B1131
Title: A unite-and-conquer-based approach that improves weak classification results
Authors: Abdoulaye Diop - University of Versailles Paris Saclay (France) [presenting]
Nahid Emad - University of Versailles (France)
Thierry Winter - Atos-Evidian (France)
Abstract: In the field of machine learning, statistical classification methods are used as a solution to multiple problems. The main idea of these methods is to build discriminative models that create decision boundaries that separate the classes. However, depending on the properties of the data studied, these models may exhibit poor class prediction performance. This problem can be a consequence of statistical issues such as class imbalances, high bias, and high variance. We propose a framework based on ensemble learning techniques and a unite and conquer approach to deal effectively with these problems. This approach makes it possible to manage the bias-variance trade-off and improve the training time and the results of the base methods composing the ensemble learner. With the detection of behavioral anomalies as a case study, we show the interest of this approach for its improvement of the prediction results and its efficiency on high-performance computing systems.
