Research On Factorization Machines Click-Through Rate Prediction Algorithm Base On Deep Learning

In the information age,people are dealing with data every day,and people need to get useful information from massive data,so the recommendation algorithm comes into being.Click-through rate prediction is an important direction of ranking in the recommendation system.With the addition of deep learning,the field of click-through rate prediction has been pushed to a new level.The deep learning algorithm makes the machine show better ability in certain scenes and specific tasks.Factorization Machines is suitable for click-through rate prediction scenarios with sparse data.However,for the problems that it cannot be extended to higher-order feature combinations and the importance of features cannot be distinguished,depth learning technology is introduced to achieve more accurate and efficient prediction.This thesis mainly focuses on the impact of data sparsity on model prediction results in click-through rate prediction scenarios,and the learning of effective feature interaction in click-through rate prediction models:First of all,the problem of sparse feature data is alleviated by adding the new nonlinear feature.In chapter 3,the feature generation neural attentional factorization machines model is proposed.A separate scaling network is added for each feature field.By using the information on the original features,a large number of nonlinear features are generated from the original features through the expansion and shrink of dimensions and the transformation of nonlinear activation functions.The original features and new features are fused to participate in the subsequent feature interaction and prediction.At the same time,the FGNAFM model uses the attentional factorization machines model combined with a deep neural network to obtain the high-order feature combination of the model,which has a stronger interaction ability than the traditional factor decomposition machine.With the addition of an attention network,the model no longer needs to rely on artificial experience to choose useful features,which can automatically weaken the influence of redundant features and useless noise,and realize the interactive combination of effective features to a certain extent.Secondly,the scaling network is essentially a fully connected neural network.Compared with the traditional fully connected layer,the convolution neural network can obtain local feature information better.To some extent,convolution+pooling can enhance the feature information of the original data and reduce the introduction of some noise.Compared with a single attention mechanism,attention parameter learning in multiple spaces can obtain more comprehensive feature information and make the expression of the feature information richer.Therefore,an improvement is made based on the FGNAFM model in Chapter 3,and a feature generation dual attentional networks model is proposed.The FGDAN model does not ignore the role of deep neural networks in processing implicit high-order feature interactions.After the features pass through the dual attentional networks,they are further sent to the full connection layer to realize high-order feature combinations.Compared with the FGNAFM model proposed in Chapter 3,the FGDAN model has a better prediction effect.