新型ε-不敏感损失函数支持向量诱导回归算法及售后服务数据模型预测系统  

A New ε-insensitivity Function Support Vector Inductive Regression Algorithm and After-sales Service Data Model Forecast System

作　　者：罗泽举[1] 朱思铭[1] 

机构地区：[1]中山大学数学与计算机科学学院,广州510275

出　　处：《计算机科学》2005年第8期138-141,154,共5页Computer Science

基　　金：国家自然科学基金(No.10371135)

摘　　要：对含有噪声的数据序列根据预测置信度进行去噪处理,将训练集和测试集及预测数据共同作为训练向量集,以此建立新型支持向量诱导回归算法。本文利用该算法对实时售后服务的“千车故障数”进行了时间序列分析,并建立了新型的ε-不敏感损失函数小样本模型预测系统。预测显示误差小于5.3％的值占了总体的98.1％,其预测署信度达到0.983,与二次和Huber损失函数相比其MAPE值只有2.3％。用计算机模拟仿真单批次预测显,当时间参量t→+∞,“千车故障数”将收敛于定值74.0601,这和实际相当吻合,表明所建预测模型的有效性。文章最后还和传统神经网络模型作了比较,说明新型SVM机比神经网络处理小样本能力更强。To filter noises according to prediction confidence level in the sequence of data that contains noises, set the training set, prediction set and testing set as the training set, we set up a new support vector inductive regression algorithm. To analyze the real-time after-sales service data time sequence of “the number of thousand cars malfunction” by using this algorithm and set up a new support vector machines models forecast system based on small sample and ε-insensitivity function. The predict value whose error is less than 5.3% is 98.3% of total. Further more, the confidence level come to 0. 981. Compareed with quadratic loss function and Huber loss function, the MAPE is only 2.3%. From the computer analog simulation the single batch, we find that when time parameter t→＋∞, the “the number of thousand cars malfunction”will converge to fixed value 74. 0601, this is correspond with the reality and show the model is very availability. In the end of this paper, contrasted with neural network, the new SVM is superior to traditional neural network in the capacity for handling small sample.

关 键 词：诱导回归算法 售后服务 预测系统 回归算法 损失函数 数据模型 支持向量 诱导 敏感 时间序列分析 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论] S165.27[自动化与计算机技术—计算机科学与技术]