Style-Based Composer Identification and Attribution of Symbolic Music Scores: A Systematic Survey

1 Introduction

The ILLIAC system by Hiller and Isaacson (1955–1956) was one of the first applications of information technology in music (Hiller and Isaacson, 1958). Around the same time, Youngblood published the first statistical analysis of compositional styles (Youngblood, 1958). The field of computational stylistic analysis has since evolved, encompassing a diverse range of literature and overlapping with tasks like genre recognition and algorithmic composition. Youngblood's work highlights the longstanding significance of this topic in music computing, with important applications in musicology and database labeling (van Nuss et al., 2017).

In this work, computational methods applied to symbolic scores are considered in the attempt to systematically analyze the literature about composer identification.

The task of identifying the composer of a musical composition based on its symbolic score has been explored from various research perspectives. Computer scientists and statisticians refer to it as ‘composer identification’ or ‘composer classification,’ while musicologists and historians call it ‘authorship attribution.’ Despite differing aims, the typical scenario remains the same: given a set of music scores, develop an algorithm to predict the composer of a new score.

The synthesis of the surveyed literature reveals several distinct research perspectives from which the composer‑identification task is approached. One reason why scientists approach the composer‑identification task is to assess the ability of a model to capture the stylistic characteristics of a composer. In this case, composer identification is often associated with genre recognition and style or epoch classification. Another application is the labeling of large databases, often mined from the web, that present a large number of music documents but with noisy or erroneous metadata. In this latter case, the task of composer identification is associated with generic document tagging. In other works, the task is presented as a benchmark for comparing different feature extractors or architectures. When one of these scenarios occurs, the main focus is usually on the computational methods, with the final task hidden behind generic nomenclature such as ‘music tagging’ or ‘music classification.’

The aim of authorship attribution is instead to infer the author of a piece of music, given a set of possible authors. This task is often associated with the re‑attribution of works whose authorship is uncertain or disputed. A more in‑depth discussion of what authorship attribution is presented in Section 2. From a machine learning perspective, the authorship attribution problem may be regarded as a particular case of the composer‑identification task.

In this work, the existing literature about composer identification from symbolic music scores is critically reviewed, providing details of the methodology and criteria adopted during the analyses. Specifically, three main research questions motivated the study:

1. What computational approaches have been developed and applied to composer identification of symbolic music scores?

2. To what extent can music composer identification methods based on compositional style be considered reliable for authorship attribution?

3. How can the research community improve the reliability and reproducibility of music authorship attribution studies?

Many studies treat ‘composer identification’ as a machine‑learning benchmark—exploiting gross stylistic differences (e.g., Baroque vs. Romantic) to compare feature extractors or architectures—without any claim to musicological authorship attribution. By contrast, the primary focus is on those works whose aim is to support musicological re‑attribution of disputed or anonymous pieces, relying on subtle, individual stylistic traits rather than broad period or genre cues. Nonetheless, the two tasks are closely related, and the distinction is not always clear‑cut. While I have attempted to navigate this distinction throughout the analysis, it is often challenging to definitively categorize studies; many works, even when framing composer identification primarily as a benchmarking task, present findings or interpretations that could be perceived as having musicological relevance or broader implications for understanding style, making a rigid separation problematic. Moreover, ‘composer identification’ models may still be used for ‘authorship attribution,’ and their study is valid for having a better understanding of the literature and in trying to answer the research questions. My critique throughout this survey, therefore, evaluates the reliability and applicability of methods primarily through this lens of musicological authorship attribution.

While a systematic approach that enforces an objective view on the existing literature was adopted, the heterogeneity of the applications and of the nomenclature used in the literature made the effort of collecting existing works particularly difficult. Nonetheless, the first comprehensive analysis of the field is provided, achieving a greater understanding of the possibilities that computational techniques open for the re‑attribution of existing works.

In Section 2, the concept of authorship attribution and its similarities and differences with the more general task of composer identification are discussed. In Section 3, the systematic collection of papers is described. In Section 4, the evaluation methodologies adopted in the literature are reviewed, and a classification among reliable and unreliable protocols is defined. In Sections 5, 6, and 7, the repertoires, models, features, and file formats used in the literature are summarized. In Section 8, the existing problems of authorship attribution approached with computational style analysis are analyzed. Finally, in Section 9, the literature is numerically analyzed, and guidelines are drawn for future research in the composer‑identification task.

