Federated learning based spatiotemporal graph prompting method and system

Abstract

The present application relates to the technical field of spatiotemporal big data, in particular to a spatiotemporal graph prompting method and system based on federated learning. By modeling the cross-platform interest point evaluation problem as a federated collaborative learning process, each platform can model and train using public interest point data without sharing local private data, thereby improving the evaluation performance of non-primary functions. In this process, external knowledge is obtained through federated training to optimize the original model. At the same time, a knowledge distillation mechanism is used to effectively integrate external knowledge, maintain the evaluation performance of the main function, and improve the accuracy of the non-main function. Federated training and knowledge distillation work together to balance the performance of the model and ultimately improve the overall effect.

Description

Technical Field

[0001] This invention relates to the field of spatiotemporal big data technology, specifically to a spatiotemporal graph prompting method and system based on federated learning. Background Technology

[0002] With the rapid development of mobile communication, Point-of-Interest (POI) rating systems have become essential tools in people's daily lives. A POI rating represents an overall assessment of a point of interest. It measures a user's preference for and satisfaction with that point of interest. POI rating systems aggregate ratings of points of interest across various scenarios, such as dining, shopping, or travel.

[0003] With the emergence of more interest-based review platforms, user habits and the dominant functions of these platforms are gradually changing. For example, Meituan primarily provides users with dining options that match their taste preferences; Dianping performs well in the areas of dining and entertainment; and Ctrip focuses more on reviews of functional interests, such as attractions, accommodations, and transportation hubs. These different focuses have led to significant differences in interest-based review practices across different platforms.

[0004] Current research primarily focuses on evaluating the interests of individual users on a single platform, that is, analyzing user preferences through interaction history and activity records to predict an individual user's evaluation of different interests. However, there is a lack of collaborative solutions that leverage the complementary nature of data across different platforms.

[0005] To address these issues, the most direct approach is to integrate data from all platforms and perform calculations and evaluations. However, relevant laws and regulations restrict direct data sharing. Furthermore, due to insufficient data on non-dominant functions of each platform, a severe imbalance in data distribution occurs. Evaluations of interest points for non-dominant functions may be unreliable. Therefore, this imbalance may prevent model aggregation methods from improving the accuracy of interest point evaluations for non-dominant functions, and may potentially reduce the evaluation performance of each platform's primary functions. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a spatiotemporal graph prompting method and system based on federated learning.

[0007] The technical solution adopted by this invention to solve its technical problem is:

[0008] A spatiotemporal graph prompting method based on federated learning is proposed, in which a set of platforms are known as participants, and each platform contains a user set, an interest point set, and an evaluation set as private data; the method includes:

[0009] S1. Each participant uses private data to train the model and obtain the original model;

[0010] S2. Each participant uses their own original model to calculate the logits output for the publicly available unlabeled point of interest data;

[0011] S3. Calculate the average logits output of all participants;

[0012] S4. During the federated training phase, each participant obtains the average logits output and combines it with the logits output of their own original model to determine the first loss value. The first loss value is then used to adjust their own original model to obtain the federated model.

[0013] S5. In the knowledge distillation stage, each participant uses their own original model as the teacher model and performs knowledge distillation training on the federated model based on private data to obtain an updated model.

[0014] S6. When the loop termination condition is met, obtain the prompt information for the interest point data using the current updated model; if the loop termination condition is not met, use the updated model as the original model and return to S2.

[0015] As a preferred embodiment, a further technical solution of the present invention is:

[0016] Preferably, the process by which each participant determines a first loss value based on the logits output of its own original model and the average logits output includes:

[0017] The similarity for each dimension is defined as follows:

[0018] ;

[0019] in, This represents the normalized value of the average logits output along the v-th dimension. Let represent the standardized value of the logits output of the i-th participant in the u-th dimension, as follows:

[0020] ;

[0021] in, This represents the logits output of the i-th participant along the u-th dimension. This represents the average value. Indicates standard deviation;

[0022] The loss function used to calculate the first loss value during the federated training phase is defined as follows:

[0023] ;

[0024] in, It is a hyperparameter; the loss function in the federated training phase aims to encourage similarity in the outputs of the logits of each participant along the same dimension.

