基于Shapelets的多变量D-S证据加权集成分类

集成学习是分类多变量时间序列的有效方法.然而集成学习对基分类器性能要求较高,基分类器组合算法优劣对分类效果影响较大.为此,提出一种基于Shapelets的多变量D-S(Dempster/Shafer)证据加权集成分类方法.首先,在单变量时间序列上学习得到基分类器Shapelets,基分类器的分类准确率确定为其在多分类器的权重.Shapelets是时间序列的子序列,不同变量Shapelets间不存在依赖关系,且单个Shapelets分类准确度较高,能得到“好而不同”的基分类器.然后,提出一种加权概率指派算法,增加分类准确率高的基分类器权重,减少分类准确率低的基分类器权重;添加了2个组合策略,即消除证据冲突,又提高了效率.在标准数据集上与多个最新算法进行比较,笔者算法取得了较好的分类结果.

Multivariate D-S Evidence Weighted Ensemble Classification Based on Shapelets

Ensemble learning is an effective method to classify multivariate time series. However, ensemble learning requires higher performance of the base classifier, and the combination of base classifier algorithms has a greater impact on the classification effect. This paper proposes a multivariate D-S(Dempster/ Shafer) evidence weighted ensemble classification method based on shapelets. First, learning the base classifier Shapelets on the univariate time series, the classification accuracy of the base classifier is determined as its weight in the multi-classifier. Shapelets are sub-sequences of time series. There is no dependency between different variable Shapelets, and the classification accuracy of a single Shapelets is high, and a “good but different” base classifier can be obtained. Then, a weighted probability assignment algorithm is proposed to increase the weight of the base classifier with high classification accuracy and reduce the weight of the base classifier with low classification accuracy. Two combination strategies are added to eliminate evidence conflicts and improve efficiency. Compared to some state-of-the-art algorithms on the standard dataset, our algorithm can obtain a better classification result.