Multi-Label And Multi-View Classifier Learning Based On Probabilistic Cooperative Representation

Multi-label classification is widely used in real life.Multi-label classification algorithms can be divided into two categories:(1)problem transformation,which transforms multi-label classification into two or multi-class classification,and then uses multi-class classification algorithm to classify.(2)Algorithm adaptation,which transforms multi-class classification algorithm into corresponding multi-label classification algorithm.Multi-label classification algorithms are all faced with three main problems:(1)how to effectively use the relevant information between labels;(2)data imbalance;(3)the effectiveness and efficiency of the algorithm for large-scale dataset.Multi-label classification algorithm based on problem transformation needs to learn multiple classifiers.When dealing with large-scale dataset,the algorithm runs slowly and needs large storage space.The model based on algorithm adaptation needs learn different high-order attribute for different labels,which requires high processing ability and is constantly very complex and easy to over-fit.In this paper,making use of problem transformation,a multi-label classifier based on probabilistic collaborative representation(Pro CRC-ML)is constructed,which has high classification accuracy,fast operation speed and high efficiency in processing large-scale dataset.In order to further improve its performance,an ensemble Pro CRC-ML(EPro CRC-ML)is designed.EPro CRC-ML has better classification performance and solves the problem of data imbalance to a certain extent.The boosting-based EPro CRC-ML has the process of feature selection,so EPro CRC-ML has better classification performance.Multi-view is quite more effective at improving the training of model than merely using single view.However,most existing multi-view learning algorithms only either pay attention to consistency or complementary among views,not making full use of multi-view dataset.Due to its high complexity,algorithm considering both complementarity and consistency has limited ability to process large-scale dataset.On the basis of Probabilistic Collaborative Representation based Classifier(Pro CRC),we propose Probabilistic Collaborative Representation based Classifier for Multi-View(Pro CRC-MV),which jointly maximizes the likelihood that a test example belongs to the co-subspace of each class.Learning subspace in the process of collaborative representation,considering consistency and complementarity concurrently,Pro CRC-MV can achieve promising classification performance.Meanwhile,it has low computational complexity,fast running speed,and can still maintain good performance when dealing with large-scale dataset.Pro CRC-MV has the ability for subspace learning based on self-representation,so we combain latent representation learning for better searching subspace with Pro CRC-MV to construct a novel classifier called LPro CRC-MV,the ability of LPro CRC-MV to process complex data is further enhanced comparing with Pro CRC-MV.