自适应子序列聚类：单层递增神经网络方法与案例分析

需积分: 9 47 浏览量更新于2024-08-26 收藏 236KB PDF 举报

"本文提出了一种基于单层自组织增量神经网络(SOINN)的自适应子序列聚类方法，旨在解决时间自组织神经网络（如TKM和RSOM）在子序列聚类中产生的碎片化问题以及多变量时间序列处理中的稳定性问题。通过引入递归滤波器，模型能够更有效地对神经元激活的量化进行建模，从而提高聚类效果和网络稳定性。" 在时间序列分析领域，自适应子序列聚类是关键任务之一，用于发现数据中的模式和结构。传统的时空神经网络，如Temporal Kohonen Map (TKM) 和 Recurrent Self-Organizing Map (RSOM)，虽然因其在线学习和解释性能力而受到青睐，但在处理子序列聚类时存在一些挑战。它们可能产生大量难以分类的碎片，这在处理复杂的时间序列数据时可能导致聚类质量下降。为了解决这个问题，作者提出了一种新的方法，即基于SOINN的自适应子序列聚类。SOINN是一种单层自组织网络，它能够自适应地学习和更新其权重，以适应不断变化的数据流。在新方法中，他们引入了一个递归滤波器来模型神经元激活的量化，将每个变量的历史活动作为一个标量进行建模。这种方法的优点在于，它能够考虑不同变量之间的相互依赖性，从而改善了多变量时间序列的处理能力，提高了网络的稳定性和聚类性能。在实际应用中，递归滤波器能够捕捉时间序列中的动态变化，并通过连续的反馈机制来优化神经元的分配，使得子序列的聚类更加连续和连贯。此外，这种方法的自适应特性使其能够在数据流中实时调整，适应新的输入，这在动态环境下的时间序列分析中尤其重要。通过案例研究，作者展示了所提方法在真实世界数据集上的有效性。这些案例可能涉及金融市场的波动分析、生物医学信号处理或工业过程监控等场景，证明了该方法在应对复杂时间序列聚类任务时的优越性。这项研究为时间序列分析提供了一种新的工具，提高了子序列聚类的准确性和效率，对于理解和挖掘大量时间序列数据的内在结构具有重要意义。

A Temporal Self-Organizing Neural Network for

Adaptive Sub-sequence Clustering and Case Studies

Dong Wang

∗

, Yanfang Long

∗

, Zhu Xiao

∗

, Zhiyang Xiang

∗

and Wenjie Chen

†

∗

Colledge of Computer Science and Electronics Engineering,

Hunan University, Changsha, China

Email: {wangd,hndx

lyf,zhxiao,z xiang}@hnu.edu.cn

†

Business college, Central South University of Forestry and Technology

Email: wendychen711@126.com

Abstract—Temporal neural networks such as Temporal Koho-

nen Map (TKM) and Recurrent Self-Organizing Map (RSOM)

are popular for their incremental and explicit learning abilities.

However, for sub-sequence clustering TKM and RSOM may

generate many fragments whose classiﬁcation membership is

hard to decide. Besides they have stability issues in multivari-

ate time series processing because they model the historical

neuron activities on each variable independently. To overcome

the drawbacks, we propose an adaptive sub-sequence clustering

method based on single layered Self-Organizing Incremental

Neural Network (SOINN). A recurrent ﬁlter is proposed to model

the quantizations of neuron activations each as a scalar instead

of a vector like in TKM and RSOM. Then it is integrated

with the single layered SOINN for adaptive clustering where

fragmented clusters in TKM and RSOM is replaced by a

smoothed clustering result. Experiments are carried out on

two datasets, namely a trafﬁc ﬂow dataset from open Caltrans

performance measurement systems and a part of the KDD Cup

99 intrusion detection dataset. Experimental results show that

the proposed method outperforms the conventional methods by

21.3% and 9.1% on the two datasets respectively.

