Transaction Data Analysis And Merchant Churn Prediction For Aggregated Payment Platforms

The aggregated payment platform collects a large amount of transaction data from offline stores every day,which is the basis for business analysis and the development of intelligent services on the platform.By analyzing and mining the massive merchant information and transaction data,the potential characteristics of merchants can be obtained,so that the personalized services to merchants can be provided and the relevant business of the platform can be optimized.In this thesis,we focus on the merchant churn prediction of aggregated payment platforms.The feature engineering methods are proposed for aggregated payment platforms from the business scenarios,which can help the business department understand the reasons for merchants' churn.And a new time series model Mf GRU(Modality-fusion GRU)is proposed for the problem of fusing static features with time series features,which improves the accuracy of the prediction results effectively.In addition,a new deep self-paced learning algorithm DSPL(Deep Self-paced Learning)is proposed for solving the class overlap problem in the data,which greatly improves the recall of churn merchants.The main contributions of this thesis are as follows:1.Proposed feature engineering methods for aggregated payment platforms:Most of the existing related works consider only the statistical features of transaction amounts,such as the maximum value,minimum value,average value,variance,etc.,in the feature extraction stage.These features reflect relatively limited information and are not closely related to business.In this work,three types of features are proposed from application scenarios.The merchant features describe the inherent attributes of merchants,the transaction features describe merchants' business models and operating conditions,and the risk control features reflect merchants' abnormal behaviors such as swiping and cashing.The features extracted in this work can help business departments better understand the reasons for merchants' churn.2.Proposed a multimodal time series model: Considering the time-series nature of transaction data,in this work,the time series model is used to model the data to improve the accuracy of prediction.In addition,considering that static features can provide information such as inherent attributes and business models of merchants,the static features and the time-series features should be merged to further improve the prediction accuracy.However,traditional time series models cannot directly input static features,and the common way is to combine static features with timeseries data or with the representations output from the model.But these methods are not efficient and do not allow static features and time-series features to fully interact.In this work,a new time series model Mf GRU is proposed,where the gating units of the model can input both time-series features and static features to fully exploit the information contained in both modes.Experimental results demonstrate that the proposed model yields better performance than other models.3.Proposed a Deep Self-paced Learning algorithm: The data in this work suffers from a serious class overlap problem,and it has been demonstrated that self-paced learning can deal with this problem.But traditional self-paced learning algorithms are difficult to apply to the training of deep learning models.In this work,a new self-paced learning algorithm DSPL is proposed,which can be combined with the training of deep learning models in the form of a ”plug-in”.DSPL can effectively deal with the class overlap problem in the data through a new self-paced sampling mechanism,and greatly improve the prediction accuracy of positive samples.The F1 score obtained by using the proposed method for merchant churn prediction is0.635,which is 31.5% higher than the method currently used by the business department.The data analysis method and churn prediction model in this thesis can effectively analyze the reasons and reduce the number of merchants churning.This is important for enterprises to reduce the operating costs,improve the competitiveness and maintain their long-term benefits.