[0025] Preferably, the knowledge distillation stage in S5 specifically includes:

[0026] Each participant uses its own federated model to calculate logits output from the private data. and output based on logits Calculate the second loss value;

[0027] Logits are calculated using the original model on private data. and combined with logits output With logits output Calculate the third loss value;

[0028] The fourth loss value is calculated using the second and third loss values, and the federated model is adjusted based on the fourth loss value to obtain the updated model.

[0029] Preferably, the loss function used to calculate the second loss value is defined as follows:

[0030] ;

[0031] in, Indicates the training labels.

[0032] Preferably, the loss function used to calculate the third loss value is defined as follows:

[0033] ;

[0034] The purpose of calculating the third loss value is to retain the knowledge gained from training on private data while maintaining the model's performance on the platform's main functions.

[0035] Preferably, the loss function used to calculate the fourth loss value is defined as follows:

[0036] ;

[0037] in, It is a hyperparameter used to balance external and local knowledge. This represents the third loss value. This represents the second loss value.

[0038] Preferably, each point of interest in the set of points of interest is described by three features: a unique ID, a geographical location, and a category.

[0039] Each rating record in the rating set represents a user's overall rating of a specific point of interest, and consists of rating ID, point of interest ID, user ID, rating time, rating score, and rating text.

[0040] The present invention also discloses a spatiotemporal graph prompting system based on federated learning, including various platforms and a central server. The platforms are participants, and each platform contains a set of users, a set of points of interest, and a set of evaluations as private data.

[0041] The platform is used to train models using private data to obtain the original model; and to calculate the logits output using its own original model on publicly available unlabeled interest point data.

[0042] The central server is used to calculate the average logits output of all participants;

[0043] The platform is used to obtain the average logits output during the federated training phase and combine it with the logits output of its own original model to determine the first loss value. The first loss value is then used to adjust the original model to obtain the federated model.

[0044] In the knowledge distillation phase, each participant uses their own original model as the teacher model and performs knowledge distillation training on the federated model based on private data to obtain an updated model.

[0045] When the loop termination condition is finally met, the updated model is used to obtain prompt information for the interest point data; if the loop termination condition is not met, the updated model is used as the original model and the next training cycle begins.

[0046] The present invention also discloses an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0047] Memory is used to store processor-executable instructions;

[0048] The processor, when executing instructions stored in memory, implements the aforementioned federated learning-based spatiotemporal graph hinting method.

[0049] The present invention, which adopts the above technical solution, has the following prominent features compared with the prior art:

[0050] The evaluation problem across different platforms is modeled as a federated learning problem. Each platform acts as an independent learning participant, optimizing its own scoring model through collaborative training while protecting its own data privacy. The model fully considers the strengths and weaknesses of each platform in different interest categories, improving the scoring accuracy of platforms in non-core functions through federated training, and maintaining their core scoring capabilities by combining knowledge distillation techniques. By analyzing the scoring deviations and global scoring trends of each platform, cross-platform information can be effectively integrated to optimize scoring results and ensure consistency and reliability of scoring across different platforms without data sharing. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the spatiotemporal graph prompting method based on federated learning in an embodiment of the present invention.

[0052] Figure 2 This is a principle block diagram of the spatiotemporal graph prompting method based on federated learning in this embodiment of the invention;

[0053] Figure 3 This is a schematic diagram of the structure of the electronic device in an embodiment of the present invention. Detailed Implementation

[0054] The present invention will be further illustrated below with reference to specific embodiments. The purpose of this illustration is solely to provide a better understanding of the invention. Therefore, the examples given do not limit the scope of protection of the present invention.

[0055] like Figure 1 and Figure 2 As shown in the figure, this embodiment presents a spatiotemporal graph prompting method based on federated learning. A group of platforms are known as participants, and each platform contains a user set, an interest point set, and an evaluation set as private data. The method includes:

[0056] S1. Each participant uses private data to train the model and obtain the original model;

[0057] S2. Each participant uses their own original model to calculate the logits output for the publicly available unlabeled point of interest data;

[0058] S3. Calculate the average logits output of all participants;

[0059] S4. During the federated training phase, each participant obtains the average logits output and combines it with the logits output of their own original model to determine the first loss value. The first loss value is then used to adjust their own original model to obtain the federated model.

[0060] S5. In the knowledge distillation stage, each participant uses their own original model as the teacher model and performs knowledge distillation training on the federated model based on private data to obtain an updated model.