2 Defining Authorship Attribution in MIR

According to Juola (2007), authorship attribution can be defined as ‘any attempt to infer the characteristics of the creator of a piece of [linguistic] data.’ In this context, ‘characteristics of the creator’ refer to the stylistic and, in the original linguistic domain, textual patterns manifest in their work, which can be used for tasks such as identifying the author (attribution) or inferring other authorial traits (profiling). The problem is usually presented in one of three possible forms:

1. In the closed‑set problem, the author is known to be one of a finite set of possible authors.

2. In the open‑set problem, the author may not be one of the authors in the training set.

3. In the profiling problem, the author is unknown, and the task is to infer the author's characteristics.

In the case of music composer identification, the literature does not report forms of types 2 and 3, so the present work is mainly focused on problems of type 1.

In practice, in an authorship‑attribution task, a sample of data with uncertain authorship is available. To infer the true attribution based on statistical analysis, researchers collect a dataset to train a model and apply it to , obtaining predictions . A key issue with this approach is determining how much trust can be placed in the predictions .

Therefore, it is important to obtain a measure of merit for , which serves as a measure of reliability for . This remains an unsolved problem because the inherent complexity and high dimensionality of artistic style mean that the dataset used to train , even if generally representative of the authors, frequently does not provide dense coverage for the specific region of the input space defined by a particular .

A key challenge in applying composer‑identification techniques to musicological authorship attribution is distinguishing between features that genuinely reflect a composer's individual style and those that are merely indicative of broader categories such as historical period, genre, or nationality. While models trained to differentiate, for example, Baroque from Romantic composers might perform well for general classification or benchmarking purposes by leveraging these broad characteristics, their utility for fine‑grained authorship attribution—such as distinguishing between two contemporaneous composers of the same school—is limited if they rely primarily on such extramusical factors rather than nuanced, individual stylistic traits. This survey critically examines the literature with this distinction in mind.

In the context of a typical machine learning experiment, the authorship‑attribution task has three key peculiarities:

1. Researchers need to create a dataset that is as similar as possible to the questionable data and should avoid introducing confounding variables or editorial choices that obscure the composer's intrinsic style. In the case of music, this means:

   • defining the possible authors to be included in ;

   • collecting works and transcriptions that are musicologically valid and include minimal musicological interpretations or editorial choices;

   • encoding the music in a digital symbolic format that does not introduce confounding variables irrelevant to compositional style;

   • ensuring that the collected data encompass all the composers' stylistic characteristics; and

   • taking care of the quantitative aspects of the dataset, making it balanced with respect to styles and composers.

2. The final inference on is meaningless without a measure of merit computed on a sufficiently large evaluation dataset that differs from the one used to train the model .

3. In the open‑set problem, the evaluation of the model is particularly difficult, since the content of a class ‘anything else’ is extremely difficult to define.

Despite these peculiarities, the authorship‑attribution task can be considered a special case of the composer‑identification task.

While composer identification and authorship attribution can be applied to various media, this survey focuses on methods applied to symbolic music scores when the goal is to attribute authorship based on intrinsic compositional style. Consequently, I exclude studies that primarily rely on MIDI performances, audio recordings, or direct images of scores as their input. This exclusion is because these media types can introduce significant confounding variables that are distinct from the composer's original stylistic intent and can obscure the features relevant for musicological attribution:

• MIDI performances are shaped by a performer's interpretation (e.g., dynamics, articulation, timing), which introduces variables related to performance choices rather than purely compositional ones. These interpretations might themselves be influenced by the performer's idea of the composer or prevailing cultural practices, further complicating the signal.

• Audio recordings are subject to performer interpretation, recording technology, room acoustics, and audio engineering choices, all of which can obscure or alter the purely compositional features.

• Images of music scores can introduce additional variables such as the quality of printing, specific editorial choices (especially in non‑urtext editions), and the graphical style of particular editions, which are not direct reflections of the composer's intrinsic style.

Such confounding variables are not merely random noise but can introduce systematic patterns (e.g., a performer's consistent interpretation shaped by their understanding of the composer, or artifacts from a specific recording process or edition). If a model learns from these patterns, it might achieve high classification accuracy, but it becomes ambiguous whether the model is identifying the composer's intrinsic style or these external factors. This ambiguity undermines the goal of genuine authorship attribution based on compositional style, as the classification task itself may no longer reliably target the composer's unique stylistic traits. For this reason, the survey focuses on symbolic representations that aim to minimize these external influences.

