Research Of Personalized Recommendation System Based On Depth BLSTM And Multi-criteria Candidates

With the popularity of intelligent device applications,various types of data information are exploding,than the recommendation system has become an important method to solve information overload.Nowadays,recommendation system has been widely used in e-commerce,big data,machine learning and other fields.How to obtain users' interest characteristics and imitate users' behavior patterns has always been its research focus.Up to now,recommendation system are mainly divided into three categories: content-based recommendation,collaborative filtering recommendation,and mixed recommendation.Content-based recommendations can accurately learn user characteristics,and it is highly explanatory,but it have cold start problems and novelty of recommended projects;collaborative filtering can effectively use other users or project information for recommendation,and alleviate novelty problems,and could face the challenges of sparsity problem and extensibility problems;The hybrid systems combine the advantages of each algorithm to improve the performance of the entire hybrid recommendation system.This paper proposes a multi-criteria fuzzy K-NearestNeighbor recommendation algorithm and a recommendation algorithm for deep bidirectional long short-term memory(LSTM)networks,aiming at the problem of data sparsity and few algorithms considering user ontology.Finally,the two recommended algorithms are combined to complement each other's advantages to reduce the impact of shallow model learning ability and cold start.The main work of this paper is as follows:1)A multi-criteria fuzzy K-NearestNeighbor(KNN)algorithm is proposed for data sparsity and without considering users' characteristics.That fuzzy mathematics and fuzzy similarity measure formula are used to process the users' rating matrix,is used to reduce the impact of user score ambiguity.The personal information similarity and user to user Jaccard sparsity are introduced to improve the comprehensive similarity measure formula between users.A list of proposed items is given using the KNN algorithm.Compared with other KNN algorithms,RMSE and MAE are reduced by about 1%-4%,and F value isincreased by about 1%-4.5%;2)A recommendation algorithm based on deep bidirectional long and short term memory(LSTM)network is proposed to solve the problem that traditional collaborative filtering cannot learn user characteristics and cannot effectively use the time factor to solve the dynamic change of user interest.The Recurrent neural network is introduced into the algorithm.The user history scoring behavior and interest cycle characteristics are learned from the two time sequences in order,and the deep network is used for superposition learning to build a rating prediction model that is more consistent with the user's rating habits.And adding the dropout strategy to solve the over-fitting problem to improve the prediction accuracy.Compared with other neural network models,the error is reduced by2%-5%,and r2 is increased by 5%-10%;3)Combine the multi-criteria fuzzy KNN algorithm with the deep bidirectional long short-term memory(LSTM)network,and use the latter to solve the problem of insufficient learning ability caused by the shallow model of the former,and use the former to solve the cold start problem of the latter.The multi-criteria fuzzy KNN algorithm is used to obtain the list of proposed items,to narrow the recommended search space,and the proposed items are predicted and scored by the deep bidirectional long short-term memory(LSTM)network to obtain the final TopN recommendation list.Finally,the comparison experiments are carried out from three perspectives: KNN improvement algorithm,shallow model recommendation algorithm and deep model recommendation algorithm.The experiment proves that the personalized recommendation algorithm proposed in this paper reduces the RMSE by about 2%-5%,the MAE by about 2%-5%,and the F value by about 2%-3.5%compared with other types of collaborative filtering algorithms.