[0061] S6. When the loop termination condition is met, obtain the prompt information for the interest point data using the current updated model; if the loop termination condition is not met, use the updated model as the original model and return to S2.

[0062] In practice, the federated point of interest (POI) rating problem presents a set of platforms P, each containing a user set, a point of interest (POI) set, and a rating set. Each platform can rate POIs using its own private data, but due to data limitations and differing rating standards, individual platform ratings may be biased. The optimization objective is to improve the rating accuracy of each platform through a federated approach without sharing the original data.

[0063] Each point of interest in the set is described by three features: a unique ID, geographical location, and category.

[0064] Each rating record in the rating set represents a user's overall rating of a specific point of interest, and consists of rating ID, point of interest ID, user ID, rating time, rating score, and rating text.

[0065] In implementation, to improve the rating performance across all platforms, the similarity between platform ratings should be encouraged, meaning the logits outputs across each dimension should be as similar as possible. Based on this, the process by which each participant determines the first loss value based on the logits output of their own original model and the average logits output includes:

[0066] The similarity for each dimension is defined as follows:

[0067] ;

[0068] in, This represents the normalized value of the average logits output along the v-th dimension. Let represent the standardized value of the logits output of the i-th participant in the u-th dimension, as follows:

[0069] ;

[0070] in, This represents the logits output of the i-th participant along the u-th dimension. This represents the average value. Indicates standard deviation;

[0071] The loss function used to calculate the first loss value during the federated training phase is defined as follows:

[0072] ;

[0073] in, It is a hyperparameter; the loss function in the federated training phase aims to encourage similarity in the outputs of the logits of each participant along the same dimension.

[0074] In implementation, during the knowledge distillation phase, each platform acquires a new federated model trained using public interest point data. This federated model integrates more external information, potentially improving the accuracy and consistency of evaluations across different platforms. However, directly using the federated model as the final model might affect the performance of local model training. To avoid this, a knowledge distillation mechanism is introduced. Specifically, the knowledge distillation phase in S5 includes:

[0075] Each participant uses its own federated model to calculate logits output from the private data. and output based on logits Calculate the second loss value;

[0076] Logits are calculated using the original model on private data. and combined with logits output With logits output Calculate the third loss value;

[0077] The fourth loss value is calculated using the second and third loss values, and the federated model is adjusted based on the fourth loss value to obtain the updated model.

[0078] Specifically, the loss function used to calculate the second loss value is defined as follows:

[0079] ;

[0080] in, Indicates the training labels.

[0081] Specifically, the loss function used to calculate the third loss value is defined as follows:

[0082] ;

[0083] The purpose of calculating the third loss value is to retain the knowledge gained from training on private data while maintaining the model's performance on the platform's main functions.

[0084] Specifically, the loss function used to calculate the fourth loss value is defined as follows:

[0085] ;

[0086] in, It is a hyperparameter used to balance external and local knowledge. The knowledge distillation stage can improve the reliability and performance of the model.

[0087] This invention models the cross-platform point of interest (POI) evaluation problem as a federated collaborative learning process. Each platform, without sharing its local private data, utilizes shared POI data for model training, thereby improving its evaluation performance in non-core functions. In this process, external knowledge is acquired through federated training to optimize its original model; simultaneously, a knowledge distillation mechanism effectively integrates external knowledge, maintaining the evaluation performance of core functions while improving the accuracy of non-core functions. The combined effect of federated training and knowledge distillation balances model performance, ultimately improving the overall result.

[0088] This invention also provides a spatiotemporal graph prompting system based on federated learning, including various platforms and a central server. The platforms are participants, and each platform contains a user set, an interest point set, and an evaluation set as private data.

[0089] The platform is used to train models using private data to obtain the original model; and to calculate the logits output using its own original model on publicly available unlabeled interest point data.

[0090] The central server is used to calculate the average logits output of all participants;

[0091] The platform is used to obtain the average logits output during the federated training phase and combine it with the logits output of its own original model to determine the first loss value. The first loss value is then used to adjust the original model to obtain the federated model.

[0092] In the knowledge distillation phase, each participant uses their own original model as the teacher model and performs knowledge distillation training on the federated model based on private data to obtain an updated model.

[0093] When the loop termination condition is finally met, the updated model is used to obtain prompt information for the interest point data; if the loop termination condition is not met, the updated model is used as the original model and the next training cycle begins.

