Research On Multi-source Data Fusion Method Based On Matrix Factorization

In the information age,the rational use and processing of the acquired data can produce great value.How to fully explore the multi-source data information and analyze it through data fusion to obtain decision-making is the current research hotspot.In this context,multi-source data fusion as an intelligent information processing technology plays an important role.This thesis studies the design and implementation of matrix factorization and fusion algorithm for multi-source data on the basis of the current mainstream matrix factorization algorithms.Moreover,data fusion is used to quantitatively infer the correlation between the two target objects.The main work is as follows:(1)Aiming at the problem that the prediction accuracy of the data fusion method is not high in the correlation prediction,this thesis proposes a matrix factorization algorithm based on similarity.Firstly,a multi-type relationship network is constructed,and the multiple relationships between data sources are decomposed factorized into low-rank matrices through a three-factor collaboration matrix.Then,the similar regular terms between the attributes of different types of objects are added to solve and optimize the objective function.Finally,the optimized low-rank matrix and similarity relationship are used to reconstruct the correlation matrix to improve the correlation prediction accuracy of the algorithm.The comparative experiments show that the algorithm has better performance in association prediction than the existing classical multi-source data fusion methods,and has certain improvement in accuracy.(2)In view of the lack of division of data source differences in fusion algorithms,this thesis proposes a fusion algorithm based on weighted matrix factorization.The algorithm introduces weights to the association networks of heterogeneous data sources and homogeneous data sources respectively,and uses the regular term to smooth the objective function to prevent overfitting.The low-rank matrices are jointly optimized,and finally the correlation matrix is reconstructed to predict potential correlations.The experimental results show that compared with other methods,the fusion algorithm based on weighted matrix factorization can improve the accuracy and reduce the error rate,which reflects the importance of distinguishing data source differences in data fusion.(3)In this thesis,the fusion algorithm based on weighted matrix factorization is applied to the field of biomedicine.Based on the multi-source gene-disease data,the relationship between three specific diseases and long non-coding RNA was mainly studied.The prediction results show good prediction ability compared with the database,which further confirms the effectiveness of the algorithm.