Research On The Recommendation Algorithm Of Fusion Of Multisource Heterogeneous Data Based On Deep Learning

With the progress of science and technology,various new technologies related to the Internet have developed rapidly.Various information on the Internet is becoming more and more complex,and has caused serious information overload.Through deep learning,a large number of heterogeneous data are integrated and deeply integrated with the recommendation system,which makes the recommendation model more suitable for the needs of users,thus improving the performance of the recommendation algorithm,which becomes a very important task of the current recommendation algorithm for fusion of deep learning.In practical application,because the interaction data between users and items is very sparse,the algorithm can not efficiently obtain the important features of new users and new items,which reduces the performance of the recommendation system.At present,it is an important solution to solve the above problems by integrating relevant auxiliary data,such as text description and label information of items,trust relationship between users,user label data and other user feedback information.The source and structure of these data are diverse.It is necessary to set up a reasonable mechanism in the recommendation algorithm in order to effectively integrate them,and ultimately improve the personalization and accuracy of the recommendation results.On the basis of summarizing the traditional machine learning recommendation algorithm and the existing deep learning methods which can be used in recommendation system,this paper establishes a deep learning model to solve the existing problems.(1)Asymmetric Depth Matrix Factorization Recommendation Model Based on User Trust Relationship(ADMFT).Based on the deep matrix factorization model,the model integrates the trust relationship between users and optimizes the model structure according to the differences in the number of users and items.A cross-entropy loss function with social regularization constraints is used for training.Experimental results show that the ADMFT model alleviates the limitations of the collaborative filtering algorithm such as data sparsity and scalability,and improves the accuracy of recommendations.(2)Aiming at the problem of data sparseness and cold start,a Matrix Factorization Model Combining Multi-sourse Information(MFCMI)is proposed that integrates multi-source heterogeneous information.The model combines the user's personal information and the auxiliary text information of the project.Firstly,the word vector and context information of the text information of the project are fused,and the sentences are fully expressed.Then,the convolution neural network with gating mechanism is used to extract the text features accurately and deeply,and MLP(MultiLayer Perceptron)is used to extract the features of personal information data.Then,according to the scoring data,the hidden features of items and users are learned through the matrix decomposition model,and the features of users and items are combined with the hidden features for recommendation.Experiments show that MFCMI model achieves the optimization of recommendation algorithm and the improvement of accuracy with the help of multi-source information fusion.