Adaptive group sparse multi-view classification method based on mutual information

Multi-view classification has been widely utilized across numerous fields, such as image recognition. Existing methods typically integrate complementary information from multiple views by assigning specific weights to each view, thereby achieving multi-view fusion. However, these approaches generally impose identical penalties on all features, overlooking the varying importance of individual features within each view. Moreover, current multi-view learning approaches perform fusion at the view level, lacking effective utilization of intra-view information. To address these issues, we propose an adaptive group sparse multi-view classification method based on mutual information (AGSMC). In the proposed framework, features within each view are first grouped using class-label-specific mutual information. Subsequently, adaptive group weights and adaptive individual feature weights are derived to accurately reflect the significance of different groups and individual features for classification. Based on these adaptive weights, an adaptive group sparse penalty is formulated and integrated with the multi-view regression loss, thereby effectively promoting the roles of informative groups and discriminative features in multi-view data fusion. Additionally, a fast-converging iterative algorithm is developed to alternately optimize the proposed model efficiently. Extensive experimental evaluations demonstrate the stability and superior classification performance of the proposed AGSMC method compared to existing state-of-the-art techniques.