Hotspots Areas Mining And Order Demand Prediction Based On Take-out Orders Data

The rapid expansion of Chinese online take-out industry shows huge potential for development.As a link between consumers,merchants and take-out platforms,delivery process still has some problems,such as the lack of response for orders and underutilized delivery staff,resulting in low efficiency and waste of human resources,which further affects the common interests of the three parties.Based on massive historical order data,researches on the law and change pattern mining of spatiotemporal distribution,the allocation and scheduling scheme optimization of delivery staff are urgently needed.Aiming at the contradiction between supply and demand caused by unreasonable distribution in a single region and untimely scheduling in multiple regions of delivery staff,the spatio-temporal distribution characteristics of take-out orders were used as the starting point for the related researches on hotspots areas mining and order demand prediction,which mainly includes the following three aspects:(1)Based on take-out order data,spatio-temporal distribution characteristics of orders were analyzed at two time-scales of one day and half an hour,and two spatialscales of administrative division and transportation analysis zone,combined with multisource data such as GDP and POI.Results show that the distribution of take-out orders changes periodically at intervals of one week and has two peaks in one day.In a single cycle,the peak heights and corresponding periods of weekdays and weekends are various,whose spatial distributions are significantly different.Take-out orders have gradually changed from mainly coming from working areas to residential and commercial areas.(2)To slove the problems of unstable clustering results and local optimal solutions of KMeans algorithm,IPSO-KMeans++ algorithm was proposed in order to accurately mine hotspots areas of take-out demand.The KMeans++ algorithm is used to initialize the position of the particle swarm to waken the influence on the clustering results caused by random selection of initial centers.What is more,the improved algorithm optimizes the updating strategy of particle’s velocity,learning factor and inertia weight in the iterative process for better ability of global search.Based on the improved algorithm,hotspots areas for take-out demand under different spatial and temporal distribution were mined.Different demand characteristics of multiple hotspots areas,non-hotspots areas and hotspots areas for part of the periods were discussed respectively to provide references for delivery staff allocation.(3)Feature sets affecting take-out demand under different spatio-temporal distribution were constructed through correlation analysis and an order demand prediction model based on adaptive deep neural network was designed.Besides,the number of neurons in hidden layers was adaptively adjusted for different feature sets and used the distributed computing framework Ray for the optimization of hyperparameters.The prediction experiments were carried out and evaluated by the quantitative indexes and comparative tests.The results show that the proposed model has outstanding performance in precision.Based on it,the quantity of take-out orders near the demand hotspots in different periods of the next day is predicted,which provides reasonable suggestions for the formulation of dispatching schemes and contributes to the construction of intelligent take-out platform.