Enhancing Completion Time Prediction Through Attribute Selection

Approaches have been proposed in process mining to predict the completion time of process instances. However, the accuracy levels of the prediction models depend on how useful the log attributes used to build such models are. A canonical subset of attributes can also offer a better understanding of the underlying process. We describe the application of two automatic attribute selection methods to build prediction models for completion time. The filter was used with ranking whereas the wrapper was used with hill-climbing and best-first techniques. Annotated transition systems were used as the prediction model. Compared to decision-making by human experts, only the automatic attribute selectors using wrappers performed better. The filter-based attribute selector presented the lowest performance on generalization capacity. The semantic reasonability of the selected attributes in each case was analyzed in a real-world incident management process.