计算机工程

• 人工智能及识别技术 • 上一篇    下一篇

一种联合的时序数据特征序列分类学习算法

史苇杭,林楠   

  1. (郑州大学 软件技术学院,郑州 450002)
  • 收稿日期:2015-05-18 出版日期:2016-06-15 发布日期:2016-06-15
  • 作者简介:史苇杭(1980-),女,讲师、硕士,主研方向为数据挖掘;林楠,副教授、硕士。
  • 基金项目:
    国家自然科学基金资助项目(61170306);国家科技型中小企业技术创新基金资助项目(10C26214102198)。

A Joint Classification Learning Algorithm for Feature Sequences of Time-series Data

SHI Weihang,LIN Nan   

  1. (College of Software Technology,Zhengzhou University,Zhengzhou 450002,China)
  • Received:2015-05-18 Online:2016-06-15 Published:2016-06-15

摘要: 针对时序数据特征学习中特征序列占用空间大和运算复杂度高的问题,提出一种联合学习特征序列和分类参数的分类算法。对时序数据进行特征变换后,采用线性分类器从最小距离矩阵中学习模型参数,以预测目标变量。在目标函数中,对分类预测的损失函数和分类器的线性权重进行联合学习,并利用随机梯度下降法求解优化问题。实验结果表明,与F-Stat和表达式变换方法相比,该算法在保持较少运算时间的前提下,具有较高的分类预测准确率。

关键词: 时序数据, 机器学习, 随机梯度下降法, 优化算法, 分类

Abstract: Aiming at the problems of monstrous space of feature sequences and high computing complexity in the process of feature learning for time-series data,this paper proposes a classification algorithm that jointly learns feature sequences and classification parameters.After feature transformation of the time-series data,it applies linear classifier to learn model parameters from the minimum distance matrix and predicts the target variables.In the objective function,it jointly learns the loss function and linear weights of the classifier in classification prediction,and applies stochastic gradient descent approach to solve the optimization problem.Experimental results show that,compared with F-Stat and expression transformation method,the proposed algorithm has higher classification prediction accuracy while keeping low execution time.

Key words: time-series data, machine learning, stochastic gradient descent approach, optimization algorithm, classification

中图分类号: