Research On Multi-source Information Fusion Recommendation Algorithm Based On Deep Learning

With the development of big data and cloud computing technologies,the emergence of various Internet applications has led to an exponential growth in network data.Large-scale data contains rich value,but it is becoming more and more difficult to find valuable information,which will lead to problems such as “information overload”.Since the recommendation system can deeply mine the interested information of users from large-scale data and recommend it to users according to user's preferences,it has become an effective strategy to overcome the problem of “information overload”.The recommendation system has been widely used in e-commerce and advertising computing,and has brought great commercial value.However,the recommendation system still faces many challenges,mainly reflected in the following aspects:Sparse data: In a business system with a large number of items,since the user usually only accesses a small number of items,this will result in a relatively sparse history containing user preferences.Therefore,how to solve the sparse data problem in the recommendation system based on massive data is a topic worth exploring.Cold start: Due to the lack of sufficient information to deeply extract user preferences,when a new user joins the recommendation system,it is difficult for the recommendation system to provide accurate recommendations for the user.Therefore,how to provide personalized recommendations to users who join the system is an urgent problem to be solved.Interpretability: The user usually expects the recommendation system to give an explanation to the prediction,rather than merely presenting the user with a "black box" recommendation.Therefore,how to improve the interpretability of the recommendation system is also an important challenge for the recommendation system.Recommended diversity: Since the user's interest preference is relatively wide,the recommendation system often recommends the same type of item to the user according to the user's historical activity record,and the homogeneous recommendation cannot satisfy the user's individual needs.Therefore,researching the diversity of recommendations can improve user satisfaction with the recommendation system.Many researchers have proposed a large number of solutions to above challenges.However,merely utilizing the user's rating information cannot essentially solve the inherent problems existing in the recommendation system.As more and more multi-source heterogeneous data(such as text,location,social relationships,images,etc.)can be obtained on the Internet,the integration of multi-source data information has become an important research field of the recommendation system.How to integrate multi-source information in the recommendation system,capture the complex internal relationships of the data,become an important research problem in the recommendation system field.Because deep learning can effectively capture non-linear and meaningful user-project relationships from accessible data sources such as context,text,and visual information,and make more complex abstract coding a higher level of data representation.Therefore,the multi-source data fusion recommendation algorithm based on deep learning has important theoretical significance and application value.Based on the National Natural Science Foundation of China,this paper conducts an in-depth study on the recommendation algorithms based on deep learning and multi-source information for the problems existing in the above existing research work.This includes the use of convolutional neural networks to fuse multi-source data and capture nonlinear user-item relationships,introducing attention mechanisms of deep learning to extract reviewing textual information,and mine strong association rules between items for recommendation.The main contributions of this paper are as follows:1.Points-of-interest recommendation algorithm based on convolution matrixfactorizationSince users only check in a few points of interest in the location social network,the user's check-in historical data are extremely sparse.Review text information can solve the sparse data problem of the recommendation system and gain a deeper understanding of user preferences.At present,researchers discuss location-based point recommendation based on review text information,but the existing method is to process the review text based on the word bag or document topic model,and only a shallow understanding of user's preferences.This paper uses the convolutional neural network to process the review text to model the latent factors of the points-of-interest.On the basis of the matrix factorization model,the user's social relationship and the geographic information of the points-of-interest are integrated,and the multi-source information is integrated into the same probability factor model,so that the user preference can be more meticulously modeled.The experimental results show that compared with the existing models that do not use the convolutional neural network to process the review text,the proposed model has achieved good results in accuracy and ranking prediction.2.Content-aware points-of-interest recommendation algorithm based on convolutionalneural networkThe above model only uses the review text information to model the latent factor of the points-of-interest.The user's latent factor is also obtained by using the traditional probability matrix factorization.To this end,based on the above,this paper uses convolutional neural networks to process review text information,modeling latent factor of points-of-interest and users.At the same time,the user's geographic location information and the obtained user's review sentiment category information are used to construct an objective function.The objective function is composed of matrix factorization and maximization of probability objective function.The experimental results show that the use of convolutional neural networks to process review texts can effectively model users and points-of-interest,and can obtain better recommendation results than other benchmark algorithms.3.A recommendation algorithm based on hierarchical attention mechanismAt present,many researchers propose to use deep neural networks to process user and product review text information to generate review text representations,thereby improving the performance of recommendations.However,some words or sentences in the reviews can strongly express user's preferences,while others tend to indicate the characteristics of the product.It is unreasonable to map user and product review information into the same feature representation.Therefore,this paper applies two independent two-way gated recurrent units to generate user and product representations,respectively.In order to make the feature representation of users and products reflect in the semantic layer rather than the word layer,this paper designs a word-level and sentence-level hierarchical representation structure.Because the attention mechanism can selectively filter out important information from a large amount of information and focus on important information,this paper introduces attention mechanism in each layer representation structure,ignores most unimportant information,and improves the performance of recommendation.Data analysis and experimental results show that the hierarchical attention mechanism can effectively model the review text information,improve the interpretability of the recommendation system,and can better capture user interest preferences compared with other benchmark algorithms.4.A recommendation algorithm based on utility association patterns miningIn real life,most users are only interested in several single categories of items.Recommending different categories of items with strong association patterns for users can improve the diversity of recommendations,but how to obtain the relationship between items is the first problem that needs to be solved in the recommendation task.In order to obtain the relationship between items,this paper combines the number and value of items in the transaction data and improves the utility association pattern mining algorithm UP-Growth.The clustering method is used to divide the similar transactions in the database into multiple data subsets and distribute the data subsets to each node of the distributed computing platform to construct a utility mode tree.The conditional pattern base of the same item in each node is assigned to the same node to mine the efficient association patterns,which provides support for subsequent recommendation based on the association patterns.Data analysis and experimental results show that the algorithm is superior to the comparison algorithm in terms of efficiency and recommendation diversity.Based on the recommendation accuracy,this paper studies three recommended algorithms from the four aspects of sparse data,cold start,interpretability and diversity.The first algorithm involves the contents of Chapters 2 and 3of this paper.Because the rating data is too sparse,and the recommendation system is prone to cold start problems without a large amount of rating data,the fusion of review text,geographic location,and user social relationship information can solve the above problem.The second algorithm involves the content of the 4th chapter of this paper.On the basis of solving the sparse data and cold start problem,the recommended interpretability problem is studied.Based on the deep neural network,a large amount of review text information is analyzed,and the feature content of the review is extracted.The attention mechanism finds important words and sentences in reviews as an explanation,increasing the interpretability of the recommendation.The third algorithm involves the content of the fifth chapter of this article,discussing the diversity of recommendations.By mining the strong association rules between different types of items,the association rule base of the item is established to provide users with diversification.In summary,this paper discusses the sparse data,cold start,interpretable and diversity issues in the recommendation system.In order to solve the above problems,the multi-source data is processed and integrated by using deep learning technology.Based on this,several recommendation algorithms are proposed,which comprehensively improves the performance of the recommendation.