Explanatable task-specific cross-city transfer learning method

Abstract

The invention discloses an interpretable task-specific cross-city migration learning method in the technical field of smart cities and cross-city migration, and the method comprises the following steps: dynamically extracting a CTS region of a source city based on a depth matrix decomposition algorithm, calculating the similarity between the source city and a target city through employing the auxiliary information of city POI and road network data, and carrying out the task-specific cross-city migration learning. According to the cross-city knowledge migration method based on the depth matrix decomposition technology, the problems that an existing cross-city knowledge migration technology is large in calculation expenditure, low in migration effect, poor in interpretability and the like are effectively solved, and the cross-city knowledge migration method based on the depth matrix decomposition technology based on the depth matrix decomposition technology has the advantages that the calculation cost is high, the migration effect is low, and the interpretability is poor. The calculation complexity is obviously reduced, and the adaptability to the target task and the migration accuracy are improved. Besides, which areas are beneficial to the target city can be clearly identified, and efficient data completion is realized under the condition of data scarcity.

Description

Technical Field

[0001] The present invention relates to the field of smart cities and cross-city migration technology, and in particular to an explainable task-specific cross-city migration learning method. Background Art

[0002] In current cross-city transfer learning, two methods are usually used: one is to match the spatial-temporal data features of the source city to the target city through feature transfer methods, and then apply the task; the other is to fine-tune the model and apply the model trained in the source city to the target city through fine-tuning. When dealing with knowledge transfer between cities, these methods often assume that the source city and the target city have similar spatial structure and data distribution, thereby ignoring the differences between different cities. Especially when facing different downstream tasks, traditional methods find it difficult to perform specific knowledge transfer based on task requirements, resulting in low migration efficiency, and may even introduce noise data that is unfavorable to the target city, resulting in poor performance. Specifically: 1. Computational overhead caused by global migration: Traditional methods rely on the 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 help for task optimization.

[0003] 2. Lack of adaptability to specific tasks: The effects of existing transfer learning methods on different downstream tasks are different, and it is impossible to transfer knowledge in specific areas according to the needs of specific tasks. Therefore, when dealing with complex cross-city tasks, traditional transfer learning models are difficult to provide efficient knowledge transfer solutions, which easily leads to a decrease in the transfer effect.

[0004] 3. Lack of explainability: Existing transfer learning methods mostly rely on the global knowledge of the source city, and it is difficult to explain why the knowledge of some regions is useful for the target task, and it is impossible to effectively distinguish the importance of specific regions between different tasks. This defect leads to poor interpretability of the model and it is impossible to effectively evaluate which knowledge is most valuable for the target city's tasks. Summary of the invention

[0005] 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 specification abstract and the invention title of this application to avoid blurring the purpose of this section, the specification abstract and the invention title, and such simplifications or omissions cannot be used to limit the scope of the present invention.

[0006] Therefore, the purpose of the present invention is to provide an interpretable task-specific cross-city transfer learning method, which significantly reduces the computational complexity and improves the adaptability to the target task and the accuracy of the transfer.

[0007] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions: S1, based on the deep matrix decomposition algorithm, dynamically extract the CTS area of ​​the source city; S2, using the auxiliary information of city POI and road network data to calculate the similarity between the source city and the target city, and select the CTS area with the highest similarity; S3: Migrate the data of the CTS area with the highest similarity to the target city, and complete the data of the target city through deep matrix decomposition technology.

[0008] As a preferred solution of the interpretable task-specific cross-city transfer learning method described in the present invention, in step S1, the specific steps of dynamically extracting the CTS area of ​​the source city based on the deep matrix decomposition algorithm are as follows: First, the space-time matrix of the source city Perform regional segmentation, where each unit corresponds to a specific area in the source city, where the space-time matrix of the source city represents the POI check-in data of people in each area of ​​the source city in each time period; Introducing the mask matrix , used to mask some elements in the source city matrix. The mask operation is defined as follows:

[0009] in, is a randomly generated value in the interval [0,1]. is the missing rate of the matrix; Through this operation, some areas can be randomly selected as missing data, thereby dynamically selecting the data in the source city matrix; After obtaining the partially missing source city matrix, it is reconstructed using deep matrix factorization technology; During the reconstruction process, the mask matrix is ​​adjusted iteratively. , so that it gradually selects the CTS area with the best effect in the reconstruction process as the CTS area of ​​the source city.

