The Wiki Music dataset: A tool for computational analysis of popular music

Fabio Celli

REVIEW 4 major objections 7 minor 23 references

The Wiki Music dataset: A tool for computational analysis of popular music

T0 review · 4 major / 7 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A stacking meta-classifier can predict the genre-scale score of future popular music with 75% average accuracy.

desk verdict A potentially useful hand-built dataset marred by a forecasting evaluation that lets predictions count against the wrong years. read the letter →

arxiv 1908.10275 v1 pith:UGVS7VJZ submitted 2019-08-27 cs.CL cs.CYcs.SDeess.AS

Fabio Celli This is my paper · ORCID

classification cs.CLcs.CYcs.SDeess.AS

keywords popularmusicgenretimeseriesforecastingdatasetWikipediamodelscalemachinelearning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks whether algorithms can find trends in the history of popular music and predict the characteristics of future genres, and answers yes to both. It contributes a hand-annotated dataset of 77 popular music genres scored across 41 features covering style, perceived sound, place, media, emotion, and the MUSIC psychological model, indexed by decade and representative year. Using time-series forecasting trained on 1900–2010, the authors report that a stacking meta-classifier predicts the genre scale for 2011–2018 with an average accuracy of 0.75, where the genre scale is a continuous score from 0 (downtempo/industrial) to 1 (techno/trance). The results also show that audiences increasingly prefer contemporary and intense music, and that the shift toward synthetic sound correlates with danceability, drug usage, and reduced band size. The dataset and method offer a quantitative, reproducible route from music history to cultural forecasting.

What carries the argument

The central object is the genre scale, a continuous 0-to-1 score defined by the authors that ranks popular music genres from downtempo/industrial (0) through metal, rock, blues, country, pop, jazz, latin, RnB, reggae, rap, DnB, electro, EDM, to techno/trance (1). This scale converts qualitative genre labels into numeric targets, making the prediction question a time-series forecasting problem. The predictive machinery is a stacking meta-classifier (SVM + MLP + IBk) with nearest-neighbor as the strongest base learner, and the evaluation uses an accuracy metric that counts a prediction as correct when it falls within 0.1 of any genre-scale label in the test years.

What would settle it

Recompute the stacking meta-classifier's average accuracy while requiring that each predicted genre-scale value be within 0.1 of the genre scale of a genre that actually emerged in the same year; if the average accuracy drops to the majority-class baseline, the forecasting claim is falsified.

Watch

Extended reading notes

Core claim

The central discovery is that the genre scale—a manually assigned 0-to-1 score ordering popular music genres from downtempo/industrial through metal, rock, jazz, rap, electro, and techno/trance—can be treated as a time series and forecasted. Two independent raters annotated 77 genres from Wikipedia descriptions (overall inter-annotator agreement α = 0.793, α = 0.957 for the genre scale), and training on 1900–2010 with linear, support-vector, neural, nearest-neighbor, and stacked classifiers, the stacking meta-classifier combining SVM, MLP, and IBk reaches an average accuracy of 0.75 for 2011–2018 under a metric that counts a hit when a prediction is within 0.1 of any genre-scale label in the window. The forecasting problem is nonlinear; IBk yields the closest predicted sequence; and reducing the features from 41 to 14 does not change performance, suggesting the signal is carried by a small set of interpretable dimensions.

Load-bearing premise

The load-bearing premise is that the year attached to a genre label is arbitrary, so a prediction counts as correct when it falls within 0.1 of any genre-scale label in the test window, even if the matching label comes from a different year.

Editorial extensions

If this is right

If the 0.75 average accuracy holds, music historians can use the same feature set to backcast and explain genre transitions over the past century.
The finding that feature reduction from 41 to 14 features does not hurt accuracy means a compact, interpretable set of cultural indicators may be sufficient for forecasting.
The confirmed shift toward synthetic, danceable, intense music gives the music industry a quantitative basis for anticipating audience tastes.
Because the problem is nonlinear, simple linear trend models are insufficient; the result argues for ensemble or instance-based methods in cultural forecasting.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension is to measure accuracy under strict temporal alignment between a prediction and the emergence year of the genre it matches; the paper's metric explicitly allows any year in the window, so this would reveal how much of the 0.75 accuracy depends on chronological precision.
The same annotation scheme could be applied to other cultural domains, such as film genres or literary trends, to test whether cultural evolution follows a similar continuous scale.
Because the genre scale is derived from Wikipedia descriptions, an acoustic validation against audio-derived features like tempo, timbre, and loudness would reveal whether the scale tracks perceptual reality or encyclopedic discourse.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The useful part of this paper is the dataset. The forecasting result, as reported, is not a forecast. The hand-crafted collection of 77 genres with 41 features, two annotators, and a Cronbach's alpha of 0.793 (0.957 for the genre scale) is a modest but real artifact. The trend analyses confirm earlier findings about optimal differentiation and urban origins, and the narrative history of popular music is a nice, if opinionated, read.

The problem is the evaluation. Section IV defines accuracy as a = sum(count(|l-h|<0.1))/count(t), with l and h allowed to be anywhere in the year series. That means a prediction for 2018 can get credit for matching a 2011 actual. The starred entries in Table I confirm the consequence: many starred predictions differ from the same-year actual by more than 0.1. The star counts do match the reported average accuracies, so the table is internally consistent with the unaligned metric, but that metric does not measure forecasting. The paper's own justification—'the year of the genre label is arbitrary, thus MAE or RMSE are not suitable'—is exactly where the predictive claim collapses. If the year is arbitrary, you are not forecasting.

Other soft spots: Table I contains out-of-range values like MLP 2013 = 1.407 for a scale defined 0-1; the trend lines have no significance tests or error bars; and the genre scale is a hand-defined numeric encoding, so the prediction target is expert opinion. The dataset's hosting on a personal URL with no code or DOI also limits reproducibility.

Who is this for? Computational musicologists who want a small, manually curated corpus of genre attributes. They should treat the scale and annotations as expert opinion, not ground truth. I would not cite it in my own work in the next year, but it deserves a serious referee: the dataset artifact is worth evaluating, and the evaluation flaw is identifiable and fixable. I would not accept the paper in its current form, but I would engage with it if the authors reframe it as a dataset description with descriptive trends, drop the forecasting claim, and properly host the data. That would be a legitimate, modest contribution.

Referee Report

4 major / 7 minor

Summary. This paper introduces the Wiki Music dataset, a hand-crafted collection of 77 popular music genres annotated with 41+ features covering genre scale, acoustic, place, media, emotion, style, and MUSIC-model dimensions, indexed by decade and representative year. Using this dataset, the authors analyze trends in music features over the 20th century and perform time series forecasting of a one-dimensional 'genre scale' for the years 2011-2018, training on data from 1900-2010. They report that a stacking meta-classifier achieves an average accuracy of 0.75, and conclude that the forecasting of music genres is a non-linear problem.

Significance. The dataset and its inter-annotator reliability statistics (Cronbach's alpha = 0.793 overall, 0.957 for the genre scale) represent a potentially useful resource for computational music analysis. The paper also makes falsifiable quantitative claims about trends in popular music (e.g., increasing synthetic sound, decreasing mellow/unpretentious content). However, the forecasting evaluation is not statistically sound: the accuracy metric intentionally permits predictions and labels from different years to be compared, which removes the temporal alignment that defines forecasting. As a result, the central predictive claim is untested, and the paper's main quantitative conclusion is not supported.

major comments (4)

[Section IV] The evaluation metric defined as a = sum(count(|l-h|<0.1))/count(t) explicitly states that 'the label (l) and the prediction (h) can be anywhere within the year serie (t).' This means a prediction for 2018 can be credited for matching an actual label from 2011, and vice versa. Under this metric, even a constant predictor emitting a value near 0.8 could match several labels in the 2011-2018 sequence (which includes 0.76, 0.82, 0.9, 0.91, 0.95). The reported 0.75 accuracy for the meta-classifier is therefore not a measure of forecast skill. The authors must evaluate with year-aligned metrics (e.g., MAE, RMSE, or proper forecast skill scores) and justify why temporal alignment is 'arbitrary'; otherwise the claim that 'a meta classifier ... is the most accurate in the prediction' is unsupported.
[Table I] The starred entries in Table I do not systematically correspond to within-tolerance predictions for the correct year. For example, for linear regression the starred predictions at 2015 (0.747) and 2016 (0.739) differ from the actual genre scale for those years (0.91 and 0.18) by far more than 0.1; they happen to be within 0.1 of the 2017 and 2018 labels (0.76 and 0.82). In addition, the MLP prediction of 1.407 for 2013 lies outside the [0,1] genre scale, indicating an invalid model output. The way stars are assigned and whether multiple matches are counted is not specified precisely enough to reproduce the reported accuracies (averages of 0.25, 0.375, 0.5, 0.5, 0.75).
[Section IV, correlation analysis] The paper claims that 'the change towards sampled and synthetic sound is correlated to dancing, to intensity/aggressiveness, to a larger drug usage and to a large variety of influences' without reporting correlation coefficients, significance tests, or confidence intervals. Similarly, the MUSIC feature trend lines in Figure 3 are presented as 'revealing' trends without statistical support. With only 77 genres, these claims need at least a permutation test or bootstrapped interval to rule out small-sample artifacts.
[Section III, genre scale] The 'genre scale' is a hand-defined numeric encoding created by the authors (e.g., 0.25=pop rock, 0.5=pop, 0.9=electro/house). Forecasting this scale is only meaningful if the scale is a validated operationalization of a musical concept; the paper provides inter-annotator reliability but no external validity evidence. The forecasting experiment therefore measures the predictability of an author-defined label rather than an objective property of music. This should be acknowledged and the scale ought to be validated or the claims carefully conditioned.

