基于离散小波变换的时间序列数据挖掘

提出了一种利用离散小波变换进行时间序列分析预测的新方法。该方法的特点主要是在小波系数的选取依据上与以往方法不同,以往方法大多是选取前k个位置的系数或者是选取数值最大的k个位置的系数,其依据是能量保持;本文方法的选取依据是各系数在训练集数据上的分类能力大小,即通过对已知类别的训练集的学习过程,找出使得类内距离最小、类间距离最大的若干系数作为特征系数。对于未知类别的时间序列,根据特征系数计算出该序列属于各个类别的隶属度,隶属度最高的类别即为预测结果。实验结果表明,本方法用于时间序列分析预测,显示出了较高的效率和准确性。

Time series data mining using discrete wavelet transform

A new method of time series data mining using discrete wavelet transform was proposed. The main property of this method lied in its feature extraction strategy, which was different from other methods before. Instead of using only the first k coefficients or the largest k coefficient, which emphasized on energy preservation, this new method decided extracted coefficients according to their classification ability on real time sequences in the training set. Generally speaking, this method tried to select coefficients from all wavelet coefficients to form a feature coefficient set, which best enlarged the distance between different classes and reduced the distance within same class. With this feature coefficient set, we can make prediction on test sequences set by calculating pertaining degree of each sequence to each class. The prediction result is the pertaining relation with largest pertaining degree. For real time series data used in our research, the efficiency and accuracy of this method is satisfying.