[0010] As a preferred solution of the interpretable task-specific cross-city transfer learning method described in the present invention, the deep matrix decomposition is specifically as follows:

[0011] The matrix of the source city Decomposed into two low-rank matrices Q and Z, the potential structure of the CTS region can be reconstructed, and then the mask matrix is ​​optimized .

[0012] As a preferred solution of the interpretable task-specific cross-city transfer learning method described in the present invention, in step S2, the similarity between the source city and the target city is calculated using the auxiliary information of the city POI and road network data, and the specific steps of selecting the CTS area with the highest similarity are as follows: Use POI data between cities, where POI data represents various points of interest in a city; Use cosine similarity to calculate the POI distribution similarity between the source city and the target city, and obtain the most similar area in the target city, that is, the CTS area of ​​the target city. The similarity calculation formula is as follows:

[0013] in and Represent the POI feature vectors of the CTS areas in the source city and the target city respectively; Select the k CTS regions with the highest similarity between the source city and the target city. The formula is as follows:

[0014] Among them, the Sort function sorts the results according to the similarity and selects the first k CTS regions with the highest similarity.

[0015] As a preferred solution of the interpretable task-specific cross-city transfer learning method described in the present invention, in step S3, the specific formula for completing the data of the target city by deep matrix decomposition technology is as follows:

[0016] in, It is the missing matrix of the target city in the CTS area. The Completion function completes the missing data of the target city based on the deep matrix decomposition technology.

[0017] Compared with the prior art, the present invention has the following beneficial effects: the present invention effectively overcomes the problems of high computational overhead, low migration effect and poor interpretability in the existing cross-city knowledge transfer technology through deep matrix decomposition technology and task-specific CTS region selection. By dynamically selecting the CTS region of the source city, the computational complexity is significantly reduced, and the adaptability to the target task and the accuracy of migration are improved. In addition, the present invention enhances the interpretability of the migration process, can clearly identify which regions are beneficial to the target city, and achieves efficient data completion when data is scarce. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] 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 in combination with the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. Among them: Figure 1 A flowchart of an interpretable task-specific cross-city transfer learning method of the present invention. DETAILED DESCRIPTION

[0019] 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.

[0020] This paper proposes a task-specific cross-city transfer learning method, which identifies the city-task-specific regions (CTS regions) in the source city and migrates them to the target city. Based on the deep matrix decomposition algorithm, the characteristic regions of the source city are dynamically selected, and the similarity between the source city and the target city is calculated using auxiliary information such as the city's points of interest (POI) and road network. Through this selective migration, the computational overhead can be effectively reduced and the accuracy and efficiency of transfer learning can be improved, especially in the case of data scarcity, such as Figure 1 As shown in Figure 1, this interpretable task-specific cross-city transfer learning method is as follows: 1. Extraction 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 in the source city during their respective time periods. In order to achieve efficient task-specific knowledge transfer, we first perform regional segmentation on the space-time matrix, where each unit corresponds to a specific area in the source city.

[0021] (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:

[0022] 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.

[0023] (2) Deep Matrix Decomposition Technology After obtaining the partially missing source city matrix, we reconstruct it using deep matrix factorization technology. Deep matrix factorization is an effective matrix completion method, especially suitable for low-rank matrices, i.e., matrices with a rank of is smaller than the number of rows and columns. Through deep matrix factorization, we transform the matrix of the source city Decomposed into two low-rank matrices Q and Z, the formula is as follows:

[0024] Through this decomposition, the potential structure of the CTS region can be reconstructed to optimize the mask matrix , ensuring that the selected CTS region is most representative for the specific task.