[0094] The platform also defines the similarity for each dimension as follows:

[0095] ;

[0096] in, This represents the normalized value of the average logits output along the v-th dimension. Let represent the standardized value of the logits output of the i-th participant in the u-th dimension, as follows:

[0097] ;

[0098] in, This represents the logits output of the i-th participant along the u-th dimension. This represents the average value. Indicates standard deviation;

[0099] The loss function used to calculate the first loss value during the federated training phase is defined as follows:

[0100] ;

[0101] in, It is a hyperparameter; the loss function in the federated training phase aims to encourage similarity in the outputs of the logits of each participant along the same dimension.

[0102] The platform is also used to compute logits output from private data using its own federated model. and output based on logits Calculate the second loss value;

[0103] Logits are calculated using the original model on private data. and combined with logits output With logits output Calculate the third loss value;

[0104] The fourth loss value is calculated using the second and third loss values, and the federated model is adjusted based on the fourth loss value to obtain the updated model.

[0105] The loss function used to calculate the second loss value is defined as follows:

[0106] ;

[0107] in, Indicates the training labels.

[0108] The loss function used to calculate the third loss value is defined as follows:

[0109] ;

[0110] The purpose of calculating the third loss value is to retain the knowledge gained from training on private data while maintaining the model's performance on the platform's main functions.

[0111] The loss function used to calculate the fourth loss value is defined as follows:

[0112] ;

[0113] in, It is a hyperparameter used to balance external and local knowledge.

[0114] This invention models the cross-platform point of interest (POI) evaluation problem as a federated collaborative learning process. Each platform, without sharing its local private data, utilizes shared POI data for model training, thereby improving its evaluation performance in non-core functions. In this process, external knowledge is acquired through federated training to optimize its original model; simultaneously, a knowledge distillation mechanism effectively integrates external knowledge, maintaining the evaluation performance of core functions while improving the accuracy of non-core functions. The combined effect of federated training and knowledge distillation balances model performance, ultimately improving the overall result.

[0115] This invention also provides an electronic device, such as... Figure 3 As shown, it includes a processor 001, a communication interface 002, a memory 003, and a communication bus 004. The processor 001, communication interface 002, and memory 003 communicate with each other via the communication bus 004.

[0116] Memory 003 is used to store computer programs;

[0117] Processor 001, when executing the program stored in memory 003, implements the aforementioned spatiotemporal graph prompting method based on federated learning, wherein a group of platforms are known as participants, and each platform contains a user set, an interest point set, and an evaluation set as private data; the method includes:

[0118] S1. Each participant uses private data to train the model and obtain the original model;

[0119] S2. Each participant uses their own original model to calculate the logits output for the publicly available unlabeled point of interest data;

[0120] S3. Calculate the average logits output of all participants;

[0121] S4. During the federated training phase, each participant obtains the average logits output and combines it with the logits output of their own original model to determine the first loss value. The first loss value is then used to adjust their own original model to obtain the federated model.

[0122] S5. In the knowledge distillation stage, each participant uses their own original model as the teacher model and performs knowledge distillation training on the federated model based on private data to obtain an updated model.

[0123] S6. When the loop termination condition is met, obtain the prompt information for the interest point data using the current updated model; if the loop termination condition is not met, use the updated model as the original model and return to S2.

[0124] The solution provided by this invention models the evaluation problem of different platforms as a federated learning problem. Each platform acts as an independent learning participant, optimizing its own scoring model through collaborative training while protecting its own data privacy. It fully considers the scoring advantages and disadvantages of each platform in different interest categories, improves the scoring accuracy of platforms in non-core functions through the federated training phase, and maintains its core scoring capabilities by combining knowledge distillation technology. By analyzing the scoring deviations and global scoring trends of each platform, it can effectively integrate cross-platform information, optimize scoring results, and ensure consistency and reliability of scoring between different platforms without data sharing.

[0125] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0126] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0127] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0128] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0129] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0130] This invention enables collaborative training among multiple platforms without sharing private data through a federated training mechanism. Each platform utilizes unlabeled public interest point data to generate its own model output in each training round. By normalizing and aligning the outputs, the semantic consistency of the output layer is improved. This mechanism effectively mitigates data distribution differences caused by varying functional preferences among platforms, allowing each platform to learn valuable external knowledge from other platforms. This improves scoring performance in non-primary functional categories and enhances the model's generalization ability across multiple platforms.