For genres like popular music, ‘symbolic representation’ for attribution purposes can encompass published scores, lead sheets, or other forms of notation that capture core compositional elements (melody, harmony, rhythm), provided the analysis focuses on these authorial aspects rather than on performance or audio production characteristics. The key criterion for inclusion in this survey is that the input data aim to represent the composer's notated or structurally intended musical information, minimizing confounding variables from performance interpretation or audio production. Studies like Glickman et al. (2019), which base their work on symbolic transcriptions of audio recordings to analyze compositional style, therefore fit within this scope.

For broader context and to delineate the survey's scope from its primary focus on composer attribution, I note that a related area involves works adopting computer vision techniques for attributing the scribe of a manuscript. While these works do not take into account the stylistic properties of the music, they are obviously of interest in the field of authorship attribution. Consequently, an additional search of documents was conducted and found 12 documents falling into this category. Of these, four make use of the dataset CVC‑MUSCIMA (Fornés et al., 2012), which contains images of handwritten music collected in modern times, thus not focusing on historical composer identification problems. Another four studies are authored by Fornés (Fornés et al., 2008, 2009, 2010; Fornés and Lladós, 2010) and the remaining four articles (Bruder et al., 2004; Göcke, 2003; Niitsuma et al., 2013, 2016) focus on historical manuscripts and set the purpose of their efforts in the automatic organization of digital archives. I found no real‑world attribution problem that was tackled with computer vision techniques. The reason probably lies in the fact that manuscripts of the real composers are rarely available and, when they are, other, more reliable techniques can be used for assessing the provenance of the manuscript.

3 Collecting Literature

3.1 Search methodology

The topic under investigation spans multiple decades, during which the lexicon of computer science and statistics has evolved. Consequently, different queries were employed for distinct time periods. To account for this variation, time blocks were selected, and academic search engines were used to locate scientific papers in the field of composer identification from symbolic scores. In particular, the growing availability of digital academic documents in recent years necessitated the use of shorter time blocks for more recent periods. The selected time blocks are detailed in Table 1.

Table 1

Queries used for each time block.

	Time blocks		Search Key
	Years	Duration	Token 1
	0–1950	‑	A AND B
	1951–1970	20	A AND B
	1971–1990	20	A AND B
			A AND C
	1991–2000	10	A AND C
	2001–2010	10	A AND C AND D ‑optical
	2011–2015	5	A AND C AND D ‑optical
	2016–2018	3	A AND C AND D ‑optical
	2019–2020	2	A AND C AND D ‑optical
	2021–2023	2	A AND C AND D ‑optical
	2024–	1	A AND C AND D ‑optical
	Legend
A	music AND composer AND style
B	computer OR information OR statistics OR algorithm
C	classification OR identification OR recognition OR attribution
D	‘symbolic level’ OR ‘music score’ OR midi OR musicxml OR kern

A set of search queries was formulated using terminology adapted to different historical periods. These queries are listed in Table 1. The terms ‘music,’ ‘composer,’ and ‘style’ were retained consistently across all queries due to their specificity to the task. However, during the era of Youngblood's contributions, terms such as ‘classification’ or ‘recognition’ were not associated with statistical learning methods. Consequently, more generic terms relevant to automation processes, such as ‘information,’ ‘statistics,’ ‘computer,’ and ‘algorithm’ were used as substitutes.

Initially, computational limitations precluded large‑scale digital analysis of sound signals. By the 2000s, however, the field of Music Information Processing (MIP) began to explore the audio domain extensively. Presently, the majority of literature in MIP focuses on audio signal processing methods. To refine the search results to studies using the symbolic level, specific terms such as ‘symbolic level,’ ‘music score,’ ‘MIDI,’ ‘kern,’ and ‘musicXML’ were included. As discussed in Section 2, this focus is essential for investigating authorship attribution in music scores, which cannot be reliably applied to performance data or audio recordings due to potential confounding variables introduced by a performer's stylistic interpretation, which are distinct from the composer's intrinsic style. Accordingly, studies employing audio recordings or MIDI performance data were excluded. A further refinement involved excluding works on optical music recognition.