Index Terms—Recurrent neural network, sub-sequence clus-

tering, adaptive clustering, self-organizing incremental neural

network

I. INTRODUCTION

Sub-sequence clustering algorithms are effective time series

data analysis techniques. Applications such as atmosphere

engineering [1], trafﬁc ﬂow predictions [2] depends heavily

on time series sub-sequence analysis techniques. Visualization

of time series data enables visual analysis which is a unique

way to combine expert experiences and automated knowledge

discovery. Moreover, time series clustering is related to an

important set of problems in data mining, namely the data

stream mining. Time domain mining is crucial for concept

drift detection in data stream environments. However, despite

the large amount of research on spatial clustering, time series

clusterings are given far less attention.

Through clustering a generalization of the data is obtained,

usually in the form of clustering centers. More detailed anal-

ysis involves classiﬁcation of the clustering centers. Unfor-

tunately, the classiﬁcation of clustering centers are proved

expensive and time consuming even they are much smaller

in number than the original data. Adaptive clustering is an

effort that can minimize the number of clustering centers that

require expert knowledge to classify.

Adaptive clustering methods work under assumptions of the

data distribution, which is similar to semi-supervised learning

(SSL) [3]. In terms of SSL, data are said to be distributed

under the smooth assumption if the samples that are similar to

each other are more likely to be classiﬁed the same. Adaptive

learning based on self-organizing incremental neural network

(SOINN) [4] propagates labels from clustering centers to

similar ones so that only one of the neurons labeled the same

needs expert inspections. Hierarchical methods for adaptive

clustering such as [5], are in fact assuming clustering centers

distributing under the smooth assumption and low density

assumption. In the low density assumption data separated by

low density areas are considered belonging to different models.

Segmentation methods that split time series data into seg-

ments each with its own properties are closely related to time

series sub-sequence clustering. The sliding window solution is

widely used to transform time series patterns into a geometric

space, then the clustering tasks are carried out by conventional

clustering methods [6]. The learning is implicit and is different

from recurrent neural network’s explicit learning. In time

series analysis, the drawback of sliding window technique is

that the dimensionality of transformed data is increased. Such

dimension increase compromises both the performance and

efﬁciency of the conventional clustering methods.

Temporal Self-Organizing Map (SOM) including Tempo-

ral Kohonen Map (TKM) [7], Recurrent SOM (RSOM) [8]

and Recursive SOM [9] are time series clustering methods

without sliding window model. In TSOM a recurrent ﬁlter

is introduced to synthesize an input with the current sample

and its relationship with historical samples. RSOM always

functions as a solution of nonlinear prediction of time series

[8], [10]. Nevertheless, for sub-sequence clustering it produces

fragments too many and their inter similarity is difﬁcult to

measure. In other words, RSOM is ineffective in terms of

adaptive sub-sequence clustering. Except for sub-sequence

clustering, RSOM is the most widely used method for time se-

ries visualization [10]. Recursive growing neural gas (RGNG)

learns the topology in a self-organized manner so that the

limitation of a pre-deﬁned neighborhood function is RSOM

is remedied. However, RGNG inherits the endless growing of

neurons drawback of growing neural gas.

In this paper we propose a neural network approach for

下载后可阅读完整内容，剩余4页未读，立即下载

weixin_38694299

粉丝: 5
资源: 948

自适应子序列聚类：单层递增神经网络方法与案例分析

自适应k均值聚类

论文研究-基于自适应蚁群聚类的入侵检测.pdf

时间序列聚类——十年回顾

R Mclust函数 根据数据集自适应地确定聚类的数量 代码示例

在三维点云骨架提取过程中，如何综合运用改进自适应k均值聚类算法来提高细节保留和减少骨架连接错误？

自适应聚类和无监督聚类

如何理解和应用改进自适应k均值聚类算法于三维点云骨架提取中，以提高细节保留和减少骨架连接错误？

在三维点云骨架提取中，改进自适应k均值聚类算法如何帮助保留更多细节并减少骨架连接错误？请提供详细的应用实例。

matlab som自适应聚类

时间序列聚类分析文献综述

最新资源

R Mclust函数根据数据集自适应地确定聚类的数量代码示例