Research On Homepage Recommendation Alogrithm Of Ride Sharing Platforms

Shared travel platforms such as Uber and Didi play an increasingly important role in public travel today,not only because of their huge market share and service volume,but also because the services they provide are hierarchical and diverse and can adapt to the public.demand.Today there are more than ten types of services provided in the Didi APP.When the user opens the APP,the platform will recommend and display the service page of one services on the home page.The initial homepage recommendation strategy is the user 's “last service”.With the growth of platform products and the scenario-based usage habits of users,this solution has been unable to adapt to current needs.According to statistics,30% of users in the Didi platform need to switch the service tab four times before placing an order.In order to reduce user operation complexity and operation time,a new algorithm is urgently needed to more accurately predict user needs and recommend services to users.Combined with specific business scenarios,the platform can obtain the user's current spatio-temporal characteristics during prediction.Traditional machine learning models have achieved certain results in predictions based on spatiotemporal features.In this paper,a new recommendation algorithm is proposed based on the service selection of Didi platform users and the correlation of spatiotemporal features.The algorithm is generated by the integration of three sub-models.The sub-models are a time series model that uses the Markov state transition matrix to record the transition probability between different services in the user's historical order,uses Gaussian mixture distribution to fit the user,uses the product time distribution,and then passes Bayer The Si's formula calculates the time model of the conditional probability and the spatial model that uses decision trees to divide the latitude and longitude features.The three sub-models are integrated into an integrated model through bagging.The prediction stage makes predictions by obtaining the spatiotemporal characteristics of users opening the APP and uses the prediction results as the service page displayed on the home page.The implementation of the prediction algorithm is based on the Python language,mathematical function library numpy,scipy and machine learning framework sklearn.In the experimental stage,the historical travel data of 20,000 users in the Beijing area within the platform for single quarter were randomly selected as the training set,and the orders of the last week of these users were used as the test set for offline comparison testing.Use Precision and Marco-F1 scores as evaluation indicators to analyze the performance of the new algorithm and each benchmark algorithm.Experimental results show that compared with the original recommendation method,the new algorithm has achieved a significant improvement under both indicators.