Semi-supervised manifold regularized multi-task learning with privileged information

Multi-task learning (MTL) represents an advanced learning paradigm that improves the generalization ability and learning efficiency of a model by learning multiple related tasks simultaneously. The fundamental principle of multi-task learning is the transfer of information between tasks. Nevertheless where data is limited in quantity, effectively modeling inter-task correlations is a significant challenge. We propose a novel method, semi-supervised manifold regularized multi-task learning with privileged information (MSMTL-PI), that effectively leverages the intrinsic geometric structure of data by enforcing manifold regularization and subspace learning techniques. Specifically, a similarity graph is constructed over both labeled and unlabeled samples, ensuring the preservation of local geometric relationships between data points, and manifold regularization is applied as a constraint. Concurrently, information sharing on low-dimensional subspace makes the relationship modeling between tasks more reasonable. Furthermore, a significant amount of privileged information is incorporated into the training phase, thereby optimizing the decision boundary and reducing the impact of insufficient labeled samples on the model. There is substantial experimental evidence that MSMTL-PI markedly enhances the performance of image and text classification tasks, achieving superior classification accuracy with minimal labeled data. Across 15 benchmark datasets, MSMTL-PI consistently outperforms existing methods, achieving an average F1-scores improvement of 1.92% compared to the best baseline, with a maximum gain of 4.17%.