The search engine used was Google Scholar, which retrieves results from major academic publishers and databases, including those indexed in Scopus and the Web of Science, while extending the corpus to potentially include other documents. This engine's capability to search within the full texts of available works, rather than being limited to reviewing only metadata such as titles and abstracts, was critical for identifying publications that, while not primarily focused on composer identification, still contained relevant information and experiments. Despite this advantage, the results often included noisy entries, and there was no assurance that all retrieved documents had undergone peer review. Only papers published in academic journals or presented at scientific conferences with a peer‑review process involving at least two field experts were included. This criterion led to the exclusion of Bachelor's, Master's, and Ph.D. theses, as well as recent preprint papers without peer review.

To further narrow the analysis, works that did not implement a full task of classifying musical compositions by composer were excluded. Specifically, papers providing only generic analysis without demonstrating the effectiveness of methodologies through classification tests with numerical results were not considered. However, studies using both composers and genres in the classification labels were included, as such investigations are potentially relevant for characterizing composers' styles.

The first 10 pages of results from each of the 10 queries listed in Table 1 were analyzed, resulting in the consideration of 1,000 papers. From these, 53 papers meeting the specified criteria—peer review, full classification tasks, and use of symbolic scores—were selected. During the peer‑review process, another five papers were added, primarily published during the review time‑frame. The Preferred Reporting Items for Systematic Reviews and Meta‑Analyses flowchart summarizing the literature review process is presented in Figure 1.

Figure 1

3.2 Search results

The statistical characterization of composers' styles has been studied since at least 1958 (Youngblood, 1958). However, the first work addressing the problem through the implementation of computational tools was published in 1993 at the Italian Colloquio di Informatica Musicale (Johnson, 1993). In the intervening period, particularly in the 1960s, several studies followed Youngblood's methodology by applying information theory to music analysis, but these did not compute proper classification performance (Fucks, 1962a,b; Gabura and Oliver, 1965; Knopoff and Hutchinson, 1983; Mendel, 1969; Siromoney and Rajagopalan, 1964) and generally did not utilize computational tools, with the exception of Mendel (1969).

Two additional studies were published in the 1990s (Hörnel and Menzel, 1998; Westhead and Smaill, 1994) prior to the work that is often cited as the earliest significant publication in the field (Pollastri and Simoncelli, 2001). Subsequently, 12 papers from the decade 2000–2009, 32 papers from the decade 2010–2019, and 11 papers since 2020 (excluding the present work) were identified. The distribution of these publications across the years is depicted in Figure 2.

Figure 2

It should be noted that, since 2020, some studies have addressed composer identification using MIDI performance data (Chou et al., 2021; Foscarin et al., 2022; Kong et al., 2020; Lee et al., 2020; Yang and Tsai, 2021). These works were intentionally excluded to avoid confounding variables related to performer interpretation, as detailed in Section 2.

In total, 58 works meeting the criteria outlined in Section 3.1 were collected. From each paper, all experiments involving composer‑identification tasks were reviewed, and the best‑performing models for each dataset used in each paper were recorded. This process resulted in the identification of 87 models.

4 Evaluation Methodologies

Given the critical importance of reliable evaluation protocols in authorship attribution (Rudman, 2012), it is essential to employ robust metrics that accurately reflect model performance. A model with unsatisfactory performance metrics is unsuitable for authorship attribution because there its predictions are not reliable. However, this principle is often overlooked in the literature, with some authors attributing works based on model parameters despite low classification effectiveness—see Section 8.

Additionally, the reliability of the measure of merit strongly depends on the evaluation protocol, as Sturm (2014) highlights: a well‑designed protocol free from confounding variables is essential, and reproducibility requires selecting experimental components based on the scientific question rather than convenience. In any case, the evaluation measure is fundamental: if it is flawed, any conclusion drawn from it is questionable. To gain a comprehensive understanding of the literature, I categorized works based on their evaluation measures and analyzed other variables of the datasets for each category.

While this methodology may seem overly critical, it is essential to obtain a deeper understanding and define strong guidelines for future works regarding experimental design—see Section 9. It is also important to consider that the evaluation strategies in the surveyed literature may reflect the original aims of the studies. For instance, a paper focused on benchmarking a new algorithm might employ a dataset with clearly distinct composer styles and report ‑measure, which, while potentially valid for its specific benchmarking goal, would be insufficient for assessing reliability in a nuanced musicological authorship attribution context. My critique primarily assesses evaluation choices from the perspective of their suitability for the latter.

While accuracy is a straightforward measure, it is often inadequate for imbalanced datasets, which are common in composer‑identification tasks. Despite its widespread use in the literature, accuracy can be misleading, as it does not account for the distribution of classes. The vast majority of the existing works in the analyzed literature adopt the standard accuracy, often as the only metric.

Some studies have used ROC curves, which plot the true‑positive rate against the false‑positive rate (FPR) across different threshold settings. ROC curves provide qualitative insights into model performance, particularly in binary classification tasks. However, their applicability is limited when dealing with multiclass problems or when true labels for are unavailable. The area under the ROC curve (AUC) shares these limitations, making it less suitable for comprehensive evaluation. Four works (Glickman et al., 2019; Herremans et al., 2015, 2016; Tan et al., 2019) used AUCs.

The ‑measure, originating from information retrieval, combines precision and recall into a single metric. While useful for highly imbalanced datasets, the ‑measure can be problematic when the negative class is the minority, potentially leading to inflated scores that do not accurately reflect model performance. Additionally, its non‑linear nature with respect to True Positive Rate (TPR) and True Negative Rate (TNR) can complicate interpretation (Christen et al., 2023). In the literature, six works (Hedges et al., 2014; McKay et al., 2018; Tan et al., 2019; Karystinaios et al., 2020; Zhang et al., 2023; Mirza et al., 2024) adopted the ‑measure in their analysis, while one work (van Nuss et al., 2017) used the AUC of the precision–recall curve.

Recent advances have highlighted the Matthews correlation coefficient (MCC) and balanced accuracy (BA) as more reliable alternatives (Chicco et al., 2021). BA, in particular, has gained traction due to its simplicity, interpretability, and linearity with respect to TPR and TNR. This metric ensures balanced error rates across classes and is equivalent to standard accuracy when the dataset is perfectly balanced. For multiclass scenarios, BA is extended by averaging the TPRs for all classes—i.e., the average recall—providing a fair evaluation across all categories and maintaining linear equivalence to the mean per class error and informedness. In the literature, three works explicitly adopted the BA (Simonetta et al., 2023; Nakamura and Takaki, 2015), while another two reported it under different names (‘overall accuracy’ or ‘average recall’) (Verma and Thickstun, 2019; Mirza et al., 2024).

In this work, BA is recommended as the preferred evaluation metric due to its robustness and interpretability, ensuring a more reliable assessment of model performance in the context of authorship attribution. Moreover, it can be derived directly from the confusion matrix and is equivalent to standard accuracy when the classes are evenly distributed. The BAs of studies providing confusion matrices was recalculated for 10 works. Additionally, 19 works used the standard accuracy for datasets perfectly balanced. Overall, 33 works out of 58 were included in the analysis using BA as measure of merit.

Regarding the hold‑out strategies, only 56 of the 88 recorded experiments use some form of cross‑validation, of which 25 use a ‘leave‑one‑out’ strategy. Of the remaining 32 experiments, 5 use random folds and 27 use a simple hold‑out strategy.

The selected studies were categorized into three groups: ‘unreliable,’ ‘questionable evaluation,’ and ‘good evaluation.’ The ‘unreliable’ category includes publications using inappropriate evaluation metrics and/or lacking cross‑validation. The ‘good evaluation’ category includes studies where balanced accuracy could be computed and a ‑fold cross‑validation strategy was employed. The ‘questionable evaluation’ category encompasses all remaining works, such as those using random‑fold cross‑validation or relying on the ‑measure. It must be noted that, in those cases in which the dataset size and model complexity are really large, cross‑validation procedures are hardly applicable.

As illustrated in Figure 4, we classified 27 experiments as having a ‘good evaluation,’ 21 as having a ‘questionable evaluation,’ and 40 as having an ‘unreliable’ evaluation. Additionally, Figure 3 shows the number of experiments per year among publications across the three evaluation categories. The ‘unreliable’ category is prevalent in the earlier years, while the ‘good evaluation’ category has gained prominence in recent years, particularly since 2010.

