Towards Building Method Level Maintainability Models Based on Expert Evaluations

The maintainability of software systems is getting more and more attention both from researchers and industrial experts. This is due to its direct impact on development costs and reliability of the software. Many models exist for estimating maintainability by aggregating low level source code metrics. However, very few of them are able to predict the maintainability on method level; even fewer take subjective human opinions into consideration. In this paper we present a new approach to create method level maintainability prediction models based on human surveys using regression techniques. We performed three different surveys and compared the derived prediction models. Our regression models were built based on approximately 150 000 answers of 268 persons. These models were able to estimate the maintainability of methods with a 0.72 correlation and a 0.83 mean absolute error on a continuous [0,10].