Adaptive collaborative fusion for multi-view semi-supervised classification

Multi-view semi-supervised classification is inherently a challenging task in multi-view learning due to the lack of label information. Existing methods generally suffer from insufficient data fusion, expensive computation cost in the solution procedure and fail in tackling unseen samples directly, intensively limiting their applicability and efficiency in real scenarios. To address these issues, we propose an adaptive collaborative fusion method, seeking for an appropriate representation and fusion for multi-view data. The main advantage of the proposed method is that it simultaneously fuses both multiple feature projections and similarity graphs to learn a joint projection subspace as well as a unified similarity graph that fully preserve the correlation and distinction among views. Meanwhile, our method can coalesce different views in an adaptive-weighting manner, making the learned subspace more discriminative and facilitating label propagation on the fused graph. Furthermore, an acceleration strategy has been designed to reduce the computational complexity, thereby making the proposed method scalable to relatively large-scale data. Finally, an alternating optimization has been adopted to solve the formulated objective function. Extensive experiments on synthetic and real-world datasets are conducted to demonstrate the effectiveness and superiority of our proposed method.