zoukankan      html  css  js  c++  java
  • PP: Triple-shapelet networks for time series classification

    Problem: time series classification

    shapelet-based method: two issues

    1. for multi-class imbalanced classification tasks, these methods will ignore the shapelets that can distinguish minority class from other classes.

    2. the shapelets are fixed after the training phase and cannot adapt to time series with deformation. 

    They propose a shapelet learning model: triple shapelet networks. 

    the imbalance of shapelets in minority class and majority class, to address this issue:

    they use category-level and sample-level shapelets to improve the performance. 

    classification is to find the best discriminating features. 

    Introduction:

    Shapelets are discriminative subsequences of time series data. They are suitable for TSC tasks since different classes often can be distinguished by their local patterns rather than their global structure.

    1. calculate the distances of shapelets and use these distances as discriminative features for classification. 

    shapelet transformation: find the top-k shapelets in a single pass. 

    to address two issues:

    1. imbalance features issue:

    they learn both types of features: dataset-level features and category-specific features. 

    2. deformation issue:

    Hence it would be useful to have shapelets that are specific to the data being processed. Here, it is reasonable to use a shapelet generator that is driven by the data itself to produce sample-specific shapelets.

    Three-types of shapelets: dataset-level; category-level; sample-specific level; use these three shapelets to conduct shapelet transformation and extract the discriminative features. 

    Thinking about:

    1. does this classification method is influenced by imbalanced datasets? and how?

    whether the method tends to ignore the feature of the minority categories? and only learns the features of majority categories?

  • 相关阅读:
    大数计算问题
    句子逆序
    字符个数统计
    提取不重复的整数
    合并表记录
    浅谈之高级查询over(partition by)
    proc之建表添加数据报错解决
    浅谈之过滤条件【or】
    浅谈之索引失效
    浅谈之表连接方法
  • 原文地址:https://www.cnblogs.com/dulun/p/12267407.html
Copyright © 2011-2022 走看看