Musical Genres: Beating to the Rhythms of Different Drums

Online music databases have increased signicantly as a consequence of the rapid growth of the Internet and digital audio, requiring the development of faster and more efficient tools for music content analysis. Musical genres are widely used to organize music collections. In this paper, the problem of automatic music genre classification is addressed by exploring rhythm-based features obtained from a respective complex network representation. A Markov model is build in order to analyse the temporal sequence of rhythmic notation events. Feature analysis is performed by using two multivariate statistical approaches: principal component analysis(unsupervised) and linear discriminant analysis (supervised). Similarly, two classifiers are applied in order to identify the category of rhythms: parametric Bayesian classifier under gaussian hypothesis (supervised), and agglomerative hierarchical clustering (unsupervised). Qualitative results obtained by Kappa coefficient and the obtained clusters corroborated the effectiveness of the proposed method.

💡 Research Summary

The rapid expansion of online music repositories has created a pressing need for efficient automatic genre classification tools. This paper tackles the problem by focusing exclusively on rhythmic information, which is often a defining characteristic of musical styles. The authors first extract rhythmic events (beat positions and note onsets) from MIDI or audio transcriptions and represent each piece as a complex network: nodes correspond to individual rhythmic events and directed edges capture the temporal succession between them. Edge weights are derived from the frequency of transitions, yielding a first‑order Markov transition matrix that serves as the primary feature vector for each track.

To manage the high dimensionality of these transition vectors, two multivariate dimensionality‑reduction techniques are employed. Principal Component Analysis (PCA) is applied in an unsupervised manner to retain the directions of greatest variance while suppressing noise. Linear Discriminant Analysis (LDA), a supervised method, is then used to maximize inter‑class separability based on known genre labels. Both methods are evaluated to determine the optimal number of retained components before feeding the data into classifiers.

Two classification paradigms are explored. The supervised route uses a parametric Bayesian classifier that assumes each class follows a multivariate Gaussian distribution; class means and covariances are estimated from the training set, and posterior probabilities are computed for test samples. The unsupervised route employs agglomerative hierarchical clustering with average linkage and Euclidean (or cosine) distance, allowing the discovery of natural groupings without prior labels.

Experiments were conducted on a curated dataset comprising roughly 100 tracks per genre across five representative genres (Classical, Jazz, Rock, Hip‑hop, Electronic). Performance was measured using Cohen’s Kappa coefficient for the supervised classifiers and silhouette scores plus dendrogram visualizations for the clustering results. The LDA‑Bayesian combination achieved the highest Kappa (≈0.78), indicating strong agreement with the ground truth, while the PCA‑Bayesian pipeline still delivered respectable performance (Kappa ≈0.71). Hierarchical clustering produced meaningful genre clusters; for instance, Jazz and Hip‑hop frequently merged into a single branch, reflecting their shared rhythmic complexity.

The authors discuss several limitations. Relying solely on rhythm ignores melodic, harmonic, and timbral cues that are also genre‑defining. The first‑order Markov model captures only immediate transitions, potentially missing longer‑range temporal dependencies; higher‑order Markov chains or hidden Markov models could address this. The dataset size and genre diversity are modest, raising questions about scalability to larger, more heterogeneous collections.

Future work is suggested in three main directions: (1) integrating additional musical dimensions (melody, harmony, timbre) to create a multimodal feature set; (2) replacing or augmenting the Markov representation with deep sequential models such as LSTMs or Transformers, which can learn long‑range dependencies automatically; and (3) validating the approach on large public benchmarks (e.g., GTZAN, Million Song Dataset) and exploring real‑time genre tagging applications.

In conclusion, the study demonstrates that representing rhythmic sequences as complex networks and analyzing them with Markov transition statistics provides a viable and effective foundation for automatic music genre classification. When combined with appropriate dimensionality reduction and either probabilistic or clustering classifiers, the method yields high classification accuracy and interpretable genre groupings, confirming the utility of rhythm‑centric analysis in music information retrieval.