基于串核的音乐风格聚类

Music style clustering with string kernel

给出了一种新的映射音乐到Rn空间的方法和基于串核的音乐风格聚类法。利用统计方法分析大量音乐的旋律轮廓线得到合适的编码模式,用它把旋律轮廓线编码为有限字母表(8个字母)的字符串。利用连续子串嵌入法把音乐串显式映射到高维Rn空间,并用核表示这一映射。通过用基于核的山方法选择聚类的适合初始点,最后使用基于核的K-means方法聚类音乐数据集,比较了3个不同串核在5个音乐数据集上的聚类性能。

A novel mapping of embedding music to R^n space and a new kernel-based clustering method is presented. First, Analyzing melody contours from a huge music dataset, we translate music datasets to strings with an alphabet including only eight letters. Second, a string kernel is obtained by mapping music strings to high dimension R^n space using contiguous substrings embedding method. Several data points gotten using mountain method on the music kernels are used as initial points of kernel clustering methods, for example K-means method. Finally, Performances of three string kernels on five music datasets are evaluated by the kernel K-means algorithm.