Figure 3

Figure 4

5 Repertoire and Datasets

Some existing studies approach composer identification primarily as a classification benchmark rather than as a musicological authorship–attribution task. Particularly in earlier decades, computational models were frequently tested on heterogeneous repertoires, often spanning from the Baroque era to the 20th century, and, in some instances, including medieval Gregorian chants (Cruz‑Alcázar et al., 2003). Such material is typically selected because, when the composers compared have markedly different cultural and aesthetic backgrounds, their musical styles can often be distinguished more easily through quantitative methods, simplifying the statistical challenge. While this approach may be suitable for preliminary evaluations or for benchmarking tasks where the goal is to test a model's ability to capture gross stylistic differences, it is insufficient to demonstrate a model's capability to identify composers' styles accurately for the purpose of musicological attribution. Effective evaluation for musicological authorship attribution must ensure the model is discriminating based on nuanced, individual musical style rather than easily separable extra musical factors that may correlate with style, such as genre, historical period, or geographical region (Rudman, 2012). Relying on these broader factors limits the model's applicability for distinguishing between, for example, contemporaneous composers within the same tradition or school.

Beyond the authorship‑attribution task, related MIR tasks include the automatic tagging of music in large databases (van Nuss et al., 2017) and comparative studies of composers' styles—often focusing on Mozart and Haydn—without addressing specific attribution problems (Backer and van Kranenburg, 2005; Hillewaere et al., 2010; Hontanilla et al., 2011; Wołkowicz and Kešelj, 2013; Velarde et al., 2016; Hajj et al., 2018; Velarde et al., 2018; Verma and Thickstun, 2019; Kempfert and Wong, 2020). Even in these cases, the material selected for evaluation should not be distinguishable based on extra musical factors to ensure the validity of the results.

As shown in Figure 6, the most frequently studied composer in composer‑identification tasks is Bach, followed by Mozart, Haydn, Beethoven, and Chopin. In recent years, there has been increased focus on Mozart, Haydn, and Beethoven, as illustrated in Figure 5. This distribution is likely influenced by the popularity of the CCARH corpora (Sapp, 2005), which are often used for music analysis in the symbolic domain. Custom datasets have been rarely employed, and, when used, they were typically unsupervised by musicologists, mined from user‑contributed archives, and biased towards German composers and Western art music. Exceptions exist, primarily involving Chinese and Japanese folk music (Shan et al., 2002; Ruppin and Yeshurun, 2006) and popular music, particularly The Beatles (Pollastri and Simoncelli, 2001; Glickman et al., 2019).

Figure 5

Figure 6

The median dataset size in the reviewed experiments is 251 instances, with a minimum of 24 and a maximum of 1.2 million. A total 85% of the experiments used a dataset with less than 1000 instances—Figure 8 shows the distribution of the dataset size. However, comparing dataset sizes is challenging because individual instances may differ in duration, number of notes, measures, or movements. While this variation is less significant during the training phase, it is critical for the test phase, where true and false positives/negatives are counted in terms of instances.

Figure 7

Figure 8

The average number of classes in the datasets is 4, with a minimum of 2 and a maximum of 19, but some authors did not report this aspect (Dor and Reich, 2011; van Nuss et al., 2017). For practical applications, the capacity to handle a large number of classes is generally advantageous but not essential if the composers are carefully selected by domain experts.

Only a few datasets have been used in more than three studies. Their details and associated accuracies are summarized in Table 2.

I attempted to evaluate the model accuracy—or balanced accuracy where it could be computed—based on the type of dataset. The resulting scatter plots are shown in Figures 7 and 8.

6 Representational Frameworks for Symbolic Music

The way symbolic music is represented significantly impacts the effectiveness of composer identification models. The literature showcases a variety of approaches, which can be broadly categorized by how features are derived and their fundamental representational form, as detailed below. These representations are typically derived from various digital music file formats (see Section 6.2).

7 Model Architectures

The task of composer identification has been approached with various models, each leveraging different strengths in statistical learning. This section categorizes these models based on their underlying architecture and discusses their suitability for different types of feature representations, as detailed in Section 6. Figure 9 shows the distribution of papers employing each approach.

Figure 9

To create these statistics, base classes of approaches were identified, and each recorded experiment was associated with a specific class. For instance, decision trees fall into the class of ‘Trees,’ random forests into the class of ‘Forests,’ and the same applies to CNNs, transformers, and feed‑forward NNs, all classified as ‘NN’. Similarly, ‘Markov Processes’ are used for any methodology that leverages the theory of Markov processes, including hidden Markov models and simple Markov chains.

