|Title:||Improving supervised music classification by means of multi-objective evolutionary feature selection|
|Abstract:||In this work, several strategies are developed to reduce the impact of the two limitations of most current studies in supervised music classification: the classification rules and music features have often a low interpretability, and the evaluation of algorithms and feature subsets is almost always done with respect to only one or a few common evaluation criteria separately. Although music classification is in most cases user-centered and it is desired to understand well the properties of related music categories, many current approaches are based on low-level characteristics of the audio signal. We have designed a large set of more meaningful and interpretable high-level features, which may completely replace the baseline low-level feature set and are even capable to significantly outperform it for the categorisation into three music styles. These features provide a comprehensible insight into the properties of music genres and styles: instrumentation, moods, harmony, temporal, and melodic characteristics. A crucial advantage of audio high-level features is that they can be extracted from any digitally available music piece, independently of its popularity, availability of the corresponding score, or the Internet connection for the download of the metadata and community features, which are sometimes erroneous and incomplete. A part of high-level features, which are particularly successful for classification into genres and styles, has been developed based on the novel approach called sliding feature selection. Here, high-level features are estimated from low-level and other high-level ones during a sequence of supervised classification steps, and an integrated evolutionary feature selection helps to search for the most relevant features in each step of this sequence. Another drawback of many related state-of-the-art studies is that the algorithms and feature sets are almost always compared using only one or a few evaluation criteria separately. However, different evaluation criteria are often in conflict: an algorithm optimised only with respect to classification quality may be slow, have high storage demands, perform worse on imbalanced data, or require high user efforts for labelling of songs. The simultaneous optimisation of multiple conflicting criteria remains until now almost unexplored in music information retrieval, and it was applied for feature selection in music classification for the first time in this thesis, except for several preliminary own publications. As an exemplarily multi-objective approach for optimisation of feature selection, we simultaneously minimise the classification error and the number of features used for classification. The sets with more features lead to a higher classification quality. On the other side, the sets with fewer features and a lower classification performance may help to strongly decrease the demands for storage and computing time and to reduce the risk of too complex and overfitted classification models. Further, we describe several groups of evaluation criteria and discuss other reasonable multi-objective optimisation scenarios for music data analysis.|
|Subject Headings:||High level audio features|
Multi-objective evolutionary feature selection
|Appears in Collections:||LS 11|
This item is protected by original copyright
All resources in the repository are protected by copyright.