minor comments (7)

[Abstract] The sentence 'Is it possible use algorithms...' should be 'Is it possible to use algorithms...'.
[Section III] The text contains typos such as 'mplying' for 'implying' and 'inﬂeunces' for 'influences'.
[Section IV] The model name is written as 'MPL' in the text and 'MLP' in Table I; please make this consistent.
[Section IV] 'RSME' should be 'RMSE' (root mean squared error).
[Section V] The section heading 'CONCLUSION ACKNOWLEDGMENTS AND FUTURE' appears malformed and should be revised.
[Section III, dataset availability] The dataset is hosted on a personal webpage; a persistent DOI or a standard data repository would improve reproducibility and preservation.
[Section III] The paper does not provide a comprehensive table of the exact features and their value ranges, nor the number of genres per decade; a data descriptor would be helpful.

Circularity Check

1 steps flagged · score 6.0 of 10

The claimed 0.75 forecast accuracy is an artifact of an evaluation metric that explicitly allows predicted and actual genre-scale labels to be matched across different years, so the central prediction claim is not tested as a forecast.

self definitional [Section IV ("Experiments"), evaluation metric definition and Table I]
"As the year of the genre label is arbitrary, predicted scores and labels can be not aligned, thus MAE or RSME are not suitable evaluation metrics. As evaluation metric we defined average accuracy as a = ∑ count(|l−h|<0.1) / count(t), where the label (l) and the prediction (h) can be anywhere within the year serie (t)."

The forecast claim is defined into the metric rather than tested. By allowing l and h to be taken from any year in 2011-2018, the accuracy count does not require predicting the right year: a score predicted for one year can receive credit by being within 0.1 of a hand-annotated genre-scale value from a different year. The targets in Table I (0.18, 0.25, 0.5, 0.76, 0.82, 0.9, 0.91, 0.95) form a dense set, so a model that emits a value near 0.8 can match several labels across different years; the starred entries are selected by cross-year proximity, not by temporal correspondence.

full rationale

The paper's own equations exhibit the reduction: the accuracy formula in Section IV removes the time index from the comparison, and the text explicitly says labels and predictions 'can be anywhere within the year serie.' That makes the central 'prediction' claim self-referential: the evaluation is defined so that predicting a value close to any future label, regardless of year, counts as success. Since the genre-scale labels are hand-defined anchors (Section III), the reported accuracy mainly reflects proximity to that hand-coded set, not skill at forecasting future music characteristics. Cronbach's alpha for annotation agreement (0.793 average, 0.957 for genre scale) is independent support for reliability of the annotations, but it does not validate the forecasting evaluation. No self-citation chain or imported uniqueness argument is present; the circularity is confined to the evaluation metric, yet it undermines the paper's headline predictive result. Hence a score of 6: one or more 'predictions' reduce by construction, while the dataset and descriptive trend analyses retain independent content.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central claim rests on hand-assigned genre scale values, a permissive accuracy metric, and subjective annotations. No external data or independent measurements support the numerical encoding of genre categories.

free parameters (4)

Genre scale values for super-genres = 0, 0.1, 0.15, 0.2, 0.25, 0.3, 0.4, 0.5, 0.55, 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95, 1
Hand-assigned in Section III; no empirical basis for the spacing between super-genre categories.
Accuracy tolerance epsilon = 0.1
Used in the accuracy metric a = sum(count(|l-h|<0.1))/count(t); the tolerance is arbitrary and not justified.
Representative year per genre = Varies by genre
Annotators selected a single 'representative year' per genre; the paper acknowledges the year is arbitrary (Section IV).
Feature subset for reduction = 14 features from 41
Feature reduction to 14 is mentioned but the selected features are not specified.

