Research On User Portrait Based On Improved DeepFM

The rapid development of science and technology is urging the vigorous development of all walks of life.As the leader of society,the Internet is connected with all walks of life all the time.With the rapid improvement of transmission,infrastructure,and computing capabilities,the intelligent Internet was born.As an important interface between people and the Internet,mobile phones leave a large amount of behavioral data on it every day.These data have extremely rich dimensions and strong timeliness.In Internet business scenarios,how to efficiently predict user portraits is one of the core issues currently studied by major companies.Traditional user portraits are mainly based on manual labeling,which is extremely inefficient.Therefore,how to use algorithm models to mark user portraits in big data scenarios is one of the hottest research nowadays.However,under high-dimensional and sparse feature data,it is difficult to dig out the multi-layer relationship between features using traditional machine learning algorithms,and related features need to be constructed artificially with the help of business logic,which limits downstream Applications related to user portraits.In order to solve the above problems,this paper optimizes on the basis of deep learning and attention mechanism.First Wide & Deep learning learning frame,the frame of the research level order features how to learn the combination,and the relevant basic algorithm described;then sequence information present in the data were studied,with the Attention and BiLSTM binding sequence information for the user's pick;and finally an improved Deep FM algorithms,data combination of high-and low-level,to solve the sparse high-dimensional features selflearning problems.The main contents are as follows:(1)Research the Wide & Deep framework and its variants and use it as a base model.The overall idea of this model is to absorb the memory function of the shallow model and the generalization function of the deep model,and generate a joint model to improve the accuracy of the model and the scalability of the function.This model framework makes different modifications for different business needs,thus providing an optimization direction for subsequent model optimization.(2)Improve the ability of mining sequence information with Attention BiLSTM.Natural language processing has a strong ability to extract features from serialized data.This article uses Attention BiLSTM for serialized mining.The user history of behavior of elements as the words entered into the BiLSTM sequence learning,save each cell output and then through the Attention Network for learning,improve the extraction sequence information.Finally,the vector output by Attention is regarded as the Embedding feature of the user's historical behavior information.(3)An improved Deep FM algorithm model in the original FM added on the basis of Attention mechanism,the use of Attention to category features show features a combination of weights learning mechanism,so not only can improve the overall performance of the model,but also to improve the model 's available Explanatory.However,the original Deep component uses a residual network instead,so that the overall network enhances the feature combination ability without increasing the amount of parameters,so that the model can achieve higher accuracy.