Multi-View Visual Classification via a Mixed-Norm Regularizer

In data mining and machine learning, we often represent instances by multiple views for better descriptions and effective learning. However, such comprehensive representations can introduce redundancy and noise. Learning with these multi-view data without any preprocessing may affect the effectiveness of visual classification. In this paper, we propose a novel mixed-norm joint sparse learning model to effectively eliminate the negative effect of redundant views and noisy attributes (or dimensions) for multi-view multi-label (MVML) classification. In particular, a mixed-norm regularizer, integrating a Frobenius norm and an ℓ

2,1

-norm, is embedded into the framework of joint sparse learning to achieve the design goals, which include selecting significant views, preserving the intrinsic view structure and removing noisy attributes from the selected views. Moreover, we devise an iterative algorithm to solve the derived objective function of the proposed mixed-norm joint sparse learning model. We theoretically prove that the objective function converges to its global optimum via the algorithm. Experimental results on challenging real-life datasets show the superiority of the proposed learning model over state-of-the-art methods.