assumptions (4)

domain assumption A single continuous genre scale from 'experimental and introverted' to 'euphoric and inclusive' captures the relations between all popular music genres.
Section III defines this scale; no empirical or theoretical justification for unidimensionality.
domain assumption Wikipedia descriptions plus listening to samples are sufficient to rate genres on 41 subjective features.
Section III describes the annotation procedure; this assumes the raters' judgments are valid measurements of genre characteristics.
standard math Cronbach's alpha is an appropriate measure of inter-rater agreement for these ratings.
Section III uses Cronbach's alpha (reference [23]); standard usage, though the assumption that the ratings are interval-scaled is unstated.
ad hoc to paper The accuracy metric that allows predicted and actual values from different years is a valid evaluation of forecasting.
Section IV: 'predicted scores and labels can be not aligned'; this is an ad hoc choice to avoid MAE/RMSE, but it invalidates temporal forecasting.

invented entities (1)

Genre scale
purpose: Converts music super-genres to numeric values for time series forecasting.
The scale is a hand-authored numeric encoding (Section III) with no external validation; the numerical spacing is arbitrary.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Wiki Music dataset: A tool for computational analysis of popular music." pith.science (2026). https://pith.science/paper/UGVS7VJZ

@misc{pith2026190810275,
  author       = {Pith},
  title        = {Pith review of: The Wiki Music dataset: A tool for computational analysis of popular music},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UGVS7VJZ}},
  note         = {Machine review of arXiv:1908.10275}
}

read the original abstract

Is it possible use algorithms to find trends in the history of popular music? And is it possible to predict the characteristics of future music genres? In order to answer these questions, we produced a hand-crafted dataset with the intent to put together features about style, psychology, sociology and typology, annotated by music genre and indexed by time and decade. We collected a list of popular genres by decade from Wikipedia and scored music genres based on Wikipedia descriptions. Using statistical and machine learning techniques, we find trends in the musical preferences and use time series forecasting to evaluate the prediction of future music genres.

Figures

Figures reproduced from arXiv: 1908.10275 by the authors.

**Figure 2.** Distributions of some of the features annotated in the dataset. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Trend lines (dashed) of the MUSIC features from 1900. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 22 canonical work pages

[1]

Instance-based learning algorithms

David W Aha, Dennis Kibler, and Marc K Albert. Instance-based learning algorithms. Machine learning, 6(1):37–66, 1991

work page 1991
[2]

Kraftwerk: Music non-stop

Sean Albiez and David Pattie. Kraftwerk: Music non-stop. Bloomsbury Publishing USA, 2010

work page 2010
[3]

What makes popular culture pop- ular? product features and optimal differentiation in music

Noah Askin and Michael Mauskapf. What makes popular culture pop- ular? product features and optimal differentiation in music. American Sociological Review, 82(5):910–944, 2017

work page 2017
[4]

Deep learning techniques for music generation-a survey

Jean-Pierre Briot, Ga ¨etan Hadjeres, and Franc ¸ois Pachet. Deep learning techniques for music generation-a survey. arXiv preprint arXiv:1709.01620, 2017

arXiv 2017
[5]

Music (ology) needs a context-re-interpreting goa trance

Sebastian Chan. Music (ology) needs a context-re-interpreting goa trance. Perfect Beat, 3(4):93–98, 2015

work page 2015
[6]

What’s that sound?: an introduction to rock and its history

John Rudolph Covach and Andrew Flory. What’s that sound?: an introduction to rock and its history . WW Norton & Company New York., 2006

work page 2006
[7]

People Get Ready!: A New History of Black Gospel Music

Robert Darden and Bob Darden. People Get Ready!: A New History of Black Gospel Music . Bloomsbury Publishing, 2005

work page 2005
[8]

Electronic dance music in the dubstep era

Mike D’Errico. Electronic dance music in the dubstep era. Oxford Handbooks, 2015

work page 2015

Show all 23 references

[9]

Lyrics-based analysis and clas- siﬁcation of music

Michael Fell and Caroline Sporleder. Lyrics-based analysis and clas- siﬁcation of music. In Proceedings of COLING 2014, the 25th Inter- national Conference on Computational Linguistics: Technical Papers , pages 620–631, 2014

2014
[10]

I Hear a Symphony: Motown and Crossover R&B

Andrew Flory. I Hear a Symphony: Motown and Crossover R&B . University of Michigan Press, 2017

2017
[11]

The song is you: Preferences for musical attribute dimensions reﬂect personality

David M Greenberg, Michal Kosinski, David J Stillwell, Brian L Monteiro, Daniel J Levitin, and Peter J Rentfrow. The song is you: Preferences for musical attribute dimensions reﬂect personality. Social Psychological and Personality Science , 7(6):597–605, 2016

2016
[12]

K-pop genres: A cross-cultural exploration

Jin Ha Lee, Kahyun Choi, Xiao Hu, and JH Downie. K-pop genres: A cross-cultural exploration. In Proceedings of the 14th Conference of the International Society for Music Information Retrieval (ISMIR) . The International Society for Music Information Retrieval (ISMIR)., 2013

2013
[13]

Classiﬁcation as culture: Types and trajectories of music genres

Jennifer C Lena and Richard A Peterson. Classiﬁcation as culture: Types and trajectories of music genres. American sociological review , 73(5):697–718, 2008

2008
[14]

The evolution of popular music: Usa 1960–2010

Matthias Mauch, Robert M MacCallum, Mark Levy, and Armand M Leroi. The evolution of popular music: Usa 1960–2010. Royal Society open science, 2(5):150081, 2015

1960
[15]

Embed- dedness and the production of novelty in music: A multi-dimensional perspective

Michael Mauskapf, Eric Quintane, Noah Askin, and Joeri Mol. Embed- dedness and the production of novelty in music: A multi-dimensional perspective. In Academy of Management Proceedings . Academy of Management Briarcliff Manor, NY 10510, 2017

2017
[16]

Signiﬁcance and popularity in music production

Bernardo Monechi, Pietro Gravino, Vito DP Servedio, Francesca Tria, and Vittorio Loreto. Signiﬁcance and popularity in music production. Royal Society open science , 4(7):170433, 2017

2017
[17]

Music genres and corporate cultures

Keith Negus. Music genres and corporate cultures . Routledge, 2013

2013
[18]

Systematic exploration of computa- tional music structure research

Oriol Nieto and Juan Pablo Bello. Systematic exploration of computa- tional music structure research. In ISMIR, pages 547–553, 2016

2016
[19]

New Orleans style and the writing of American jazz history

Bruce Boyd Raeburn. New Orleans style and the writing of American jazz history. University of Michigan Press Ann Arbor, 2009

2009
[20]

The song remains the same: A replication and extension of the music model

Peter J Rentfrow, Lewis R Goldberg, David J Stillwell, Michal Kosinski, Samuel D Gosling, and Daniel J Levitin. The song remains the same: A replication and extension of the music model. Music Perception: An Interdisciplinary Journal, 30(2):161–185, 2012

2012
[21]

Popular music: The key concepts

Roy Shuker. Popular music: The key concepts . Routledge, 2017

2017
[22]

Strange sounds: Music, technology and culture

Timothy D Taylor. Strange sounds: Music, technology and culture . Routledge, 2014

2014
[23]

Cronbach alpha (v1

P Wessa. Cronbach alpha (v1. 0.5) in free statistics software (v1. 2.1). Ofﬁce for Research Development and Education , 2017. 2https://musicmap.info/

2017

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Instance-based learning algorithms

David W Aha, Dennis Kibler, and Marc K Albert. Instance-based learning algorithms. Machine learning, 6(1):37–66, 1991

work page 1991

[2] [2]

Kraftwerk: Music non-stop

Sean Albiez and David Pattie. Kraftwerk: Music non-stop. Bloomsbury Publishing USA, 2010

work page 2010

[3] [3]

What makes popular culture pop- ular? product features and optimal differentiation in music

Noah Askin and Michael Mauskapf. What makes popular culture pop- ular? product features and optimal differentiation in music. American Sociological Review, 82(5):910–944, 2017

work page 2017

[4] [4]

Deep learning techniques for music generation-a survey

Jean-Pierre Briot, Ga ¨etan Hadjeres, and Franc ¸ois Pachet. Deep learning techniques for music generation-a survey. arXiv preprint arXiv:1709.01620, 2017

arXiv 2017

[5] [5]

Music (ology) needs a context-re-interpreting goa trance

Sebastian Chan. Music (ology) needs a context-re-interpreting goa trance. Perfect Beat, 3(4):93–98, 2015

work page 2015

[6] [6]

What’s that sound?: an introduction to rock and its history

John Rudolph Covach and Andrew Flory. What’s that sound?: an introduction to rock and its history . WW Norton & Company New York., 2006

work page 2006

[7] [7]

People Get Ready!: A New History of Black Gospel Music

Robert Darden and Bob Darden. People Get Ready!: A New History of Black Gospel Music . Bloomsbury Publishing, 2005

work page 2005

[8] [8]

Electronic dance music in the dubstep era

Mike D’Errico. Electronic dance music in the dubstep era. Oxford Handbooks, 2015

work page 2015