.png){kind=icon}
.png){kind=icon}
Return
Robust twin learning framework with adaptive loss
.png){kind=icon}
PRE
AI
DOI：10.1016/j.dsp.2026.106079.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Twin Extreme Learning Machine (TELM) classifiers have demonstrated remarkable success in various classification tasks. Nevertheless, the loss functions adopted in existing TELM models are relatively restricted and may not be sufficiently flexible to handle diverse types of data effectively. To overcome this limitation, this paper introduces a semi-smooth bi-parametric adaptive loss function. By adjusting its two parameters, the proposed function can be reduced to several well-established loss functions such as the hinge loss and pinball loss thereby enabling flexible and effective control over the impact of noise and outliers. In addition, compared with conventional loss functions, the proposed loss function possesses several desirable properties, including robustness, boundedness, and non-convexity. In light of these advantages, we incorporate this novel loss function into the TELM framework and develop robust twin learning architectures for both linear and nonlinear cases, aiming to improve the classification performance of TELM. To address the non-convexity introduced by the new loss function, the Concave Convex Procedure (CCCP) is adopted for model optimization, and its convergence behavior is theoretically analyzed. Finally, extensive numerical experiments are conducted on synthetic datasets, image datasets, and 9 benchmark tabular datasets. The experimental results validate that the proposed classification framework achieves superior performance across a range of classification tasks.
Keywords:
Twin extreme learning machine
Concave-convex procedure
Loss function
Robust classification
Lagrange function