Similarity based city data transfer framework in urban digitization.

Cross-city transfer learning aims to apply the knowledge and model from data-rich cities to data-poor cities to solve the cold start problem. Existing methods directly transfer the model constructed from developed cities to underdeveloped cities without considering the similarity between them, which leads to a potential transfer mismatch problem, and in turn, decreases the performance of transfer results. Meanwhile, existing transfer learning methods cannot effectively extract the time series features of the data, resulting in the inability to achieve adaptive positive migration across cities. To solve this problem, we propose a similarity-based cross-city transfer learning method named TransCSM, which embeds the urban similarity into an adaptation transfer learning framework to achieve desired data transfer. Specifically, we first constructed an urban similarity model, which utilizes the urban POI (Point Of Interest) data to group the cities with similar characteristics into the same cluster. Then, we build a feature extractor network, that uses convolution neural network (CNN) and Gated Recurrent Unit (GRU) to extract more representative features of time series data. Afterwards, we build an adaptation transfer learning framework to achieve data transfer within the same city cluster, which ensures the reliability of cross-city data transferring results. Finally, we evaluate our proposed method in many public POI datasets from Baidu Map API, and enormous results have demonstrated that our proposed method can achieve superior performance against state-of-the-art methods.