Research On Dynamic Hybrid Recommendation Algorithm Based On Deep Learning And Review Mining

The recommendation system models user's interest preference based on user's behavior data,and filters effective information personalized.The classical collaborative filtering method is simple and effective,but it often faces the problems of data sparsity and cold start.The introduction of deep learning and recommendation system brings a new development trend to solve the above problems.Deep learning can extract deepseated nonlinear features of users and projects from massive data by building a deepseated nonlinear network structure,which has a strong ability to mine the essential features of data sets from samples.At present,the recommendation algorithm based on deep learning has achieved some research results to a certain extent,but there are still some problems: first,many deep learning algorithms only use the user's rating data to model,but ignore many different types of context information,such as comments,tags,geography,social and other information,which can not solve the cold start problem well Question.Second,in mining a user's or product's comment set,although RNN(deep learning model)can well retain word order information,the contribution of each word to comment feature extraction is regarded as equally important.In fact,if the contribution of each word can be distinguished,the recommendation performance can be further improved.Thirdly,the algorithm ignores the time characteristics of user interest changes.Based on the above problems,this paper proposes an improved algorithm,the specific work is as follows:(1)The traditional collaborative filtering algorithm,which only relies on the user rating data,has the problems of low accuracy of recommendation results and cold start.In this paper,a new dynamic hybrid recommendation algorithm is proposed,which integrates stacked denoising autoencoder into the collaborative filtering based on users,learns the deep-seated features of users,and merges with the similarity based on users' project attribute preferences.In the prediction generation stage,the time decay term is set,the access probability is predicted dynamically,and the user interest changes are updated in time,so as to improve the recommendation quality.The experimental results on movielens dataset show that the algorithm has better accuracy and recall than UBCF,AE and sdae-ia algorithms.(2)In view of the high sparsity of the user interest check-in matrix and the lack of the importance of the comment text.In this paper,Re Ge So model is proposed.First of all,we use the pre trained Bert model to replace the previous word embedding model,and directly send the comment information of interest points to the Bert model to get the implicit expression of comment information.We use the bidirectional LSTM and the attention mechanism to measure the contribution of each word to the comment text mining of interest points,so that the model can understand the semantics and emotions of the comment more accurately,and model the location interest Potential factor for the point.Secondly,on the basis of matrix decomposition model,the user's social information and geographical influence factors of location interest points are fused,and the multi-source heterogeneous data is integrated into the unified probability factor model to solve,and then the user's interest preferences are more accurately mined.Experimental results show that compared with the algorithm without using bidirectional LSTM attention mechanism network to process comment text information,the performance of this algorithm is significantly improved.There are 33 figures,7 tables and 88 references in this paper.