Modeling User Profile Model Based On Mobile Terminal Data

With the development of Internet technology and the rapid spread of personal mobile devices,the use of smartphones is gradually changing people's lifestyles.Users can not only use smartphones for basic operations such as phone calls and text messages,but also use apps for more intelligent operations such as social networking,entertainment and shopping.As a result,users generate a huge variety of personal history information on their smartphones,including gender,age,interests and other user behaviour information.These data provide us a good opportunity to explore the user attributes and to efficiently build a user profile,which is conducive to the realisation of sophisticated marketing and accurate recommendations.However,although conventional learning models are effective,most of them are based on the shallow learning and do not further explore the intrinsic links between data features,especially when facing with high-dimensional sparse features,the predictive effect still needs to be improved.To address the issues above,this thesis uses the Deep Neural Network method to predict user profile labels,uncovering the deep interconnection between multi-dimensional features from the perspective of feature learning,and selecting representative features at high latitudes to further improve the effectiveness of user profile label prediction,bringing further possibilities for improving personalised experience and precision marketing.Specifically,the research content and innovations in this thesis are as follows.(1)To improve the prediction results based on the original characteristics,a gain-ofinformation based approach is first proposed to measure the importance between user labels and APP lists.The aim is to assess which APPS are more important for distinguishing specific user labels so that we can select the most relevant APPS for constructing an effective APPbased user representation.Next,on the basis of the original features,two new types of features were reconstructed: user features based on APP categories and features based on user duration of use.(2)To further explore the intrinsic connection between the data features,a new method for predicting user label prediction is proposed by analying the traditional shallow learning model and deep neural network model,i.e.the combined model of deep neural network and LightGBM,in which the Deep Neural Network acts as a feature extractor and the gradientimprovement decision tree acts as a classifier,which is a new input of the gradientimprovement decision tree to predict the features of the Deep Neural Network after layers of processing until the final implicit layer,and to clarify the theoretical basis.(3)To verify the rationality of the proposed combined model,a single model experiment is conducted to compare the performance of the Deep Neural Network model and the traditional shallow learning model by selecting a dataset with high-dimensional sparse features,which verifies that the new model proposed in this thesis is composed of two models with good prediction performance.This thesis further explore the parameter sensitivity of Deep Neural Networks,including the number of layers of the neural network,dropout regularisation and convolutional kernels,to provide a basis for parameter optimisation as a feature extractor.(4)To verify the performance of the combined model,the trained Deep Neural Network model as well as the LightGBM model are used,and the two are combined for prediction to verify that they have better results than a single model and to prove the validity of the combined model proposed in this thesis.All in all,this thesis proposes a combined model algorithm for Deep Neural Networks and LightGBM on the basis of reconstructing the original features and combining knowledge from the fields of machine learning and deep learning,and provides theoretical proof and experimental verification of the rationality and performance of the algorithm.