User Purchase Prediction Based On Feature Learning And Ensemble Algorithm

With the continuous improvement of e-commerce platform and online trading system and the rapid development of logistics industry,more and more users choose to buy goods online.Due to the lack of face-to-face communication between the shopping platform and users,the platform cannot fully understand the needs of users,so that it cannot well get the recent purchase intention of users.However,the huge amount of user behavior data accumulated by e-commerce platform makes it possible to predict users' future purchase behavior.In recent years,with the rapid development of ensemble learning,various ensemble models are widely used in the user purchase prediction competition held by various platforms at home and abroad,and have achieved excellent performance.In the field of recommend area,the method of commodity recommendation based on user purchase prediction becomes more and more popular.However,most of the user's purchase prediction is based on the commodity as the prediction target.With the increasing variety of commodities and the exponential growth of the number of platform commodities,the problem of user's purchase prediction is faced with unstable prediction and the decline of precision.Based on the real data set provided by the recommendation algorithm competition held by JD.com in 2019,this dissertation selects larger granularity commodity categories and shops as the target of user purchase prediction,and studies the user's purchase intention of commodity categories and shops in the next seven days.The main work is as follows:(1)This dissertation has deeply analyzed the real data set provided by JD platform on the user purchasing behavior and finds out the deep factors that affect the purchase of users.(2)This dissertation proposes a feature learning method based on Word2Vec model and a feature dimension reduction method based on LSTM self-coding model.These two methods have powerful feature extraction ability,which are mainly used to learn the time series generated by users in continuous time,and automatically extract the deep information in the time series through the learning of the model.In addition,this dissertation also constructs the basic features and interaction features from the perspective of users and commodity categories(shops).(3)This dissertation proposes a feature selection method based on Pearson coefficient and CatBoost model.The Pearson coefficient describes the degree of linear correlation between features,and the feature importance returned by the CatBoost model indicates the help of the feature to model training.The feature selection algorithm proposed in this dissertation fully measures the linear correlation between features and the feature importance score returned by the CatBoost model,and obtains the optimal feature subset for model training.(4)Aiming at the problem of user category purchase prediction,this dissertation proposes a two-layer ensemble model.The two-layer ensemble model uses double-layer XGBoost model,CatBoost model and logistic regression model as the base classifier of the first layer.The second layer uses the linear model-based XGBoost model to fit first layer's prediction results to learn the final user category purchase probability.This model can effectively alleviate the problem of low precision of individual model and homogeneous ensemble model.(5)Based on the categories list that user will purchase in the future,this dissertation further studies which shop under the target category users will place orders,and proposes a deep ensemble model based on ensemble algorithms.The proposed deep ensemble model draws on the idea of deep neural network,by constructing a multi-layer learner,the input features are learned layer by layer to get the probability of future users placing orders in the shop.Many experiments have verified the effectiveness of the deep ensemble model to solve the problem of user shop orders.