On October 07, Alexandre G. L. Fernandes defended his master’s thesis, which was supervised by Prof. Dr. Marcelo Fantinato and co-supervised by Prof. Dr. Sarajane Marques Peres. In this work, the presence of non-fitting cases in a testing dataset or online scenarios in a predictive task using process mining was observed and analyzed.
His thesis explored one specific scenario within the so-called Predictive Process Monitoring (PPM), which proposes mechanisms to anticipate and prevent undesirable situations. In the studied scenario, a process-aware remaining time prediction method that uses an Annotated Transition System (ATS) was applied to a real-world dataset and the occurrence of non-fitting cases was analyzed in terms of its impact on prediction. Preliminary results showed that the occurrence of non-fitting cases is heavily increased as new descriptive attributes are added to ATS, aiming to make it more precise. To address this issued, based on several studied alternatives, Alexandre proposed the use of similarity search techniques to tackle the non-fitting cases. The results obtained showed a significant improvement to the overall assertiveness of the prediction technique. Moreover, another contribution was the generation of categorical attributes from date-time variables, bringing a seasonality perspective to the model. The slides used in the defense session can be found here (in Portuguese).
The evaluating board of this work had the following researchers:
