Similarity based city data transfer framework in urban digitization.

Area of Science:

• Artificial Intelligence
• Data Science
• Urban Computing

Background:

• Cross-city transfer learning addresses the cold start problem by transferring models from data-rich to data-poor cities.
• Existing methods often fail due to transfer mismatch and inadequate time series feature extraction, hindering performance.
• The inability to adaptively migrate data across cities limits the effectiveness of current transfer learning approaches.

Purpose of the Study:

• To propose TransCSM, a similarity-based cross-city transfer learning method for effective data transfer.
• To embed urban similarity into an adaptive transfer learning framework to mitigate transfer mismatches.
• To enhance the extraction of time series features for improved cross-city learning.

Main Methods:

• Constructed an urban similarity model using Point Of Interest (POI) data to cluster cities with similar characteristics.
• Developed a feature extractor network employing Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU) for robust time series feature extraction.
• Implemented an adaptive transfer learning framework for data transfer within identified city clusters, ensuring reliable cross-city migration.

Main Results:

• The proposed TransCSM method demonstrated superior performance compared to state-of-the-art methods in cross-city transfer learning.
• Clustering cities based on urban similarity effectively reduced transfer mismatches.
• Enhanced time series feature extraction led to more accurate and adaptive data transfer.

Conclusions:

• TransCSM offers a significant advancement in cross-city transfer learning by incorporating urban similarity and adaptive feature extraction.
• The method provides a reliable framework for transferring knowledge between cities, particularly beneficial for data-poor urban areas.
• Empirical evaluations confirm the effectiveness and superiority of TransCSM over existing approaches for POI data analysis.