7.3 Markov processes

Markov processes are based on the assumption that music is a Markov process, where the probability of a musical symbol depends solely on the preceding one. This approach aligns with sequential and pattern‑based descriptors. The musical symbol can vary but often includes discrete elements like pitches, intervals, durations, or combinations thereof. This category includes Markov chains (Kaliakatsos‑Papakostas et al., 2011; Nakamura and Takaki, 2015; Neto and Pereira, 2017), ‑grams (Cruz‑Alcázar et al., 2003; WoŁkowicz et al., 2008; Hillewaere et al., 2010; Hontanilla et al., 2011, 2013; Wołkowicz and Kešelj, 2013; Hedges et al., 2014; Alvarez et al., 2024; Ruppin and Yeshurun, 2006; Shan et al., 2002), and hidden Markov models (Pollastri and Simoncelli, 2001). Markov models often struggle with representing music polyphony. Methods dealing with polyphonic music range from considering only the highest part, simplifying it and automatically extracting the melody (Simonetta et al., 2019), to concatenating the various parts sequentially and modeling each part with a different model. As shown in Figure 10, the use of Markov processes has declined in recent times, likely being replaced by NNs with long short‑term memory or attention methods (Chou et al., 2021; Zhang et al., 2023).

Figure 10

8 Authorship Attribution Problems

Only three cases of unknown or questionable authorship have been approached with composer identification models. The first case involves a set of keyboard fugues originally attributed to Johann Sebastian Bach but then recently re‑attributed to Krebs, supported by various statistical analyses (van Kranenburg and Backer, 2005; van Kranenburg, 2008; Hontanilla et al., 2011). The second case pertains to more than 200 works of questionable attribution by Josquin des Prez (Brinkman et al., 2016; McKay et al., 2018; Verma and Thickstun, 2019). Finally, the third case concerns eight songs (or portions of songs) with disputed attribution to John Lennon and Paul McCartney (Glickman et al., 2019). A synthesis of the existing authorship attribution problems is shown in Table 3.

In the Bach fugue problem, the initial work (van Kranenburg and Backer, 2005) identified a set of possible alternative authors based on historical observations and musicological considerations, thus restricting the inference problem to three authors: Johann Sebastian Bach, Wilhelm Friedemann Bach, and Johann Ludwig Krebs. The authors collected music scores from the CCARH collection (Sapp, 2005). The same authors later included Johann Peter Kellner as an option in a more extended analysis (van Kranenburg, 2008). Other researchers approached the problem in the second formulation, obtaining similar results (Hontanilla et al., 2011). It should be noted that the full text of the latter work was not accessible, and the analysis was based on the conference slides made available by the authors and on a subsequent publication describing the same method in full detail but excluding the authorship attribution problem (Hontanilla et al., 2013). Upon examining the evaluation procedures of these three works, it was found that the first work did not adopt proper evaluation measures for the imbalanced dataset at hand, while the third work showed poor classification performance, thus undermining the reliability of the model. The second work (van Kranenburg and Backer, 2005), however, allowed for proper evaluations based on the confusion matrix reported in the original publication, showing a balanced accuracy greater than 90% in a leave‑one‑out evaluation procedure. Therefore, it is suggested that this latter work (van Kranenburg, 2008) contains the most definitive evaluation up to now. Regarding the construction of the dataset, not much attention was paid by the authors on the elimination of potential confounding factors stemming from editorial choices in the CCARH collection. The authors acknowledged the limits of the closed set formulation, stating ‘the possibility that a composer not represented in the dataset wrote the piece should be kept open’ (van Kranenburg, 2008).

