Multi-view Discriminant Analysis with Posterior Probability Graph Weighting

Multi-view Discriminant Analysis is a supervised multi-view learning method widely used in pattern recognition fields. When one view’s data gets affected, the labels may contain errors, making the label reliability of that view’s dataset questionable. To address the above issue, the Multi-view Discriminant Analysis with Posterior Probability Graph Weighting (MvDA-PPG) method is proposed for reliable multi-view learning tasks in unreliable labeling environments. This method optimizes the posterior probability matrix by introducing the sample posterior probability graph. Moreover, it reduces the impact of data from unreliably labeled views on the performance of multi-view learning by adjusting the weights of some views. Experimental results demonstrate that the KMvDA-PPG method helps reduce the influence of unreliable labels by applying posterior probability weighting to samples containing unreliable labeled views. This improves classification accuracy and robustness. The paper conducted experiments on three widely used datasets. Theoretical analysis and experimental results indicate that the improved KMvDA-PPG algorithm exhibits excellent classification recognition performance.