【作者】 王迪；

【导师】 王萍；

【作者基本信息】 天津大学 ， 模式识别与智能系统， 2018， 博士

【摘要】 一致性预测作为一个集合预测框架,具有预测结果错误率可控这一优良特性,并受到诸多高风险低容错应用领域的青睐。然而,由于其计算框架的限制,原始的一致性预测计算速度慢,难以应用到需要实时处理数据的应用场合。本文从一致性预测的算法框架和底层算法入手,将近年来提出的刀切法一致性预测和具有快速学习能力的随机权网络相结合,并利用随机权网络可以快速计算训练集标签的留一交叉估计这一特性,提出了基于随机权网络、流形正则化随机权网络和递归随机权网络的一致性预测算法,同时证明了基于随机权网络、流形正则化随机权网络的一致性预测算法在大样本下的可靠性,并对算法的有效性进行了说明。本文算法在仿真数据集和实际数据集上进行了检验和应用。对于基于随机权网络的一致性预测算法,实验在经验上证明了算法在加快一致性预测学习速度的同时,保持了预测的可靠性和有效性,适用于高风险低容错且需要实时处理数据的应用场景中;对于基于流形正则化随机权网络的一致性预测算法,实验在经验上证明了算法的可靠性,且其中的一致性分类算法可以使预测结果更有效;对于基于递归随机权网络的一致性预测算法,实验在经验上证明了算法的可靠性,并展现出算法在混沌时间序列区间预测中的诸多优势。更多还原

【Abstract】 As a framework of a set prediction,conformal prediction is capable of controlling prediction error.It is welcome by high-risk application areas.However,the original computational framework of conformal prediction is too slow to be applied to a lot of real-time applications.This work improves the learning speed and ability of conformal prediction from employing a proper learning framework and a fast underlying algorithm.Specifically,this paper combines recently proposed jackknife conformal prediction and random weight networks which have fast learning ability and can compute leave-one-out predictions very fast.We propose conformal prediction with random weight network,with random weight network based on manifold regularizion and with recurrent random weight network respectively,and give some theoretical analysis of the convergence aspects of conformal predictions with random weight network and random weight network based on manifold regularizion and some comments on their efficiency.The algorithms are applied to both synthetic data sets and real-world data sets.For conformal prediction with random weight network,the experiments showed that the algorithms are very fast and have the property of validity and efficiency inherited from conformal prediction,which can be applied to high-risk applications.For conformal prediction with random weight network based on manifold regularization,the experiments showed that the algorithms are valid and the corresponding conformal classifier can improve efficiency of prediction.For conformal prediction with recurrent random weight network,the experiments showed validity of the algorithm and advantages of appling it to chaos time series interval prediction.更多还原