Research On A Novel CTR Prediction Model Based On Deep Feature Fusion And Selection Network

The click-through rate prediction model can predict the click probability of users to advertising,and it is the center of advertising recommendation field.The existing CTR prediction models generally use the method of parallel combination of wide part and deep part to learn feature interaction,and deep neural network is usually used in the deep part.However,these CTR prediction models mainly have three problems.Firstly,the effect of deep neural network of deep part needs to be improved.Theoretically,the performance of deep neural network can be improved by using more neural layers in the DNN network or using more neurons in each layer,but it will cause problems such as gradient vanishing,gradient explosion and excessive parameters,and the performance of DNN will decrease rapidly after reaching saturation with the increase of its layers.Secondly,with further research,the wide part of these models have better choices.Thirdly,not all interactive features learned by the model are valid,and the importance of interactive features is inconsistent.If traditional methods are used to retain interactive features,the model will not be able to achieve better results.Based on the above three problems,this paper makes the following research contents and results:(1)In order to solve the problem of gradient disappearance caused by traditional methods to improve the performance of DNN,a Dense DNN network is proposed.Dense DNN network introduces the idea of Dense Net,takes the outputs of all previous layers as the input of current layer,which strengthens feature propagation,achieves feature fusion and alleviates the gradient disappearance caused by increasing the number of layers or neurons.(2)In order to select a better wide model to improve the overall effect of the model,the DDNNFMFM model is proposed in this paper.DDNNFMFM follows the structure of the current mainstream CTR prediction model,and adopts the parallel structure of deep part and wide part.The deep part uses the Dense DNN network proposed in this paper,and the wide part selects the mainstream Fm FM model.By simultaneously learning high-order and low-order feature interactions,the accuracy of the CTR prediction model is further improved.(3)In order to select interaction features,we propose an optimized TDDNNFMFM model.Based on DDNNFMFM model and the idea of three-way decision,an interaction feature selection layer is added between Fm FM model and output layer to strengthen important interaction features,eliminate redundant interaction features,and give weight to general features,so that the effect of the model is significantly improved.(4)Based on the classic advertising datasets Criteo and Avazu,a large number of comparative experiments are carried out.Firstly,the wide part Fm FM and the deep part Dense DNN are compared separately,and the accuracy of both are better than similar models.Secondly,Dense DNN is replaced by the DNN part of the mainstream model,and the results are better than the original model.Thirdly,DDNNFMFM is compared with similar models such as Deep FM,x Deep FM,and Deep Light to verify the effectiveness of DDNNFMFM proposed in this paper.Finally,the optimized TDDNNFMFM model is compared with DDNNFMFM,which shows that the interactive feature selection method based on the three-way decision idea can improve the prediction effect of the model.(5)The TDDNNFMFM model proposed in this paper is applied to the prototype system to realize the personalized recommendation function in the platform.The feasibility and effectiveness of the model proposed in this paper are proved.