基于SVM主动学习的音乐分类

摘要

An improved SVM(Support Vector Machine)active learning method is proposed. By providing the user with the most informative samples which are put into training set through several iterations, the cost of manually labelled samples can be greatly reduced. In the experiment, to evaluate the performance of the classifier, it classifies 801 songs according to five kinds of genres(including dance, lyric, jazz, folk, rock), and verifies the effectiveness of SVM active learning in two aspects which are the accuracy convergence rate and the number of samples need to be labelled to achieve the same accuracy.%提出了一种改进的SVM（支持向量机）主动学习方法，通过多次迭代提供给用户信息量最大的样本并将其加入训练集，可以大大减少人工标记样本所耗费的代价。为了评估分类器的性能，实验中对包含了五种音乐流派类别（舞曲、抒情、爵士、民乐、摇滚）的801首音乐样本进行了分类，并在分类准确率的收敛速度和达到同等准确率下需要标注的样本数目两个方面验证了提出的SVM主动学习方法的有效性。