| The rapid development of abundance audio data retrieval technologies pose severe challenges to the existing feature selection algorithms. It makes a strong demand for feature selection methods with higher performance which are fit for the audio dataset. This dissertation mainly studies on feature selection based on high-dimensional music dataset. Contributions in this dissertation mainly include:1. We compared the existing feature selection algorithms. Wrapper and Filter methods are compared through experiments.2. Most Wrapper methods can not evaluate features in a measurable way, based on this demerits, the CCRS algorithm is proposed. This method evaluates the contribution values for every feature using classification accuracies of series feature subsets. This provides important information for feature analysis and classification model construction.3. ReliefF + Correlation Analysis and ReliefF + PCA/LDA methods are used to improve ReliefF algorithm against its default that ReliefF can not eliminate redundancy information. Then ReliefPCA algorithm is proposed, which uses PCA feature space transformation to eliminate redundancy information, while retaining the intelligibility of features.4. Based on the merits and demerits of Wrapper model and Filter model, a two phase feature selection method ReliefGA with Filter-Wrapper model is proposed, which uses the feature evaluation of ReliefF to instruct the initialization of genetic population, the coupling model aims to improve the efficiency of genetic algorithm which use the performance of the classifier as evaluation of feature subsets. Experiments show that the algorithm has good comprehensive performance with respects to accuracy, size of feature subsets, and time complexity.This paper takes the high redundancy and high correlation of audio data into account, and studies the feature selection technologies on music dataset. We get good results in retaining the intelligibility of features and improving the classification accuracies. |
PDF Full Text Request