Information Fusion Method Based On Symmetric Nonnegative Matrix Factorization And The Applied Study

The development of information technology makes the online data show the characteristics of big data,mainly manifested in large-scale and multi-source heterogeneous.These issues pose serious challenges to scientific management and calculation methods,urgently requiring us to integrate data from multiple sources to obtain more accurate and reliable knowledge than any single information source.In view of this,aiming at the deficiencies in the existing information fusion methods,this paper conducts some research on the information fusion model based on symmetric non-negative matrix factorization,symmetric non-negative matrix factorization fusion models considering graph regularization and prediction model based on symmetric non-negative matrix factorization,and first applied to cross-modal retrieval tasks.The specific work is as follows:1.For the assumption that there may be a "consistent" clustering pattern among multiple perspectives,an information fusion model(Multi-view SNMF)based on symmetric nonnegative matrix factorization is established.By normalizing the cluster indicator matrices obtained from different views,the distance between the "consistent"clustering pattern and these matrices,a constraint item,is introduced into the SNMF objective function.This solves the problem brought by assigning the same clustering model shared among different views.The existing methods donot consider the thought,namely,integrating this "consistency" assumption into symmetric non-negative matrix decomposition to model the fusion of information from distinct sources.The experimental results show that the proposed Multi-view SNMF model has better clustering performance in terms of accuracy and normalized mutual information.2.To solve the problems of low clustering accuracy and stability in information fusion,Laplacian regularized symmetric non-negative matrix factorization fusion model,called as LJ-SNMF,is proposed.In the LJ-SNMF model,the consistency of the manifold that is the cluster structures from different views tend to be consistent is maintained by constructing a robust Laplacian graph.Thus,the finally obtained cluster consensus matrix not only maintains the potential relationship existed in the original data,but also makes full use of the complementary and compatibility information carried by each view.Experimental results show that the introduction of graph regularization can significantly improve the performance of the SNMF algorithm.To some extent,the proposed LJ-SNMF model can also solve the problem of information incompatibility.3.Due to the fact that Laplacian graph regularization cannot used to infer unknown data effectively,Hessian regularization based symmetric non-negative matrix factorization fusion model is proposed:HJ-SNMF.This method uses the second-order information of the original data to describe the relationship among the samples,allowing the geodesic function to linearly infer the data,avoiding bias predicting unknown data against constants.The results show that HJ-SNMF has good performance in some tasks.These two fusion models above can be used to calssify the target accurately.They may play an important supporting role in some scenarios,such as information recommendation,user management and so on.4.In order to obtain categories and representation of real-time samples,a real-time sample prediction method based on "consistent" clustering model is proposed.When the representation of new sample in a certain modal is given,the prediction model can be used to determine the approximate representation of the sample in other modality;this can also be used to determine the category to which the sample belongs.This prediction method can incrementally identify real-time samples instead of reusing historical data,avoiding cumbersome iterative calculation processes,and greatly improving computational efficiency.This method will have great application prospects in the fields of text classification and information push.Finally,we study the application of SNMF in cross-modal information retrieval,such as searching texts with images and searching images with texts.This method combines multi-view clustering into the associations between different modalities,improves the performance of the cross-modal retrieval system to some extent.The validity and practical value of the proposed theoretical method are further verified.