基于马氏距离的分段矢量量化时间序列分类

提出一种基于马氏距离的分段矢量量化时间序列分类(Mahalanobis distance-based time series classification using PVQA, MPVQA)算法。该算法在继承传统算法时间复杂度的基础上,引入马氏距离,克服了欧氏距离容易受模式特征量纲影响的缺点,提高了算法精度。首先,在训练时采用分段矢量量化近似方法获得码本,然后以马氏距离为相似性度量对时间序列进行分段重构。对重构后的时间序列,同样基于马氏距离为相似性度量进行判别。在4个时间序列数据集上进行的试验结果验证了所提方法在时间序列表示和分类上的优越性。

Time series classification using piecewise vector quantized approximation based on Mahalanobis distance

A Mahalanobis distance-based time series classification using PVQA(MPVQA)algorithm was developed. On the basis of inheriting the time complexity of the traditional algorithm and by exploiting Mahalanobis distance, the algorithm could easily overcome the default that the Euclidean distance was easily influenced by the mode characteristic dimension and improve the accuracy. PVQA was first used to generate a codebook using training samples, and then the Mahalanobis distance was taken as the measure of similarity and used to reconstruct time subsequences. For an unseen time series, the Mahalanobis distance was also adopted to find the most similar one to it. Experimental results on four time series datasets demonstrated that our method was more powerful to classify the time series.