考虑模型相关性的组合预测过程中单项模型筛选研究

发布时间：2018-12-20 17:01

【摘要】：针对组合预测过程中单项模型筛选难以刻画模型之间相关性的问题,采用模糊测度和模糊积分来刻画模型的相关性,并对单项模型预测结果进行集成,进而提出一种考虑模型相关性的组合预测过程中单项模型筛选方法.采用2可加模糊测度来刻画不同模型之间的相关性,并利用Choquet积分依据模糊测度值,将单项模型的预测值集成起来,形成组合预测结果.在这个组合预测过程中,采用基于模糊测度定义的Shapley值和交互作用指标来对单项模型进行筛选.为了验证文章提出的考虑模型相关性的组合预测单项模型筛选方法的有效性,选择软件工程领域的软件成本估算问题进行算例分析,选择基于案例推理方法(CBR)、最小二乘回归(OLS)、支持向量回归机(SVR)、分类回归树(CART)、人工神经网络(ANN)等数据驱动模型作为软件成本组合预测过程中的单项模型.选择常用的Desharnias数据库来验证模型的有效性.实证结果表明文章提出的单项模型筛选方法是一种有效方法,经过筛选后的组合预测模型能有效提高软件成本估算的精度,此外,研究结果还表明组合估算过程中最重要的模型(Sharply值最大)并不是估算精度最高的模型,即单个模型的重要性与该模型的估算精度没有必然联系,说明传统的以单个模型估算精度为依据的组合预测模型存在着一定的缺陷.
[Abstract]:In order to solve the problem that it is difficult to depict the correlation between models by single model selection in the process of combinatorial prediction, fuzzy measures and fuzzy integrals are used to depict the correlation of models, and the results of single model prediction are integrated. Furthermore, a single model selection method considering model correlation is proposed. Two additive fuzzy measures are used to describe the correlation between different models, and the Choquet integral is used to integrate the prediction values of single models according to the fuzzy measure values to form the combined prediction results. In the process of combination prediction, the Shapley value and the interaction index defined by fuzzy measure are used to screen the single model. In order to verify the validity of the single model selection method of combinatorial prediction considering model correlation, the software cost estimation problem in software engineering field is selected for example analysis, and the case-based reasoning method (CBR), is selected. The least square regression (OLS), support vector regression machine (SVR), classification regression tree (CART), artificial neural network (ANN) and other data-driven models are used as single models in the process of software cost combination prediction. Select the commonly used Desharnias database to verify the validity of the model. The empirical results show that the single model screening method proposed in this paper is an effective method, and the combined forecasting model after screening can effectively improve the accuracy of software cost estimation. The results also show that the most important model (maximum Sharply value) in the process of combined estimation is not the model with the highest estimation accuracy, that is, the importance of a single model is not necessarily related to the estimation accuracy of the model. It shows that the traditional combined prediction model based on the estimation accuracy of a single model has some defects.
【作者单位】：中国科学院科技政策与管理科学研究所;
【基金】：国家自然科学基金(71201156,71571179,71425002) 中国科学院青年创新促进会(2013112)资助课题
【分类号】：O159

【相似文献】