| Nowadays,in the Internet era,recommendation system is the research focus of various Internet companies.Facing the massive data waiting to be processed,the recommendation algorithm based on deep learning came into being.However,the model structure of most deep learning algorithms is based on the combination of multi-layer perceptron,and some problems of multi-layer perceptron itself are left in the recommendation system model.Firstly,the nonlinear changes that activation functions can simulate are limited,and some activation functions will have problems such as complex calculation and gradient disappearance in the process of back propagation.Secondly,nowadays,most recommendation algorithms based on deep learning rely on features to find the relationship with the prediction target.When some features are missing or the eigenvalues are wrong,it will affect the accuracy of prediction.Finally,based on the current data set,when calculating the relationship between multiple features and prediction targets,the recommended algorithm ignores the function change when the prediction target takes the features outside the data set or a single important feature in the data set as independent variables.It lacks the idea of considering the change of prediction objective function from the perspective of mathematics when training some regular data sets.In view of the above problems,this paper proposes to use function fitting to improve.In this way,we can mine the prediction target vectors that may have the function change law.A variety of mathematical functions are used to make up for the limitations of the activation function itself,and it is not necessary to extract all the features of the data set in the fitting process,which reduces the dependence of the algorithm on the data set.In this paper,the output probability of AFM model is used as the input of function fitting to fit a smooth and continuous mathematical function and mine the prediction target vector with regular changes.In the experiment,several recommendation system models based on deep learning are built,and compared according to their respective characteristics and evaluation indexes.Finally,the AFM model with attention mechanism and the Deep FM model with good experimental effect are selected for the construction of the new model.And this paper uses a variety of fitting functions to fit the target vector,and constantly tries to find the fitting function with better prediction effect.Finally,a new model composed of AFM model,fitting function and Deep FM model is constructed,and the following points are achieved.(1)The function change of the objective vector is simulated by a smooth everywhere derivable and everywhere continuous fitting function.(2)Consider the prediction of targets from the perspective of mathematics.The existing features can be used to mine new features with practical meaning,and the function change between the new features and the target vector can be processed by fitting.(3)The new model combines the advantages of AFM model and Deep FM model and function fitting.(4)For the data set used in the experiment,the improved algorithm proposed in this paper can play a better prediction effect.Compared with the improved Deep FM model,the AUC index of the improved algorithm is increased by 6%. |
