| In order to solve the problems of "cold start" and "sparse score data" in existing collaborative filtering algorithms,some researchers have applied deep learning technology to collaborative filtering algorithms in recent years,and the accuracy of the algorithm has been greatly improved.However,the existing deep network recommendation methods are still faced with such problems as insufficient feature extraction in deep network,mostly single matrix,and insufficient use of the correlation between users and products,which seriously affect the accuracy of recommendation results.Based on this,this thesis carries out the following two works:Firstly,this thesis proposes a recommendation algorithm based on improved factorization machine model of SDAE.At present,deep learning is mainly applied to automatically learn deep representations from user and item information,but its ability to mine the potential deep relationship between users and items is ignored.At the same time,the design of user representation in existing studies is not accurate enough.At present,the existing research mainly focuses on the analysis of the user’s short-term behavior,but neglects the user’s long-term behavior in the past.Therefore,this thesis proposes an improved method for the traditional factorization machine model.In the traditional factorization machine model,the SDAE model is integrated.Firstly,the neural network is used to extract the hidden features of the object,and then the extracted feature vectors are input into the SDAE-FM model for training.It can better capture the linear relationship between users and items,strengthen the correlation between features,and finally obtain the final user preference results through the activation function.Finally,AUC and Recall are used as evaluation indicators to verify that the improved model is better than the traditional FM model.Secondly,a recommendation algorithm DSF based on deep learning is proposed.The neural network can learn the nonlinear relationship between users and items,and SDAE-FM can capture the linear relationship between users and items;therefore,the model can learn more information from the perspective of feature combination.Firstly,feature extraction is carried out through the neural network to obtain the feature vector,and the obtained feature vector is input into two different modules of the preference generation part,namely the SDAE-FM module and the neural network module,and then the results generated by the two sub-modules are input into the loss function at the same time for joint training.Finally,the obtained results are used as the final user preference prediction results.The experimental results show that the proposed model is higher than the basic model in AUC and Recall evaluation indicators,which proves that the proposed model can improve the quality of the recommendation algorithm,and this deep mixture is helpful to improve the generalization ability of the model. |
