A team of researchers, working with collaborator Fatma S. Abousaleh, has introduced a data-driven framework that models and predicts music popularity patterns using Spotify audio features. The study, published in the International Journal of Data Science and Analytics, addresses a gap left by earlier work that focused narrowly on track-level prediction. Instead of treating every song as an isolated case, the team examined higher-level structure in the audio-feature space and tracked how popularity varies across release cohorts, grouping tracks by the month in which they appeared.
The dataset covers tracks released between 2000 and 2022, with popularity scores retrieved once during data collection in 2023. The authors are careful to note that these values therefore represent collection-time popularity rather than popularity at the moment of release, an important limitation when interpreting the results.
As a first step, Pearson correlation analysis revealed only weak marginal linear associations between individual audio features and popularity. In other words, no single feature could meaningfully explain popularity on its own, which motivated the search for richer representations.
The framework subsequently derives latent acoustic profiles, compact groupings in the audio-feature space that capture combinations of characteristics rather than single dimensions. The team then modeled the mean collection-time popularity of each profile across chronologically ordered release-month cohorts.
One profile stood out: Profile 2, characterized by relatively high acousticness and low intensity, held the highest mean collection-time popularity in 25 of the 36 release-month cohorts from 2020 to 2022, roughly 69 percent. The researchers stress that this finding describes the more frequent dominance of Profile 2 within the retrieved sample, not a universal law about what makes music popular.
On the prediction side, the evaluation compared recurrent neural architectures, and the best-performing model varied across profiles. Long short-term memory networks, or LSTM, achieved the lowest error for Profile 1, while gated recurrent units, or GRU, performed best for Profile 2, and a vanilla recurrent neural network was strongest for Profile 3. This heterogeneity suggests that the temporal dynamics of each acoustic profile differ, and that no single recurrent architecture is uniformly superior when modeling popularity patterns across release cohorts.
The harder task proved to be identifying which profile would dominate a given cohort. In dominant-profile identification, both LSTM and GRU predicted Profile 2 for every test cohort, which meant they only matched a training-selected majority-profile baseline. The authors present this honestly: the recurrent models did not outperform a simple baseline that always picks the historically most frequent dominant profile. The contribution therefore lies primarily in the latent representation framework itself and in the descriptive finding about Profile 2’s recurring dominance, rather than in demonstrated predictive gains over naive baselines.
For the science of music analytics, the study offers a technically grounded lesson. Weak marginal correlations do not preclude useful structure at a higher level of abstraction, and latent profiles can expose patterns that single features miss. At the same time, the work illustrates the boundaries of platform-based popularity research: scores captured at a single collection time, a finite sample, and baseline-matching predictive performance all temper the conclusions. The framework provides a reusable template for analyzing release-cohort popularity, while leaving open questions about how collection-time snapshots relate to the longer trajectory of a song’s reception.
Subject of Research: Predicting music popularity from Spotify audio features using latent acoustic profiles
Article Title: A latent representation framework for modeling and predicting music popularity across release cohorts
Article References: Hidayati, S. C., Joedhiawan, M. A., Subakti, M. M. I., Amaliah, B., Ahmadiyah, A. S., & Abousaleh, F. S. (2026). A latent representation framework for modeling and predicting music popularity across release cohorts. International Journal of Data Science and Analytics, 22(1), Article 358. https://doi.org/10.1007/s41060-026-01317-0
Image Credits: AI Generated
DOI: 10.1007/s41060-026-01317-0
Keywords: latent, representation, framework, modeling, predicting, music, popularity, across, release, cohorts, scientific research
News Source: Denise Maddox. (October 25, 2026). Latent acoustic profiles reveal patterns in music popularity across release cohorts. Scienmag.
