Research On Urban Similarity And Development Trend Prediction Model Based On POI Data

In 2008,the concept of urban computing was put forward,hoping to use big data and artificial intelligence to sort out each stage of urban development,and effectively combine the virtual world connected by the Internet and the physical world in reality through data and models to establish a smart city.So as to provide people with a better,harmonious and convenient urban life.Therefore,in urban computing,using urban data to solve urban development planning problems,such as urban similarity and urban development trend,has become a research hotspot.Transfer learning and federated learning are two cutting-edge data processing and learning methods in urban data processing.With similarity as the core and similarity measurement criteria as the means,transfer learning can solve the problem of small and medium-sized sample training in urban computing.Federated learning can improve the training efficiency of the model in urban computing by using the multidimensional and high similarity of the data of all parties participating in the training.This paper puts forward new implementation methods on the application of urban migration learning,federal learning and urban similarity in urban development trend prediction.POI is "point of interest",a POI can be a house,a supermarket,a bus stop,etc.Firstly,this paper collects 19 dimensions of POI data of more than 60 cities in China through Baidu map developer platform,makes visual analysis and correlation calculation on the distribution and internal relationship of POI data of each city,and puts forward the urban similarity model CBCS based on clustering algorithm CKM.However,in the process of model establishment,due to the uneven urban development and the limitation of data privacy,it will hinder the establishment of effective model.Although many cities can collect massive data through a variety of data sources,there are still cities with extreme lack of data due to actual conditions.In this case,the neural network model can not be trained effectively,resulting in the cold start problem.Therefore,this paper proposes a Trans-UDTP model based on transfer learning to optimize the cold start problem to a certain extent.In addition,in order to train a better neural network model,we need not only massive data,but also the integration of multiparty data.In many scenarios,the data belong to sensitive private data,such as bank data,financial data and medical data,which can not be fully exposed,resulting in the problem of data Island.Therefore,this paper proposes a HFL-UDTP model based on Federated learning to protect data privacy,and uses multi-party urban data for joint training of the model.Both Trans-UDTP model and HFL-UDTP model use POI data to predict the urban development trend.Trans-UDTP model focuses on solving the problem of cold start,and HFL-UDTP model focuses on solving the problem of data island.At the same time,both of them are effectively combined with the CBCS model proposed above to further improve the prediction accuracy.On the one hand,CBCS model provides the basis for judging the transfer effectiveness of Trans-UDTP model,and improves the accuracy of the model after transfer.On the other hand,it optimizes the training effect of HFL-UDTP model.The experimental data prove the effectiveness of CBCS,Trans-UDTP and HFLUDTP models,and the prediction results can provide effective reference for urban decision-makers.