In the advertisement recommendation system,the prediction of the click-through rate(CTR)of advertisements is very important,and the feature crossing based on user behavior plays an important role in improving the prediction rate of the recommendation system.At present,most end-to-end neural network models convert high-dimensional sparse features to low-dimensional dense features through the embedding layer,and learn the cross-features of user behavior through the intersection of low-dimensional dense features,but this end-to-end model lack of migration.Therefore,this paper proposes a multi-fields embedding(MFE)model.The multi-fields embedding model is a pre-trained model that can simultaneously embed the sparse features of different domains into the same vector space.Then,we use the Gradient Boosting Decision Trees(GBDT)model to predict the click-through rate.By comparing with the GBDT model and the currently popular end-to-end neural network models FNN,NFM,Wide & Deep,Deep & Cross,DeepFM,and xDeepFM,gradient promotion decision trees model based on multi-fields embedding model was obtained the best results in two data sets.The results verify that the multi-fields embedding model can learn the better representation of high-dimensional sparse features belonging to different domains,so that the model we proposed can achieve good results on the task of click-through rate (CTR) prediction of advertisements. |