Robust Multi-view Multi- Modal Ensemble for Improved Image Classification with Feature Selection

various machine learning applications frequently incorporate multi-view data, which is characterized by distinct features representing different points of view or aspects. However, multi-view data can often include outliers, noisy and irrelevant features. The unique characteristics of multi-view data pose challenges in feature selection due to their complexity. To address this challenge, we propose a new supervised method called Robust Multi-view Multi-Modal Ensemble and Feature Selection (RMME-FS). With the aim of improving the performance of multi-view image classification, our approach simultaneously takes advantage of correlation preservation and complementarity between views to select relevant features for subsequent learning. Specifically, to combine multi-view information, features with higher weights are selected as model inputs. To this end, the least squares regression model with a Frobenius norm is used, accompanied by a penalty term l
2,1
-norm to assign a different optimal weight to each feature. Consequently, hierarchical feature selection is performed during the learning process. The F-norm penalty performs top-level feature selection to retain the essential information of the views, while the l
2,1
-norm penalty performs bottom-level feature selection to eliminate outliers and noisy features. Experimental results demonstrate the effectiveness of the method, showcasing superior accuracy compared to traditional classification approaches. This innovative ensemble methodology, enriched with feature selection, offers a promising avenue for improved image classification across diverse and challenging datasets.