Igor Vatolkin

Multi-objective feature selection in music genre and style recognition tasks

Feature selection is an important prerequisite for music classification which in turn is becoming more and more ubiquitous since entering the digital music age. Automated classification into genres or even personal categories is currently envisioned even for standard mobile devices. However, classifiers often fail to work well with all available features, and simple greedy methods often fail to select good feature sets, making feature selection for music classification a natural field of application for evolutionary approaches in general, and multi-objective evolutionary algorithms in particular. In this work, we study the potential of applying such a multi-objective evolutionary optimization algorithm for feature selection with different objective sets. The result is promising, thus calling for deeper investigations of this approach.

Multi-objective evolutionary feature selection for instrument recognition in polyphonic audio mixtures

Instrument recognition is one of the music information retrieval research topics. This task becomes very challenging if several instruments are played simultaneously because of their varying physical characteristics: inharmonic attack noise, energy development during attack–decay–sustain–release envelope or overtone distribution. In our framework, we treat instrument detection as a machine-learning task based on a large amount of preprocessed audio features with target to build classification models. Since classification algorithms are very sensitive to feature input and the optimal feature set differs from instrument to instrument, we propose to run a multi-objective feature selection procedure before building of classification models. Two objectives are considered for evaluation: classification mean-squared error and feature rate (smaller amount of features stands for reduced costs and decreased risk of overfitting). The analysis of the extensive experimental study confirms that application of an evolutionary multi-objective algorithm is a good choice to optimize feature selection for music instrument identification.

Selecting small audio feature sets in music classification by means of asymmetric mutation

Classification of audio recordings is often based on audio-signal features. The number of available variables is usually very large. For successful categorization in e.g. genres, substyles or personal preferences small, but very predictive feature sets are sought. A further challenge is to solve this feature selection problem at least approximately with short run lengths to reduce the high computational load. We pursue this goal by applying asymmetric mutation operators in simple evolutionary strategies, which are further enhanced by mixing in greedy search operators. The resulting algorithm is reliably better than any of these approaches alone and in most cases clearly better than a deterministic greedy strategy.

Design and comparison of different evolution strategies for feature selection and consolidation in music classification

Music classification is a complex problem which has gained high relevance for organizing large music collections. Different parameters concerning feature extraction, selection, processing and classification have a strong impact on the categorization quality. Since it is very difficult to design a deterministic approach which provides the efficient parameter tuning, we haven chosen a heuristic approach. In our work we apply and compare different evolution strategies for the optimization of feature selection and consolidation using three pre-defined personal user categories. Concepts of local search operators with domain-specific knowledge and self-adaptation are examined. Several suggestions based on an empirical study are discussed and ideas for future work are given.

Multi-Objective Evaluation of Music Classification

Music classification targets the management of personal music collections or recommendation of new songs. Several steps are required here: feature extraction and processing, selection of the most relevant of them, and training of classification models. The complete classification chain is evaluated by a selected performance measure. Often standard confusion matrix based metrics like accuracy are calculated. However it can be valuable to compare the methods using further metrics depending on the current application scenario. For this work we created a large empirical study for different music categories using several feature sets, processing methods and classification algorithms. The correlation between different metrics is discussed, and the ideas for better algorithm evaluation are outlined.

Huge Music Archives on Mobile Devices

The availability of huge nonvolatile storage capacities such as flash memory allows large music archives to be maintained even in mobile devices. With the increase in size, manual organization of these archives and manual search for specific music becomes very inconvenient. Automated dynamic organization enables an attractive new class of applications for managing ever-increasing music databases. For these types of applications, extraction of music features as well as subsequent feature processing and music classification have to be performed. However, these are computationally intensive tasks and difficult to tackle on mobile platforms. Against this background, we provided an overview of algorithms for music classification as well as their computation times and other hardware-related aspects, such as power consumption on various hardware architectures. For mobile platforms such as smartphones, a careful balance of algorithm complexity, hardware architecture, and classification accuracy has to be found to provide a high quality user experience.

Improving supervised music classification by means of multi-objective evolutionary feature selection

1

Partition Based Feature Processing for Improved Music Classification

Identifying desired music amongst the vast amount of tracks in today’s music collections has become a task of increasing attention for consumers. Music classification based on perceptual features promises to help sorting a collection according to personal music categories determined by the user’s personal taste and listening habits. Regarding limits of processing power and storage space available in real (e.g. mobile) devices necessitates to reduce the amount of feature data used by such classification. This paper compares several methods for feature pruning– experiments on realistic track collections show that an approach attempting to identify relevant song partitions not only allows to reduce the amount of processed feature data by 90% but also helps to improve classification accuracy. They indicate that a combination of structural information and temporal continuity processing of partition based classification helps to substantially improve overall performance.

Computational Prediction of High-Level Descriptors of Music Personal Categories

Digital music collections are often organized by genre relationships or personal preferences. The target of automatic classification systems is to provide a music management limiting the listener’s effort for the labeling of a large number of songs. Many state-of-the art methods utilize low-level audio features like spectral and time domain characteristics, chroma etc. for categorization. However the impact of these features is very hard to understand; if the listener labels some music pieces as belonging to a certain category, this decision is indeed motivated by instrumentation, harmony, vocals, rhythm and further high-level descriptors from music theory. So it could be more reasonable to understand a classification model created from such intuitively interpretable features. For our study we annotated high-level characteristics (vocal alignment, tempo, key etc.) for a set of personal music categories. Then we created classification models which predict these characteristics from low-level audio features available in the AMUSE framework. The capability of this set of low level features to classify the expert descriptors is investigated in detail.

Optimization of Feature Processing Chain in Music Classification by Evolution Strategies

In this paper a new method based on evolution strategies (ES) is presented to optimize a classifier for personal music categories. The user assigns songs to multiple personal music categories: Examples from each category are selected in order to train a category-specific classifier using musical features as input. The classifier then ranks all songs according to their similarity to the category examples. Since an exhaustive search for parameters maximizing the classifier performance is not feasible an ES is applied. The experiments show a significant performance increase for various music categories due to the ES optimization.