Universum driven adaptive robust Adaboost twin extreme learning machine imbalance learning framework for pattern classification

Abstract

Pattern recognition and machine learning research has demonstrated that Universum, as a third class distinct from the positive and negative classes, can be integrated with prior knowledge to improve the generalization performance of a model. This approach enables the incorporation of prior knowledge into the learning process, facilitating the development of more accurate models. This paper proposes a novel learning framework, termed the Universum-driven adaptive robust AdaBoost twin extreme learning machine imbalance learning framework (ARATELM), for addressing class imbalance classification problems. In this framework, a new generalized smooth uncapped adaptive robust loss function called $L_{\theta }\left(u\right)$ is designed to improve the robustness of ARATELM. The generalized smooth uncapped adaptive robust loss function $L_{\theta }\left(u\right)$ aims to address the problems caused by the capped loss function $L_{\theta \varepsilon }$, which ignores normal data points and introduces non-differentiability. Concurrently, $L_{\delta }\left(u\right)$ effectively inherits the adaptability of $L_{\theta \varepsilon }$ through the utilization of the adaptive parameter, $\delta$, during the learning process. This enables the selection of diverse robust loss functions for different learning tasks, thereby enhancing the generalization performance of our method. Furthermore, the Universum data are taken into account in the proposed method, and prior information regarding the distribution of said data is provided; this enhances the generalization performance of the model. Additionally, the learning impact of our approach has been optimized through the integration of AdaBoost into ARATELM. Comprehensive experimental results across various class imbalance scenarios demonstrate that our presented method outperforms other methods in terms of robustness, classification accuracy, and other critical performance metrics.