Regarding the case of Josquin's works, the originating publication (Brinkman et al., 2016) did not perform careful musicological research to restrict the set of possible authors. Instead, reliance was placed on a musicologically curated dataset containing various well‑known Renaissance composers collected by the Joquin Research Project (JRP).¹ This approach enabled the automatic attribution of labels to more than 200 questionable works. While this method is more suitable for large‑scale problems, it lacks the reliability of a meticulous musicological analysis that narrows the possible attribution labels. Observing the evaluation procedure, the originating work employed a single hold‑out strategy, resulting in a balanced accuracy of 50%. A subsequent study (McKay et al., 2018) adopted a 10‑fold cross‑validation procedure and evaluated the ‑measure, which partially accounts for dataset imbalance. This study achieved 91% and 82% accuracy for two different two‑classes problems (Josquin vs. Ockeghem and Josquin vs. La Rue, respectively). Additionally, the average balanced accuracy of a 10‑fold cross‑validation performed on the six classes was computed by Verma and Thickstun, yielding a value of 77% (Verma and Thickstun, 2019). Overall, recent improvements allow for a more rigorous approach to the problem, but the attribution of suspicious Josquin works remains an open challenge. Even in this case, the authors rely on a weak process for data collection, with only one musicologist responsible for the transcription—thus, no external review was performed—and with no musicological variants encoded. From the information available in the JRP documentation, it is not clear how dubious passages have been transcribed and encoded and how potential biases from the transcriber (e.g., systematic interpretation or error patterns) have been mitigated. However, McKay et al. (2018) placed particular attention on the automatic handling of confounding factors deriving from encoding choices, which can partly mitigate the problem.

In the case of the disputed attribution of Lennon–McCartney song portions, the originating work by Glickman et al. (2019) approached the problem by analyzing symbolic representations. Specifically, they used published music score editions of 70 songs by The Beatles as their primary source material. From these symbolic scores, they manually extracted a set of hand‑crafted features designed to capture compositional style, distinct from performance or recording artifacts. These features were then used to train a logistic linear regression model with both and regularization factors. The model's performance was evaluated using standard accuracy in a leave‑one‑out validation scheme, achieving 76% correct predictions for disputed song excerpts. This study is notable as one of the few prominent computational authorship attribution investigations in popular music that operates on symbolic data, employing a methodology of feature extraction from scores comparable to approaches used for classical music. However, the result indicates that the re‑attribution of these song excerpts remains an open challenge. The authors, interested in establishing ‘the stylistic fingerprint of a songwriter based solely on a corpus of songs' musical content,’ based their training on existing published collections. The study did not delve into a detailed investigation of potential editorial choices within these music transcriptions, which can be a complex factor for popular music that often has a strong oral tradition alongside published forms. Future work could enhance the validation and feature extraction using more recent tools and methodologies.

9 Takeaways for Future Research in Composer Identification

This section summarizes the major findings of the research that may be useful to the MIR community for further investigations in the analysis of musical styles and especially for improving composer identification models.

10 Conclusions

This systematic review has meticulously surveyed the landscape of composer identification and authorship attribution from symbolic music scores, drawing critical insights from 58 peer‑reviewed studies spanning decades of computational musicology. My analysis underscores a pivotal challenge: while technological advancements have introduced sophisticated machine learning models, the reliability and musicological validity of attribution claims often remain compromised by inconsistent evaluation practices and inherent dataset limitations.

Key findings highlight the critical importance of robust evaluation metrics, particularly BA and MCC, over traditional accuracy, especially when dealing with imbalanced datasets common in this domain. I identify a pervasive reliance on simple hold‑out validation and a concentration on a limited repertoire of Western classical composers, often overlooking crucial aspects of data integrity and musicological curation. The case studies on Bach, Josquin des Prez, and Lennon–McCartney illustrate the complex interplay between computational predictions and musicological interpretation, revealing that even high classification scores do not automatically translate to definitive authorship.

To address these gaps, this work proposes a comprehensive set of guidelines for future research. These recommendations emphasize the necessity of transparent, musicologically informed data‑collection protocols, the adoption of rigorous cross‑validation strategies, and the judicious interpretation of model outputs, ideally within a Bayesian framework to quantify uncertainty. Ultimately, while computational methods offer powerful tools for stylistic analysis, their true potential for authorship attribution can only be realized through interdisciplinary collaboration and a cautious, evidence‑based approach that integrates machine learning insights with deep musicological expertise.

Funding Statement

This work has been funded by the European Union (Horizon Programme for Research and Innovation 2021–2027, ERC Advanced Grant ‘The Italian Lauda: Disseminating Poetry and Concepts Through Melody (12th–16th century),’ acronym LAUDARE, project no. 101054750). The views and opinions expressed are, however, only those of the author and do not necessarily reflect those of the European Union or the European Research Council. Neither the European Union nor the awarding authority can be held responsible for such matters.

Competing Interests

The author has no competing interests to declare.

Additional Files

The additional files for this article can be found using the links below: