基于加权偏最小二乘回归的软件成本估算方法

Software Cost Estimation Method Based on Weighted Partial Least Squares Regression

针对软件成本估算中影响因素较多且自变量间存在多重相关性的特点,提出一种基于加权偏最小二乘回归(PLSR)的软件成本估算方法。定义属性权重,得到描述软件历史数据相似度的加权相似离度。通过计算样本相似度自适应地为样本分配权值,采用遍历搜索的方式确定最优主成分及权值分配参数。实验结果表明,该方法的估算误差比多元线性回归方法减少73.61%,比全局PLSR方法减少32.34%。

For the excessiveness of operating factors and multi-correlation of variables in software cost estimation, this paer proposes a software cost estimation method based on Partial Least Squares Regression（PLSR）. After the variables are weighted, an integrated index named analogue deviation is defined to describe the approximation of data samples. Then an adaptive weight is assigned to sample according to the approximation, and the optimal partial least squares latent variables and weight parameters are calculated by traversal searching. Experimental results show that the prediction error is reduced by 73.61% and 32.34% than multiple linear regression and PLSR respectively.