Software Development Effort Prediction Based on Collaborative Filtering
Masateru Tsunoda, Naoki Ohsugi, Akito Monden, Kenichi Matsumoto, Shinichi Sato · 2005
††† To predict software development effort, this paper proposes an effort prediction method based on theCollaborative Filtering (CF ) which uses as input various software metrics recorded in past software development projects. The CF has an advantage that it can conduct a prediction using “defective” input data containing a large amount of missing values. There are, however, no researches which propose a method for applying the CF to Software effort prediction. Our proposal consists of three steps. In the first step, we normalize values of metrics to equalize their value range. In the next step, we compute the similarity between target (current) project and past (completed) project using normalized values. In the last step, we estimate the effort of target project by computing the weighted sum of efforts of high-similarity projects (that are similar to the target project) using the similarity of each project as a weight. In a case study to evaluate our method, we predicted the test process effort using 1,081 software projects including 14 metrics whose missing value rate is 60%, which have been recorded at NTT DATA Corporation. As a result, the accuracy of our method showed better performance than conventional methods (stepwise multiple regression models); and, the average accuracy per project was improved from 22.11 to 0.79.