Multi-View Multi-Label Learning Based on Improved Fusion Strategy

In multi-view multi-label classification task, each sample is described by features from multiple views and contains multiple semantic information. Previous methods established a separate classifier for each view and combined the prediction results and contribution weights of all classifiers to make the final prediction. However, these methods tended to overlook possible interactions among multiple views and did not consider the shared information among multiple views. Therefore, we propose Multi-view Multi-label Learning based on Improved Fusion Strategy (MMIFS). Firstly, we learn a shared subspace and utilize it as a supplementary view. Then we construct a separate classifier for each view and learn the corresponding contribution weights. We introduce digital labels instead of logical labels and maintain label co-occurrence dependency based on the smoothing assumption. Finally, we improve the performance of MMIFS by converting linear model to non-linear model. Based on extensive experiments with five datasets, MMIFS exhibits favorable performance and effectiveness.