music.mcgill.ca This paper presents ACE (Autonomous Classification Engine), a framework for using and optimizing classifiers. Given a set of feature vectors, ACE experiments with a variety of classifiers, classifier parameters, classifier ensembles and dimensionality reduction techniques in order to arrive at a good configuration for the problem at hand. In addition to evaluating classification methodologies in terms of success rates, functionality is also being incorporated into ACE allowing users to specify constraints on training and classification times as well as on the amount of time that ACE has to arrive at a solution. ACE is designed to facilitate classification for those new to pattern recognition as well as provide flexibility for those with more experience. ACE is packaged with audio and MIDI feature extraction software, although it can certainly be used with existing feature extractors. This paper includes a discussion of ways in which existing general-purpose classification software can be adapted to meet the needs of music researchers and shows how these ideas have been implemented in ACE. A standardized XML format for communicating features and other information to classifiers is proposed. A special emphasis is placed on the potential of classifier ensembles, which have remained largely untapped by the MIR community to date. A brief theoretical discussion of ensemble classification is presented in order to promote this powerful approach.

This paper experimentally investigates the classification utility of combining features extracted from separate audio, symbolic and cultural sources of musical information. This was done via a series of genre classification experiments performed using all seven possible combinations and subsets of the three corresponding types of features. These experiments were performed using jMIR, a software suite designed for use both as a toolset for performing MIR research and as a platform for developing and sharing new algorithms. The experimental results indicate that combining feature types can indeed substantively improve classification accuracy. Accuracies of 96.8 % and 78.8 % were attained respectively on 5 and 10-class genre taxonomies when all three feature types were combined, compared to average respective accuracies of 85.5 % and 65.1 % when features extracted from only one of the three sources of data were used. It was also found that combining feature types decreased the seriousness of those misclassifications that were made, on average, particularly when cultural features were included. 1.

A library of 160 high-level features is presented along with jSymbolic, a software package that extracts these features from MIDI files. jSymbolic is intended both as a platform for developing new features as well as a tool for providing features to data mining software that can be used to automatically classify music or evaluate musical similarity. 1

Hierarchical taxonomies of classes arise in the analysis of many types of musical information, including genre, as a means of organizing overlapping categories at varying levels of generality. However, incorporating hierarchical structure into conventional machine learning systems presents a challenge: the use of independent binary classifiers for each class in the hierarchy can produce hierarchically inconsistent predictions. That is, an example may be assigned to a class, and not assigned to the parent of that class. This paper applies a Bayesian framework to combine, or aggregate, a hierarchy of multiple binary classifiers in a principled manner, and consequently improves performance over the hierarchy as a whole. Furthermore, such an approach allows for an arbitrarily complex hierarchy, and does not suffer from classes that are too broad or too refined. Experiments on the MIREX 2005 symbolic genre classification dataset show that our Bayesian Aggregation algorithm provides significant improvement over independent classifiers, and demonstrates superior performance compared to previous work. Our method also improves similarity search by ranking songs by similarity of hierarchical predictions to those of a query song. 1

jMIR is a free and open-source software suite designed for applications related to automatic music classification. jMIR includes the jAudio, jSymbolic and jWebMiner feature extractors, the ACE meta-learning framework, the ACE XML information exchange file formats, the jMusicMetaManager musical dataset management software and the Codaich, Bodhidharma MIDI and SAC musical datasets. The primary goals underlying jMIR are the provision of powerful ready-to-use software tools to music researchers with diverse ranges of technical backgrounds, the encouragement of research combining features derived from audio, symbolic and cultural sources of data, and the provision of a framework for iteratively and collaboratively developing further music information retrieval (MIR) tools and performing original music classification research. 1.

Music classification continues to be an important component of music information retrieval research. An underutilized tool for improving the performance of classifiers is feature weighting. A major reason for its unpopularity, despite its benefits, is the potentially infinite calculation time it requires to achieve optimal results. Genetic algorithms offer potentially sub-optimal but reasonable solutions at much reduced calculation time, yet they are still quite costly. We investigate the advantages of implementing genetic algorithms in a parallel computing environment to make feature weighting an affordable instrument for researchers in MIR.

mail.mcgill.ca This paper describes ACE, a framework for automatically finding effective classification methodologies for arbitrary supervised classification problems. ACE performs experiments with both individual classifiers and classifier ensembles in order to find the approaches best suited to particular problems. A special emphasis is placed on classifier ensembles, as they can be powerful tools, yet are currently rarely used in MIR research. In addition to evaluating various classification methodologies in terms of success rates, ACE also allows users to specify constraints on training and classification times. The input to ACE is an arbitrary taxonomy accompanied by training feature vectors and their model classifications. ACE then outputs comparisons of the effectiveness of different classification methodologies, including information relating to feature weightings, dimensionality reduction and classifier combination techniques. The user may then select any of these configurations, after which s/he will be presented with trained classifiers that can be used to classify new feature vectors. Although designed to be used easily with any existing feature extraction system, ACE is also packaged with MIDI and audio feature extraction sub-systems. In addition, plans are underway to make use of distributed computing in order to decrease processing times. 1.

This paper discusses the results of the MIREX 2005 symbolic genre classification contest and describes the Bodhidharma system, which attained the highest classification success rates in all four of the evaluated categories. Five systems were submitted to this contest, which was conducted independently at the University of Illinois at Urbana-Champagne (UIUC). Each system was evaluated in two different experiments, one involving thirty-eight genre classes and one involving nine classes. Success rates were measured in two ways: one based only on how well each system was able to find the single correct genre of each recording, and the other giving partial scores to incorrect classifications that were relatively close to the correct genre. Evaluations were performed using stratified cross-validation. Bodhidharma is a sophisticated system that utilizes a combination of flat, hierarchical and round-robin classification strategies based on classifier ensembles consisting of feedforward neural networks and k-nearest neighbour classifiers. Bodhidharma bases its classifications on 111 high-level features that it extracts from MIDI recordings. Each classifier ensemble uses genetic algorithms to evolve a weighted sub-set of the features that are appropriate for that particular ensemble.

Automatic music classification is a wide-ranging and multidisciplinary area of inquiry that offers significant benefits from both academic and commercial perspectives. This dissertation focuses on the development of jMIR, a suite of powerful, flexible, accessible and original software tools that can be used to design, share and apply a wide range of automatic music classification technologies. jMIR permits users to extract meaningful information from audio recordings, symbolic musical representations and cultural information available on the Internet; to use machine learning technologies to automatically build classification models; to automatically collect profiling statistics and detect metadata errors in musical collections; to perform experiments on large, stylistically diverse and well-labelled collections of music in both audio and symbolic formats; and to store and distribute information that is essential to automatic music classification in expressive and flexible standardised file formats. In order to have as diverse a range of applications as possible, care was taken to avoid tying jMIR to any particular types of music classification. Rather, it is designed to be a

This paper 1 describes both a newly developed method for manual annotation for aspects of melodic similarity and its use for evaluating melody features concerning their contribution to perceived similarity. The second issue is also addressed with a computational evaluation method. These approaches are applied to a corpus of folk song melodies. We show that classification of melodies could not be based on single features and that the feature sets from the literature are not sufficient to classify melodies into groups of related melodies. The manual annotations enable us to evaluate various models for melodic similarity. 1