小时间序列动态完全Bayesian集成分类器研究

发布时间：2019-01-02 15:30

【摘要】：提高连续属性小时间序列分类的可靠性重要且具有挑战性.由于小时间序列所蕴含的信息不充分和时间序列数据具有时序依赖性,使得优化分类器与数据的拟合程度非常困难,而且非时间序列数据分类器的许多成熟技术都不具有实用性.针对这种情况,本文采用动态完全Bayesian分类器来增加属性为类提供的信息量,以实现时序与非时序信息的融合,并将基于具有对角平滑参数矩阵的多元Gaussian核函数估计属性条件联合密度、平滑参数的区间划分、时序递进分类准确性标准、平滑参数配置树的构建和分类器选择与平均等相结合来建立小时间序列动态完全Bayesian集成分类器.使用宏观经济小时间序列数据集进行实验,实验的结果显示,经过优化的动态完全Bayesian集成分类器具有良好的分类准确性.
[Abstract]:It is important and challenging to improve the reliability of small time series classification of continuous attributes. Because the information contained in the small time series is not sufficient and the time series data are time-dependent, it is very difficult to optimize the fitting degree between the classifier and the data. Moreover, many mature techniques of non-time series data classifier are not practical. In this paper, a dynamic complete Bayesian classifier is used to increase the amount of information provided by attributes to realize the fusion of temporal and non-temporal information. Based on the multivariate Gaussian kernel function with diagonal smoothing parameter matrix, the conditional density is estimated, the interval partition of smoothing parameters and the accuracy standard of sequential progressive classification are proposed. The construction of smooth parameter collocation tree and the combination of classifier selection and average are combined to build a dynamic complete Bayesian integrated classifier for small time series. The experimental results with macroeconomic small time series data sets show that the optimized dynamic complete Bayesian integrated classifier has good classification accuracy.
【作者单位】：上海立信会计金融学院信息管理学院;上海立信会计金融学院统计与数学学院;
【基金】：国家自然科学基金(批准号:61272209) 上海市自然科学基金(批准号:15ZR1429700)资助项目
【分类号】：TP18

【相似文献】