Kernel-Shapelets:基于卷积网络的特征子序列学习方法

Kernel-Shapelets:Approach to Learning Shapelets Based on CNN

在时序分类领域中,基于特征子序列的方法因能给出个体分类的局部依据,从而具有较强的可解释性,但计算十分耗时,且分类表现相较于深度学习方法并不占优。深度学习方法虽在分类表现上优于其他方法,但缺乏可解释性。为了提高基于特征子序列方法的分类表现,增强深度学习方法的可解释性,提出基于CNN网络提取特征子序列的方法Kernel-Shapelets。该方法通过提取CNN的卷积核对应权重并作进一步筛选,从而提取出特征子序列,利用全局最大池化层的输出找出CNN网络的判断依据,从而提高模型的可解释性。通过在UCR时序数据集上进行实验,Kernel-Shapelets方法的平均分类准确率为82%,相比基于特征子序列的最优基准模型提高了15.4%,证明了Kernel-Shapelets能够利用CNN网络的学习能力提取出更有效、更具有辨识性的特征子序列,提取出的特征子序列也提高了CNN网络的可解释性。

In the field of time series classification,Shapelets-based methods are interpretable as they can target the key parts for classifica⁃tion.However,high time consumption and poor performance compared with deep learning methods impedes its application.While deep learn⁃ing methods could achieve better performance,they are lack of interpretability.To improve the performance of Shapelets-based methods and enhance the interpretability of deep learning models,we proposed a new method called Kernel-Shapelets.Based on the weight corresponding to kernels of CNN and the output of Global Max Pooling Layer,it could generate shapelets and locates the key basis of CNN for classification.Through experiments on the UCR time-series dataset,the average classification accuracy of the Kernel-Shapelets is 82%,improving 15.4%compared to the optimal benchmark Shapelets-based model.The results demonstrate that Kernel-Shapelets can extract more effective and dis⁃criminative Shapelets by using the learning ability of the CNN network,while improving the interpretability of CNN.