An improved method to simplify software metric models constructed with incomplete data samples

Software metric models are useful in predicting the target software metric(s) for any future software project based on the project´s predictor metric(s). Obviously, the construction of such a model makes use of a data sample of such metrics from analogous past projects. However, incomplete data often appear in such data samples. Worse still, the necessity to include a particular continuous predictor metric or a particular category for a certain categorical predictor metric is most likely based on an experience-related intuition that the continuous predictor metric or the category matters to the target metric. However, in the presence of incomplete data, this intuition is traditionally not verifiable “retrospectively” after the model is constructed, leading to redundant continuous predictor metric(s) and/or excessive categorization for categorical predictor metrics. As an improvement of the author´s previous work to solve all these problems, this paper proposes a methodology incorporating the k-nearest neighbors (k-NN) multiple imputation method, kernel smoothing, Monte Carlo simulation, and stepwise regression. This paper documents this methodology and one experiment on it.