Task-specific cross-city migration method

Abstract

The invention discloses a task-specific cross-city migration method in the technical field of smart cities and cross-city migration. The method comprises the following steps: iteratively screening a CTS region based on a reconstruction error of spatio-temporal data of a source city, calculating a matching degree between the source city and a target city region by using multi-source city auxiliary information, and screening out a region with the highest similarity with the CTS region of the source city in the target city according to the matching degree calculation. The method comprises the following steps: screening a target city CTS area, then complementing the target city by using the screened target city CTS area, verifying the effectiveness of the extracted CTS area in the aspect of capturing key information of the target city, and quantitatively verifying and proving the importance of the selected CTS area by using the whole process of data complementation by adopting a dynamic optimization method of depth matrix decomposition and random mask. Therefore, the most appropriate area is dynamically selected for different downstream tasks, the calculation complexity is effectively reduced while the effect is ensured, and a more visual explanation basis is provided for knowledge migration.

Description

Technical Field

[0001] The present invention relates to the field of smart cities and inter-city migration technologies, and in particular to a task-specific inter-city migration method. Background Art

[0002] Current cross-city transfer learning approaches typically employ two approaches: one is feature transfer, where the spatial-temporal data features of a source city are matched to the target city before being applied to the task; the other is fine-tuning, where a model trained in the source city is fine-tuned and applied to the target city. These approaches often assume that the source and target cities have similar spatial structures and data distributions when handling knowledge transfer between cities, thus ignoring the differences between cities. This is particularly true when faced with diverse downstream tasks, as traditional approaches struggle to tailor knowledge transfer specifically to the task requirements. This results in low transfer efficiency and may even introduce noisy data that is unfavorable to the target city, leading to poor performance.

[0003] 1. Computational overhead caused by global migration: Traditional methods rely on global data of the source city for migration, which not only leads to high computational overhead, but also many global features are not closely related to the target task and cannot provide effective assistance for task optimization.

[0004] 2. Lack of adaptability to specific tasks: Existing transfer learning methods have varying effectiveness across different downstream tasks, making them unable to transfer knowledge to specific areas based on task-specific requirements. Consequently, traditional transfer learning models struggle to provide efficient knowledge transfer solutions for complex cross-city tasks, resulting in reduced transfer effectiveness.

[0005] 3. Lack of explainability: Existing transfer learning methods often rely on global knowledge of the source city, making it difficult to explain why certain regional knowledge is useful for the target task and effectively distinguishing the importance of specific regions across different tasks. This flaw results in poor model explainability and an inability to effectively assess which knowledge is most valuable for the target city's tasks. Summary of the Invention

[0006] The purpose of this section is to summarize some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of this application to avoid blurring the purpose of this section, the abstract and the title of the invention, and such simplifications or omissions should not be used to limit the scope of the present invention.

[0007] Therefore, the purpose of this invention is to provide a task-specific cross-city migration method. By employing a dynamic optimization method based on deep matrix factorization and random masking, the method focuses on using data completion to quantitatively verify and prove the importance of selected CTS regions, thereby dynamically selecting the most appropriate regions for different downstream tasks. This technical approach, which uses "partial representation of the whole" rather than directly pursuing target data integrity, can effectively reduce computational complexity while ensuring effectiveness and provide a more intuitive explanation for knowledge transfer.

[0008] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions: A task-specific cross-city migration method, the steps are as follows: S1, iteratively screening CTS areas based on the reconstruction error of the source city spatiotemporal data; S2, using multi-source city auxiliary information to calculate the matching degree between the source city and the target city area; S3. Based on the above matching calculation, the area in the target city that is most similar to the CTS area of ​​the source city is screened out. Then, the target city is completed using the screened CTS area of ​​the target city to verify the effectiveness of the extracted CTS area in capturing the key information of the target city.

[0009] As a preferred solution of the task-specific cross-city migration method of the present invention, in step S1, the specific steps of iteratively selecting the CTS area based on the reconstruction error of the source city spatiotemporal data are as follows: Randomly mask the source city spatiotemporal matrix to generate missing data; Reconstruct missing data through deep matrix factorization and calculate regional reconstruction error; The mask matrix is ​​updated based on the reconstruction error to determine the CTS region.

