| In recent years,with the rapid rise of the Internet,o2o e-commerce model continues to develop,with the popularity of smart phones,penetrated into all areas of people's work and life.Among them,o2o takeout platform is also growing.In the early stage of o2o delivery platform,through burning money and subsidies,developed wildly,and obtained an objective number of users,and its market cake was also expanding.However,with the increase of users,the operation and resource allocation of the platform are also tense.Therefore,in order to improve the support capacity of the platform,in order to improve the support capacity of the platform,it is a relatively common method to predict the future consumer behavior according to a large number of existing historical data,so as to provide more accurate services for customers with different needs,and improve consumer satisfaction and service quality.It can also allocate platform resources more reasonably and improve the operation ability of the platform.Therefore,from this point of view,this paper uses the historical data accumulated by the platform to mine the characteristics of o2o order geographical regions,and tries to divide the order geographical regions into different sub regional spaces according to the distribution of customers' addresses.At the high end(e.g.in different regions at the low end).Refine the customer hierarchy.Then,the demand forecast is carried out from three aspects:geographical region level,sub region level and customer level.The hierarchical prediction algorithm is used to coordinate the prediction results to improve the accuracy and robustness of the prediction.So as to improve the service of each sub region.It can also fully mine the scheduling ability of the platform according to different levels of customer demand,and reduce the operation cost.Based on the takeaway order data of a certain platform,this paper studies the geographical characteristics of o2o orders,the regional division and the demand forecasting of customers at different levels.Starting from these two problems,the specific contents and conclusions of this paper are as follows:(1)The order area will be divided into the same size as the actual grid area through quantitative analysis.For the evaluation of grid size.In this paper,we use the order information to construct the basic features,and use the support vector regression(sqssvr)to fit the average order price in the grid area.The mesh size is evaluated by the fitting results,which is the basis of mesh adjustment.(2)On the basis of the determined mesh generation.Data mining is used to extract the mean value of grid orders.In this paper,the order geographical region under grid is studied.Through the density clustering of the grid,the o2o order geographical region is divided into several sub regions.On this basis,we use RFMD model and fuzzy c-means algorithm to cluster customers in each sub region.(3)From the above clustering,three levels of data can be obtained.They are the order data of the total region,the order data of the sub region,and the customer data of different levels under each sub region.According to the descriptive statistics,the relevant feature dimensions are constructed.Input sqssvr model to obtain the prediction results of different levels.At the same time,we also choose linear regression,support vector regression,neural network,and the mainstream ensemble learning random forest and xgboost for comparative experiments.(4)Based on the prediction of sqssvr,considering that the data of different levels of customers are noisy and unstable,the hierarchical forecasting method based on historical proportion from top to bottom is selected to coordinate the orders of different levels of customers,so as to improve the accuracy and stability of prediction.The main conclusions of this work include:(1)Based on the historical data,the grid size of o2o order geographical region is determined according to sqssvr fitting,the characteristics and rules of grid are obtained,and the relationship between customer level and demand region is established and explained.Experiments show that the average order price of grid area has a great relationship with latitude,longitude,user number,and grid location,that is,from the perspective of model,The number of users is related to the order price in the region.(2)Through the grid density clustering,the average price in the grid,the average number of orders per day,.By adjusting the EPS and epsion parameters,the whole order geographical area can be divided into several sub regions.It is not only geographically classified according to customers' real orders,but also changes the limitations of previous administrative division.(3)In terms of o2o takeout prediction,sqssvr is used to solve this problem.The nonlinear and robust characteristics of sqssvr make the prediction result better than the traditional machine learning model,and it has good performance compared with ensemble learning random forest and xgboost.In addition,from the perspective of different levels,the prediction effect of sqssvr is generally higher for the whole o2o order geographical area,which indicates that the prediction effect of the whole data is relatively high.However,with the level down,no matter which machine learning method,the prediction is unstable.It shows that the noise of the bottom data increases and affects the prediction.(4)In different levels of customer demand forecasting.Due to the increase of the fluctuation of the lowest level data of different customer demand,the prediction results of machine learning method are distorted.It is no longer of reference significance.Therefore,the historical average top-down hierarchical prediction method is used to modify the bottom prediction results.Some prediction results can be improved.Improve the effect of overall bottom prediction.Through the design of hierarchical forecasting method,we can ensure the integrity of customer demand forecast value at different levels,and improve their prediction quality through coordination. |
PDF Full Text Request