«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

34

期数:

2024年08期

页码:

9-16

栏目:

大数据与云计算

出版日期:

2024-08-10

文章信息/Info

Title:

Weighted Deep Forest Algorithm Based on Dynamic Sliding Window

文章编号:

1673-629X(2024)08-0009-08

作者:

赵艺臻; 周立婵; 杨雨晴; 赵建军

太原科技大学 计算机科学与技术学院,山西 太原 030024

Author(s):

ZHAO Yi-zhen; ZHOU Li-chan; YANG Yu-qing; ZHAO Jian-jun

School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China

关键词:

时间序列分类; 深度森林; 窗口变化值; 迭代最优; 变化趋势

Keywords:

time-series classification; deep forest; window change value; iterative optimum; trends of change

分类号:

TP311

DOI:

10.20165/j.cnki.ISSN1673-629X.2024.0154

摘要:

深度森林是一种典型的机器学习方法,被广泛用于分类任务中。 但其在时间序列分类中,往往容易忽视时间序列变化趋势对其特征提取的积极作用;且在级联森林中的特征向量更新时,将各子分类器同等对待,使不同子模型的分类能力无法得到充分利用,最终使得时间序列分类陷入局部最优。 为了解决上述问题,该文提出了一种基于动态滑动窗口的加权深度森林方法,称为 AWGE-gcForest,用于时间序列数据的分类。 AWGE-gcForest 算法首先根据时间序列的变化趋势,定义了窗口变化值 WCV,实现窗口动态调整的同时减少多粒度扫描次数,以提高特征提取的效率、分类的准确率和泛化能力;其次,通过迭代最优对级联森林进行加权,为分类准确率高的森林赋予更大权重,从而降低分类性较弱的子树对整个模型的影响。 上述操作从全局考虑级联森林的分类性能,避免陷入局部最优,以减少级联层数并降低时间复杂度。该算法在 UCR 数据集上与 TS-CHIEF 算法、MultiRocket 算法、DF21 算法和 OS-CNN 算法进行对比,其分类精度以及时间效率优于目前先进的时间序列分类方法,是一种相对高效的时间序列分类算法。

Abstract:

Deep forest is a typical machine learning method that is widely used in classification tasks. However, in time series classification,it is often easy to ignore the positive effect of time series change trend on its feature extraction. In addition,when the feature vectors in the cascade forest are updated, each subclassifier is treated equally, so that the classification ability of different submodels cannot be fully utilized,and finally the time series classification falls into the local optimum. In order to solve the above problems,we propose a weighted deep forest method based on dynamic sliding window,called AWGE-gcForest,for the classification of time series data. Firstly,according to the change trend of the time series,the AWGE-gcForest algorithm defines the window change value(WCV),which realizes the dynamic adjustment of the window and reduces the number of multi-granularity scans,so as to improve the efficiency of feature extraction,the accuracy of classification and the generalization ability. Secondly,the cascading forests are weighted by iterative optimum,and the forests with high classification accuracy are given more weight,so as to reduce the influence of the subtrees with weak classification on the whole model. The above operations consider the classification performance of the cascaded forest globally,and avoid falling into the local optimum,so as to reduce the number of cascading layers and reduce the time complexity.Compared with the TS -CHIEF algorithm,the MultiRocket algorithm, the DF21 algorithm and the OS -CNN algorithm on the UCR dataset,the classification accuracy and time efficiency of the proposed algorithm are better than those of the current advanced time series classification methods,and it is a relatively efficient time series classification algorithm.

常用功能

更新日期/Last Update: 2024-08-10