[0010] As a preferred solution of the task-specific cross-city migration method described in the present invention, the random mask operation is implemented by the following formula: ; in, is a randomly generated value in the interval [0,1]. is the missing rate of the matrix.

[0011] As a preferred solution of the task-specific cross-city migration method described in the present invention, the deep matrix decomposition is achieved by minimizing the reconstruction loss, and the formula is as follows: ; in, To reconstruct the matrix, is the F-norm.

[0012] As a preferred solution of the task-specific cross-city migration method described in the present invention, the mask matrix is ​​updated by the following formula: ; in, and It is an adjustable parameter.

[0013] As a preferred solution of the task-specific cross-city migration method of the present invention, in step S2, the auxiliary information includes: POI 3D matrix , including 18 categories of interest point distribution features; Road network feature vector ,Connectivity and density structure features are extracted through graph convolutional networks.

[0014] As a preferred solution of the task-specific cross-city migration method of the present invention, in step S2, the specific steps of calculating the matching degree between the source city and the target city area using the multi-source city auxiliary information are as follows: Extract city auxiliary information to construct regional feature vector; The cosine similarity is used to calculate the matching degree between the source city CTS area and the target city area. The formula is as follows: ; in, and represents the POI feature vector of the CTS area in the source city, It represents the POI feature vector of all areas of the target city.

[0015] As a preferred embodiment of the task-specific cross-city migration method of the present invention, in step S3, the region with the highest similarity between a CTS region of the source city and the target city is selected based on the matching degree calculation in step S2 as follows: the target city regions are sorted according to the similarity, and the top k regions with the highest similarity are selected as candidate regions. The formula is as follows: .

[0016] As a preferred solution of the task-specific cross-city migration method of the present invention, in step S3, the data of the target city is completed as follows: ; in, is the missing matrix of the target city containing the CTS region, The function is a data completion function based on deep matrix decomposition technology.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. Compared to traditional meta-learning-based region selection methods, this method optimizes CTS region selection by simply iteratively updating a random mask matrix, eliminating the need for multi-level internal and external loop training. This significantly reduces computational complexity and training costs. This simplified selection process not only improves the system's overall operational efficiency but also makes it more suitable for large-scale urban applications, reducing implementation costs.

[0018] 2. This invention uses matrix decomposition technology to verify the effectiveness of the selected regions in the data completion process. Specifically, the reconstruction error in these regions is significantly lower than that in other regions, thus demonstrating their advantage in capturing key urban information and spatiotemporal characteristics. Based on this verification concept, although this embodiment uses matrix completion to verify the CTS region, similar dynamic region selection strategies can be adopted in other downstream tasks (such as traffic prediction and environmental monitoring). By replacing the data completion module with the prediction or analysis module of the corresponding task, the data can be nearly complete, alleviating the problem of data scarcity in the target city.

[0019] 3. By quantitatively calculating the reconstruction error, the present invention can intuitively verify and explain the importance of the selected CTS area in capturing the core functions and spatiotemporal characteristics of the city, and also makes the entire cross-city migration process highly interpretable.

[0020] In summary, the present invention effectively solves the problems of high computational overhead, low migration effect and poor interpretability in the existing technology through a low-complexity and high-efficiency dynamic CTS region selection strategy, and provides a knowledge transfer method with universal applicability and promotional application value, which can be widely used in data analysis and prediction of different downstream tasks in smart cities. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort. Among them: Figure 1 This is a flowchart of a task-specific cross-city migration method of the present invention. DETAILED DESCRIPTION

[0022] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0023] like Figure 1As shown in the figure, this paper proposes a task-specific cross-city transfer method. This method identifies specific city-task-specific regions (CTS regions) in a source city and transfers them to a target city. Based on a deep matrix factorization algorithm, this method dynamically selects characteristic regions in the source city and calculates the similarity between the source and target cities using auxiliary information such as points of interest (POIs) and road networks. This selective transfer effectively reduces computational overhead and improves the accuracy and efficiency of transfer learning, especially in data-scarce environments. The specific steps of this method are as follows: 1. Extraction and optimization of CTS region In the present invention, the space-time matrix of the source city represents the POI check-in data of people in their respective areas of the source city during their respective time periods. To achieve efficient task-specific knowledge transfer, we first perform regional segmentation on this space-time matrix, where each unit corresponds to a specific area in the source city.

