Personalized Advertising Recommendation Based On Goods' Category

With the advent of the Internet Age,e-commerce has experienced an explosive growth and an increasing number of people tend to buy what they want through the network.Free from the space limitation of the traditional shelf,those B2 C e-commerce platforms such as Amazon,Tmall,JD can present numerous goods to their users on the Internet so as to meet the growing needs of different people.However,a mass of goods also mean difficult choices.In order to help the users quickly find what they want,help the sellers get their target users,and enhance the trading volume of the e-commerce platform,the recommender system has to be adopted.Only in this way can the users,on-line sellers and the B2 C company achieve a win-win situation.This thesis focuses on how to realize personalized advertising recommendations according to the needs of users.To solve the problem of large computing consumption and bad user experience in traditional recommender system,a personalized advertising recommender system based on goods' category is proposed.Through the mining of the users' click log,a model is constructed to predict what kind of goods the user are interested in recently,and the B2 C company can provide marketing services based on the users' preference so as to meet the marketing demands of different merchant in various scenarios.On the basis of the e-commerce data of Alibaba,the thesis carries out researches on the feature construction and the optimization of recommending model and finally builds up a preliminary model for the recommender system.Feature construction begins with the analysis of users' click log,and then the data are properly divided into the training set,test set and validation set.Next,we design and constructs a large number of features from the perspective of users,categories and user categories,which refines the connection between the user and the goods' category and reduces the information loss.Finally,with the aid of the Map Reduce parallel computing framework,numerous feature extraction codes are written,and the feature flow is extracted on the Hadoop pseudo-distributed cluster,which provides a good data base for the follow-up recommendation model construction and optimization.Model optimization firstly uses the XGBoost algorithm to build a model.For a betterperformance,the grid search and greedy strategy are integrated to work out a super-parameter optimization model.Then,in light of the excessive and redundant features,a feature selection algorithm based on frequent feature pairs is proposed in this thesis.By exploring the feature links of the decision tree from the root node to the leaf nodes,we obtains the frequent feature pairs and use them to measure the importance of features.In this way,the rapid reduction of dimension can be achieved without affecting the model's performance,which greatly reduces computation consumption of model training and prediction.In the end,a multi-model fusion scheme is made where the predictive values of the logistical regression model,extreme random trees model,random forests model,GBDT model and XGBoost model are used as features.Combined with the important features obtained by the feature selection algorithm,the XGBoost algorithm is applied into the construction of the composite model.Therefore,the classification performance will be improved on the basis of the former single model,and the generalization of the model will be enhanced greatly as well.