Music Classification Based On SVM Active Learning

In recent years, the rapid development of the internet technology and multimedia information technology officially declared that the age of big data has been coming in twenty-first Century, how to retrieve the useful information from the information of the mass data will have great research significance.Multimedia information on the Internet certainly includes the digital music with the rapid growth of the number, a lot of singers is emerging, the massive album and network songs have become available, also influenced by the development of world cultural diversity, many kinds of music style have emerged too, in order to satisfy the people according to their different preferences to accurately and quickly find which song they want to query, this requires a music retrieval system witch is more efficient and fast. However, the traditional music classification system will train music samples firstly to obtained the classification model, and then predict the class of unknown music samples,the effect of this classifier with this kind of traditional classification method often depends on the number of training samples. To label the mass of training samples manually is clearly not realistic, active learning can be a very good solution for this problem.SVM(Support Vector Machine) is a kind of excellent machine learning methods, active learning method is combined with SVM in this paper, and this method is applied to the music genre classification. The traditional sample selection strategy of active learning method based on SVM is only confined to the uncertainty of sample, that if a sample is nearest to the classification hyper plane, means the value of this sample is maximum.The improved method proposed in this paper based this algorithm is described as follows:(1)When we select the most valuable samples, we will consider to choose the samples those distance to classification hyper plane is near, and at the same time ensure the diversity of them. Because the dimension of music samples characteristics is relatively high, the angle between samples is selected as the sample diversity measure in this paper, and thus established the ultimate criterion of sample value: score;(2) "1-v-r one versus the rest" is a common method of applying SVM to multi-classification, but this method will cause the deflection of data sets artificially, it will have a certain bad influence to the final classification effect, so the sample balance judgment standard parameter:b is established in this paper,when the number of valuable samples selected bythe active learning method do not satisfy the equilibrium conditions, the balance adjustment for them will be done.