[0024] (1) Definition of mask operation: In order to dynamically select the CTS region, we introduce the mask matrix , It is used to mask some elements in the source city matrix to simulate the scenario of missing data. The mask operation is defined as follows:

[0025] in, is a randomly generated value in the interval [0,1]. is the missing rate of the matrix. Through this operation, we can randomly select some areas as missing data, thereby dynamically selecting the data in the source city matrix.

[0026] (2) Deep data reconstruction After obtaining the partially missing source city matrix, we use matrix decomposition technology to deeply reconstruct it. Deep matrix decomposition is an effective matrix completion method, especially suitable for low-rank matrices, i.e., matrices with a rank of Less than the number of rows and columns. Through deep matrix decomposition, we transform the matrix of the source city into Decomposed into two low-rank matrices Q and Z, under the action of the mask matrix, the reconstruction matrix is ​​obtained , and calculate each area The reconstruction error is:

[0027] (3) Dynamic random mask update In order to adaptively screen areas that contribute more to downstream tasks, the following iterative update formula is designed:

[0028] The parameters and It can be adjusted according to the experiment. The meaning of this formula is: if the area The reconstruction loss If it is smaller, it means that the data in this area is easier to reconstruct and the features are more obvious, then the M value tends to 1 after the update; on the contrary, if If it is large, it tends to 0. After multiple iterations, the optimal mask matrix is ​​finally obtained. The corresponding area is determined as the CTS area of ​​the source city. This process realizes the dynamic selection of key areas.

[0029] It's important to note that the matrix completion technique used here is merely a validation tool, used to quantitatively measure the information capture capabilities of each region during the reconstruction process. Experimental results show that within the selected CTS regions, the reconstruction error is significantly lower than that of other regions, fully demonstrating their superiority in capturing key urban information and spatiotemporal characteristics.

[0030] Knowledge transfer process After successfully extracting the CTS region of the source city, we then perform knowledge transfer by calculating the similarity between the source city and the target city regions. The specific process is as follows: (1) Use of city auxiliary information In order to better measure the similarity between the source city and the target city, we use the point of interest (POI) data between cities. POI data is used to represent various points of interest in the city, such as buildings, parks, commercial areas, etc. In this invention, POI data is represented as a three-dimensional matrix , where the POI type dimension is 18, indicating 18 different types of points of interest.

[0031] In addition, if the dataset supports it, we can also use other city auxiliary information such as road network information, weather characteristics, etc. Here we use the road network as an example: we can use methods such as graph convolutional networks (GCN) to extract regional land network density, connectivity, average path length and other structural features to obtain the road network feature vector If there are multiple auxiliary information, we can also combine them into a unified region representation vector To perform the following similarity calculation.

[0032] Since the public dataset we use only contains POI data, we use Take calculation as an example, the auxiliary information of other cities is the same.

[0033] (2) Regional similarity calculation In order to measure the matching degree between the CTS area in the source city and the various areas in the target city in terms of city auxiliary information, we need to calculate the POI distribution similarity between the two, so as to obtain the most similar area in the target city, which is the CTS area of ​​the target city we are looking for. Therefore, we use cosine similarity to calculate the CTS area of ​​the source city. Any area in the target city The similarity between them is calculated as follows:

[0034] in and represents the POI feature vector of the CTS area in the source city, It represents the POI feature vector of all areas of the target city. This formula can quantitatively reflect the similarity between the two areas in the distribution of points of interest.

[0035] It should be pointed out that this method is also applicable to other auxiliary information (such as road network data, weather characteristics, etc.). If supported by the dataset, regional similarity calculation and dynamic region selection can also be performed through similar feature extraction and fusion methods.

