基于图正则化自编码回声状态网络的时间序列分类算法

Echo State Network with Graph Regularized Autoencoder for Time Series Classification

回声状态网络(Echo State Network,ESN)能够为解决时间序列问题提供有效的动态解决方法,然而大多数情况下ESN模型主要用于预测而不是分类,ESN在时间序列分类任务的应用尚未得到充分的研究。传统ESN由于存在随机生成的输入权重,使得其性能并不能保证最优。随机生成的权重在特征映射时,可能会破坏有用的特征。针对这些缺点,提出了一种针对时间序列分类任务的基于图正则化自编码的回声状态网络模型(GRAE-ESN),利用流形学习考虑数据内在的流形结构,来约束输出权重使得相似样本的输出在新的空间中更加接近,之后将ESN结构中的输入权重用解码层获得的权重所替换,以学习到丰富的输入特征。在基准数据上的实验表明,所提出的GRAE方法能够有效的改进ESN分类器,在与多个主流算法和深度学习算法相比,该算法具有更好的性能和鲁棒性。

Echo state networks(ESNs)can provide effective dynamic solutions for time series problems;however,in most cases,ESN models are mainly used for prediction rather than classification,and the application of ESNs in time series classification tasks has not been fully researched.Traditional ESNs suffer from randomly generated input weights,which do not guarantee optimal performance.These randomly generated weights can disrupt useful features during the feature mapping process.To address these drawbacks,a Graph Regularized Autoencoder based Echo State Network model(GRAE-ESN)for time series classification tasks is proposed.This model uses manifold learning to consider the intrinsic manifold structure of the data,constraining the output weights so that the outputs of similar samples are closer in the new space.Subsequently,the input weights in the ESN structure are replaced by with weights obtained from the decoding layer to learn richer input features.Experiments on benchmark data show that the proposed GRAE method effectively improves the ESN classifier.Compared to multiple mainstream algorithms and deep learning methods,this algorithm exhibits better performance and robustness.