Data-driven prediction of change propagation using Dependency Network

Change propagation is a central aspect of complex system developments. The prediction of change propagation is necessary to prevent further changes and to perform an assessment of the cost of planned changes. Bayesian Network has been applied to extract co-change patterns from the historical change log and to predict the probability of further changes caused by the change of other components. Due to the complexity of the Bayesian Network, however, its application to large scaled system can be limited. Also, Bayesian Network cannot represent the bi-directional relationship between system components. To address these limitations, this article proposes an alternative method using Dependency Network, which is an approximated version of the Bayesian Network. Detailed procedure for learning the DN from the data, as well as probabilistic inference algorithm using DN is explained. To show the feasibility of the model, a case study is conducted with empirical data obtained from the open-sourced software, Azureus. To validate the effectiveness of the proposed method, several experiments incorporating different parameters were conducted. The result confirms that our model can produce reliable and accurate estimation of change propagation probabilities.