Predicting master's applicants performance using KDD techniques

The selection process of master's applicants has serious implications for the University. Students must contribute to science through scientific publications in congresses and journals, increasing the institution levels. However, predicting its performance demand advisor's experience and time. The advisors must select from a group of projects which one is the best for both sides. This procedure can lead to wrong choices. With the purpose of enhancing the selection process of masters' students, this paper presents an analysis of the academic characteristics of master's education applicants. It was used to identify set of features that might help to predict an applicant's scientific production. Through the process of Knowledge Discovery in Database, was obtained a set of correlation rules between features and the expected applicant's performance. Data collection was made on the Lattes Platform (A Brazilian platform to Scientific Curricula). Then, after pre-processed, it was optimized using the Partition Around Medoids with Gower distance. Experiments were conducted by training seven classification algorithms in two approaches. The first was a naive approach, making the analysis as simple as possible and the second one, develop an optimized process of DM. The optimization used the clustering process to create a new main class for the classifiers. Results demonstrate that the optimized approach was able to achieve an accuracy around 85% using the Random Forest, while the naive has approximately 44%. Through the PAM algorithm, it was possible to determine the best attributes and highlight the main rules. The conclusion is that the category of University and their previous production was the most relevant attributes. It shows that is possible to predict the productivity of master's degree students.