotify-based dataset (51,010 albums, 548,852 tracks) and features Valence, Energy, Loudness, and Tempo. It models adjacent transitions with Parsons coding to create feature-specific transition matrices $Tm_j$ and evaluates albums via the mean log-likelihood $\,\mathcal{L}(a|Tm)$, while also analyzing absolute positioning with segment-based feature means and overall trajectories with Spearman correlations. The results show that original album sequences are non-random, with higher likelihoods under the learned transition patterns, beginning segments enriched in high valence/energy/loudness, and a tendency toward global down ramps, alongside an above-chance automated sequencing performance (≈17% correct adjacency vs ≈9% baseline). These findings support a partial, context-dependent influence of sequencing on listening experiences and offer practical insights for playlist generation and album curation, while acknowledging perceptual validation remains an open question. Future work could integrate deeper magnitude-aware models and perceptual experiments to bridge the global/local perceptual debate in album sequencing using richer optimization and validation methods.

Abstract

Based on a review of anecdotal beliefs, we explored patterns of track-sequencing within professional music albums. We found that songs with high levels of valence, energy and loudness are more likely to be positioned at the beginning of each album. We also found that transitions between consecutive tracks tend to alternate between increases and decreases of valence and energy. These findings were used to build a system which automates the process of album-sequencing. Our results and hypothesis have both practical and theoretical applications. Practically, sequencing regularities can be used to inform playlist generation systems. Theoretically, we show weak to moderate support for the idea that music is perceived in both global and local contexts.

Figures (6)

• Figure 1: Melody A is different from melody B, even though they are composed of the exact same notes. This illustrates how sequential factors might determine perceptually relevant musical features, such as contour and interval.
• Figure 2: Likelihood of original albums versus random sequences (i.e., $a$ versus $'a$) under the same empirically derived transition matrices (i.e., $Tm$)
• Figure 3: Feature values throughout album segments. Error bars represent Standard Error of the Mean.
• Figure 4: Proportions of down ramps for original and randomized albums.
• Figure 5: Normalised accuracy score represents the number of correct sequences divided by the length of the album. Our system presents 17% of correct sequences, whereas the random bootstrapped model presents an average of 9%.