[0036] (3) Select the CTS area of ​​the target city After obtaining the similarity between each region of the target city and the CTS region of the source city, we adopt a dynamic region selection strategy. Specifically, the region with the highest similarity between the target city and a CTS region of the source city is selected. This can be achieved through the following steps: a) For each source city CTS area , calculate its similarity with all areas of the target city b) Sort the target city areas by similarity and select the top k areas with the highest similarity as candidate areas, that is:

[0037] Data verification and task support Finally, we migrated the data of these selected CTS regions to the target city and supplemented the data of the target city using deep matrix decomposition technology to verify the effectiveness of the extracted CTS regions in capturing the key information of the target city. The specific formula is as follows:

[0038] in, It is the missing matrix of the target city in the CTS area, and the Completion function is a data completion function based on deep matrix decomposition technology.

[0039] It's important to note that this data completion module serves only as a validation tool, demonstrating that the CTS region can achieve near-complete data when reconstructing the target city's data. Therefore, in other downstream tasks (such as traffic prediction and environmental monitoring), this data completion module can be replaced with task-specific prediction or analysis modules, while the CTS region extraction and verification process remains applicable. This provides a unified and interpretable knowledge transfer method for various tasks, effectively alleviating the scarcity of target city data.

[0040] Although the present invention has been described above with reference to embodiments, various modifications may be made thereto and equivalent components may be substituted without departing from the scope of the present invention. In particular, as long as there are no structural conflicts, the various features of the embodiments disclosed herein may be combined with each other in any manner, and the omission of an exhaustive description of such combinations in this specification is solely for the sake of space and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A task-specific cross-city migration method, characterized in that: Here are the steps: S1, iteratively screening CTS areas based on the reconstruction error of the source city spatiotemporal data; S2, using multi-source city auxiliary information to calculate the matching degree between the source city and the target city area; S3. Based on the above matching calculation, the area in the target city that is most similar to the CTS area of ​​the source city is screened out. Then, the target city is completed using the screened CTS area of ​​the target city to verify the effectiveness of the extracted CTS area in capturing the key information of the target city.

2. A task-specific cross-city migration method according to claim 1, characterized in that: In step S1, the specific steps of iteratively selecting the CTS area based on the reconstruction error of the source city spatiotemporal data are as follows: Randomly mask the source city spatiotemporal matrix to generate missing data; Reconstruct missing data through deep matrix factorization and calculate regional reconstruction error; The mask matrix is ​​updated based on the reconstruction error to determine the CTS region.

3. The task-specific cross-city migration method according to claim 2, characterized in that: The random mask operation is implemented by the following formula: ； in, is a randomly generated value in the interval [0,1]. is the missing rate of the matrix.

4. The task-specific cross-city migration method according to claim 2, characterized in that: The deep matrix decomposition is achieved by minimizing the reconstruction loss, which is formulated as follows: ； in, To reconstruct the matrix, is the F-norm.

5. The task-specific cross-city migration method according to claim 2, characterized in that: The mask matrix is ​​updated by the following formula: ； in, and It is an adjustable parameter.

6. The task-specific cross-city migration method according to claim 1, characterized in that: In step S2, the auxiliary information includes: POI 3D matrix , including 18 categories of interest point distribution features; Road network feature vector ,Connectivity and density structure features are extracted through graph convolutional networks.

7. The task-specific cross-city migration method according to claim 1, characterized in that: In step S2, the specific steps of calculating the matching degree between the source city and the target city area using the multi-source city auxiliary information are as follows: Extract city auxiliary information to construct regional feature vector; The cosine similarity is used to calculate the matching degree between the source city CTS area and the target city area. The formula is as follows: ； in, and represents the POI feature vector of the CTS area in the source city, It represents the POI feature vector of all areas of the target city.

8. The task-specific cross-city migration method according to claim 7, characterized in that: In step S3, the region with the highest similarity between a CTS region of the source city and the target city is selected based on the matching degree calculation in step S2. Specifically, the target city regions are sorted according to the similarity, and the top k regions with the highest similarity are selected as candidate regions. The formula is as follows: 。 9. The task-specific cross-city migration method according to claim 1, characterized in that: In step S3, the data of the target city is completed as follows: ； in, is the missing matrix of the target city containing the CTS region, The function is a data completion function based on deep matrix decomposition technology.

Images