Cross-City Crowd Flow Prediction Based On Transfer Learning

With the development of sensing technology and the maturity of large-scale computing environment,massive data in cities can be collected.Therefore,exploring urban laws through big data has attracted much attention in smart city construction at the present stage.However,the massive raw data in cities need to be processed before they can be used.Due to the imbalance of urban development,data have not been processed in most cities,and they still face data cold-start problem,making it difficult to explore urban laws with their own data.However,it is an indispensable part of modern city construction to extract useful information from data to interpret human mobility patterns and explore the deep knowledge of human movement.Therefore,in order to solve the data cold-start problem,researchers introduced the concept of transfer learning.Urban transfer learning aims to extract knowledge from source cities with rich data to help target cities with poor data predict the law of crowd flow.When using urban transfer learning to solve the crowd flow problem,the source city and target city with high similarity degree can be used as much as possible,which can effectively improve the final prediction result.However,at present,most studies are based on the assumption of sufficient similarity between cities and rely on empirical and experimental methods to select source cities.The former is usually empirically based on the level of urban development,which has proved unreliable;The latter trialand-error approach is a huge waste of time and resources.In the process of knowledge transfer,blind selection of source cities will even lead to negative transfer.At the same time,the costs of making the wrong decisions in real life are unbearable.Therefore,in view of the current situation of urban data cold-start problem,how to choose the appropriate source city to achieve cross-city crowd flow prediction has become an important research problem in urban transfer learning.In addition,in the related work of cross-city crowd flow prediction,the current prediction model is still inadequate in modeling.The work of this paper will focus on the above two problems.In urban space,the intention behind human mobility is closely related to the structural distribution of urban space.Inspired by this idea,we proposed a cross-city crowd flow prediction framework based on transfer learning,called Area Transfer.It includes two parts: source city selection and model training.In the framework,the source city selection part mainly uses the correlation between urban layout characteristics and human mobility to calculate the similarity between areas,and uses the KM algorithm to achieve the optimal match between cities,and constructs the evaluation algorithm of inter-city mobility to select the most suitable city for transfer.Feature extraction unit in model training part,through the dilated convolution and standard convolution method to extract the characteristics of urban crowd flow data,and SE attention module was added to improve the modeling ability of urban layout features.In the output unit,we use one-by-one convolutional structure module to reduce the number of parameters and pay attention to the same area information aggregation of different channels.In terms of the training process,firstly,the deep learning network model is pre-trained according to the sample data of the selected source city.Secondly,with the help of a binocular optimization function and a small number of target city data to optimize the model,the prediction model of the target city is finally obtained to complete the crowd flow prediction in scarce-data city.In this paper,the real dataset of Chinese cities are used to evaluate the proposed framework.The experimental results show that the proposed method can accurately and effectively select appropriate city for transfer knowledge,and has a large performance improvement of 9-13% compared with the traditional prediction model and advanced deep learning model.Therefore,the cross-city crowd flow prediction framework based on transfer learning proposed in this paper can evaluate the city transferability in a small number of alternative cities.Combined with the proposed network model,the crosscity crowd flow prediction problem under the background of scarce data can be effectively solved.