| As a branch of recommendation algorithm,collaborative filtering model mainly focuses on model-based collaborative filtering recommendation.Among them,the collaborative filtering recommendation based on matrix decomposition model has the most profound impact.Matrix decomposition model is one of the research hotspots in the field of recommendation algorithms.On the basis of the traditional matrix factorization model,in recent years,many improvers have been proposed to comprehensively use the scoring matrix and auxiliary information to establish a matrix factorization model for users and projects to improve the recommendation effect and mine effective recommendation basis.However,the auxiliary information is not the same type as the scoring matrix data and the scoring matrix data implies a lot of potential information.At present,there is less research on the type and association process of the underlying data of deep mining scoring matrices.Fewer models pay attention to matrix decomposition which will lead to data loss.In view of the above problems,this paper proposes a user-item alternating feedback decomposition model based on feature reconstruction.The model utilizes the idea of feature reconstruction to deeply dig out the explicit and implicit features and their relationships between users and projects,and compensates for matrix decomposition data loss problems.To realize the predictive ability of a matrix factorization model,it is necessary to learn the parameters of the model,and iterative learning methods based on the gradient descent method are generally used to update the parameters of the matrix factorization model.However,the gradient descent method has the problem of low iteration efficiency,which leads to too long a parameter learning time of the matrix factorization model.According to the nature of the gradient descent method,it is known that the gradient of the error curve gradually decreases,which will cause the convergence rate to slow down when the loss function approaches the minimum error,and the convergence rate becomes slower and slower when the error is getting smaller.Compared with the gradient descent method,the Newton method has a significant advantage in the convergence speed because of its second derivative characteristic.However,the calculation of the matrix parameters requires the calculation and storage of a Hessian matrix,which brings high algorithm complexity and storage requirements.According to the above analysis,this paper proposes a two-order fusion accelerated learning algorithm.The second-order partial derivative is used as the criterion to fuse the gradient descent with the Newton method,which improves the problem of slowing down the convergence rate of the gradient descent method,and also alleviates the matrix parameters calculated by the Newton method.The experimental results show that the feature mining ability of the alternating feedback matrix decomposition model is improved significantly,and the training time of the model is effectively reduced.In addition,it shows good results in the model update iteration and the stability of parameter changes.At the same time,the convergence speed of the two-stage fusion acceleration algorithm is significantly improved compared with the gradient descent method,and at the same time the calculation and storage overhead of the Newton method is reduced,which effectively improves the learning efficiency of the model parameters. |
