A Dynamic Gaussian Mixture Classification Method Based on Trusted Collaboration of Multiple Views

Abstract

This invention proposes a dynamic Gaussian mixture classification method for multi-view reliable collaboration, belonging to the field of semi-supervised multi-view classification. This invention trains a multi-view classification network consisting of a dynamic density-aware module, a consistency-specific representation reconstruction module, and a decision fusion classification module to perform classification prediction on multi-view data with a few labels. In the training phase, a KNN graph is first constructed for each view, and graph structure features are extracted using a view-specific Gaussian mixture model. Then, a learnable Gaussian mixture model is used in the feature space to estimate the sample-level view reliability, and the representations of each view are weighted accordingly to generate a sample-level consistency representation. Simultaneously, a shared mapping network is used to extract the specific and consistency representations of each view, and these are jointly input into the decoder for multi-view reconstruction. In the classification phase, a dual-branch classifier is constructed based on the specific and consistency representations respectively, and cross-entropy supervision is applied to labeled samples. Finally, the classification loss, reconstruction loss, density negative log-likelihood term, and covariance constraint term are weighted and summed to form the total loss function, and the entire network parameters are jointly optimized through backpropagation. During testing, multi-view data is input into the model to obtain multi-view features, and data classification prediction is performed. Compared with other methods, the present invention significantly improves the accuracy of classification results with a small amount of labeled data.

Images