Hidden Semantic Space Learning Fusion Of Multi-view And Multi-label Information

With the development of sensor technology and the increase of storage devices,the data generated by individuals and enterprises has grown exponentially.Big data has penetrated into all aspects of life.Data from multiple sources of information such as text,images,audio and video,and web pages are also Increasingly diversified,multi-view multi-label data is one of the data types that exist widely.How to perform fusion learning on multi-view and multi-label information to improve the prediction ability of algorithms has become a research hotspot in the field of machine learning in recent years.Existing multi-view and multi-label learning methods are also facing the following challenges while making progress: How to learn the complementary information between them from multiple different perspectives,while taking into account the correlation between different labels is the first problem to be solved.Multi-view information and multi-label data often exist in a high-dimensional,sparse space.How to prevent data sparseness and difficulty in pattern recognition is another problem that multi-view multi-label learning needs to solve.In view of the above problems,based on the idea of orthogonal subspace reconstruction,this paper proposes supervised and unsupervised learning models that fuse multi-view information.The main work is as follows:(1)On the basis of multi-label implicit semantic index model and based on orthogonal subspace reconstruction,a supervised learning model combining multi-view multi-label information is studied.The main idea is to learn the internal relationship between them by projecting the input multi-view feature information and the output multi-label data into potential orthogonal subspaces,and at the same time assuming that different multi-view information and multi-label output data can It is associated with the shared projected orthogonal subspace through linear transformation.(2)Since the above model is only applicable to multi-view and multi-label information complete supervised scenarios,considering some unsupervised scenarios lacking multi-label information,this paper also studies a latent semantic orthogonal sub-component that only merges multi-view information Spatial learning model.The model obtains complementary information between different perspectives by mapping multi-view feature information to orthogonal shared subspaces.The shared subspaces are related to different multi-view information through a linear matrix.(3)The two models proposed above need to map the input feature information to the orthogonalsubspace.The coupling variables and orthogonal subspace constraints introduced thereby make the optimization of the model difficult.Therefore,this paper proposes a method to optimize the model by combining the Bregman iterative algorithm and the alternating direction multiplier method,and at the same time derives the analytical solution of the model.The comparison between multi-label data sets and unsupervised data sets shows that the proposed solution has a significant improvement in accuracy compared to other algorithms in the multi-view multi-label field.