[0025] (3) CTS region selection process During the reconstruction process, we iteratively adjust the mask matrix , so that it gradually selects those CTS areas that have the best effect in the reconstruction process. These areas are the areas in the source city that are most helpful to the target task.

[0026] 2. 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 urban point of interest data 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.

[0027] (2) Similarity calculation In order to identify which areas in the target city are most similar to the CTS area of ​​the source city, we use cosine similarity 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. The similarity calculation formula is as follows:

[0028] in and Represent the POI feature vectors of the CTS regions in the source city and the target city, respectively. By calculating the similarity of each CTS region, we can identify which CTS regions of the source city are most similar to those of the target city.

[0029] (3) Select the CTS region with the highest similarity Next, we select the k CTS regions with the highest similarity between the source city and the target city, using the following formula:

[0030] Among them, the Sort function sorts the results according to the similarity and selects the first k CTS regions with the highest similarity.

[0031] 3. Data completion Finally, we migrate the data of these selected CTS areas to the target city and complete the data of the target city through deep matrix decomposition technology. The specific formula is as follows:

[0032] in, It is the missing matrix of the target city in the CTS area. The Completion function completes the missing data of the target city based on the deep matrix decomposition technology.

[0033] Although the present invention has been described above with reference to the embodiments, various modifications may be made thereto and parts thereof may be replaced by equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. 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. An interpretable task-specific cross-city transfer learning method, characterized in that Here are the steps: S1, based on the deep matrix decomposition algorithm, dynamically extract the CTS area of ​​the source city; S2, using the auxiliary information of city POI and road network data to calculate the similarity between the source city and the target city, and select the CTS area with the highest similarity; S3: Migrate the data of the CTS area with the highest similarity to the target city, and complete the data of the target city through deep matrix decomposition technology.

2. The interpretable task-specific cross-city transfer learning method according to claim 1, characterized in that: In step S1, the specific steps of dynamically extracting the CTS area of ​​the source city based on the deep matrix decomposition algorithm are as follows: First, the space-time matrix of the source city Perform regional segmentation, where each unit corresponds to a specific area in the source city, where the space-time matrix of the source city represents the POI check-in data of people in each area of ​​the source city in each time period; Introducing the mask matrix , used to mask some elements in the source city matrix. The mask operation is defined as follows: ； in, is a randomly generated value in the interval [0,1]. is the missing rate of the matrix; Through this operation, some areas can be randomly selected as missing data, thereby dynamically selecting the data in the source city matrix; After obtaining the partially missing source city matrix, it is reconstructed using deep matrix factorization technology; During the reconstruction process, the mask matrix is ​​adjusted iteratively. , so that it gradually selects the CTS area with the best effect in the reconstruction process as the CTS area of ​​the source city.

3. The interpretable task-specific cross-city transfer learning method according to claim 2, characterized in that: The details of deep matrix decomposition are as follows: ； The matrix of the source city Decomposed into two low-rank matrices Q and Z, the potential structure of the CTS region can be reconstructed, and then the mask matrix is ​​optimized .

4. The interpretable task-specific cross-city transfer learning method according to claim 1, characterized in that: In step S2, the similarity between the source city and the target city is calculated using the auxiliary information of the city POI and the road network data, and the specific steps of selecting the CTS area with the highest similarity are as follows: Use POI data between cities, where POI data represents various points of interest in a city; Use cosine similarity to calculate the POI distribution similarity between the source city and the target city, and obtain the most similar area in the target city, that is, the CTS area of ​​the target city. The similarity calculation formula is as follows: ； in and Represent the POI feature vectors of the CTS areas in the source city and the target city respectively; Select the k CTS regions with the highest similarity between the source city and the target city. The formula is as follows: ； Among them, the Sort function sorts the results according to the similarity and selects the first k CTS regions with the highest similarity.

5. The interpretable task-specific cross-city transfer learning method according to claim 1, characterized in that: In step S3, the specific formula for completing the data of the target city using deep matrix decomposition technology is as follows: ； in, It is the missing matrix of the target city in the CTS area. The Completion function completes the missing data of the target city based on the deep matrix decomposition technology.

Images