[0131] This invention employs a knowledge distillation mechanism, effectively integrating the advantages of federated and original models. During the distillation phase, each platform uses its own point-of-interest data while referencing the outputs of both the federated and original models. A dual loss function guides the model to achieve a performance balance between dominant and non-dominant functions. This mechanism avoids the forgetting phenomenon that occurs when introducing external knowledge, maintaining the scoring ability of the original dominant functions, while simultaneously enhancing the model's performance in weaker function categories, thereby improving the overall stability and accuracy of the scoring system.

[0132] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the invention. All equivalent changes made based on the description and drawings of the present invention are included within the scope of the present invention.

Claims

1. A spatiotemporal graph prompting method based on federated learning, characterized in that, Given a group of platforms as participants, and each platform containing a user set, an interest set, and a rating set as private data; the method includes: S1. Each participant uses private data to train the model and obtain the original model; S2. Each participant uses their own original model to calculate the logits output for the publicly available unlabeled point of interest data; S3. Calculate the average logits output of all participants; S4. During the federated training phase, each participant obtains the average logits output and combines it with the logits output of their own original model to determine the first loss value. The first loss value is then used to adjust their own original model to obtain the federated model. S5. In the knowledge distillation stage, each participant uses their own original model as the teacher model and performs knowledge distillation training on the federated model based on private data to obtain an updated model. S6. When the loop termination condition is met, use the current updated model to obtain prompt information for the interest point data; if the loop termination condition is not met, use the updated model as the original model and return to S2. The process by which each participant determines the first loss value based on the logits output of their own original model and the average logits output includes: The similarity for each dimension is defined as follows: ； in, This represents the normalized value of the average logits output along the v-th dimension. Let represent the standardized value of the logits output of the i-th participant in the u-th dimension, as follows: ； in, This represents the logits output of the i-th participant along the u-th dimension. This represents the average value. Indicates standard deviation; The loss function used to calculate the first loss value during the federated training phase is defined as follows: ； in, It is a hyperparameter; the loss function in the federated training phase aims to encourage similarity in the outputs of the logits of each participant along the same dimension. The knowledge distillation stage in S5 specifically includes: Each participant uses its own federated model to calculate logits output from the private data. and output based on logits Calculate the second loss value; Logits are calculated using the original model on private data. and combined with logits output With logits output Calculate the third loss value; The fourth loss value is calculated using the second and third loss values, and the federated model is adjusted based on the fourth loss value to obtain the updated model; Each point of interest in the set is described by three features: a unique ID, geographical location, and category. Each rating record in the rating set represents a user's overall rating of a specific point of interest, and consists of rating ID, point of interest ID, user ID, rating time, rating score, and rating text.

2. The spatiotemporal graph prompting method based on federated learning according to claim 1, characterized in that, The loss function used to calculate the second loss value is defined as follows: ； in, Indicates the training labels.

3. The spatiotemporal graph prompting method based on federated learning according to claim 1, characterized in that, The loss function used to calculate the third loss value is defined as follows: ； The purpose of calculating the third loss value is to retain the knowledge gained from training on private data while maintaining the model's performance on the platform's main functions.

4. The spatiotemporal graph prompting method based on federated learning according to claim 1, characterized in that, The loss function used to calculate the fourth loss value is defined as follows: ； in, It is a hyperparameter used to balance external and local knowledge. This represents the third loss value. This represents the second loss value.

5. A spatiotemporal graph prompting system based on federated learning, characterized in that, This includes various platforms and a central server. The platforms are participants, and each platform contains a set of users, a set of points of interest, and a set of evaluations as private data. The platform is used to train models using private data to obtain the original model; and to calculate the logits output using its own original model on publicly available unlabeled interest point data. The central server is used to calculate the average logits output of all participants; The platform is used to obtain the average logits output during the federated training phase and combine it with the logits output of its own original model to determine the first loss value. The first loss value is then used to adjust its own original model to obtain the federated model. In the knowledge distillation phase, each participant uses their own original model as the teacher model and performs knowledge distillation training on the federated model based on private data to obtain an updated model. When the loop termination condition is finally met, prompt information for the interest point data is obtained using the current updated model. If the loop termination condition is not met, the updated model will be used as the original model before entering the next training cycle. The process of determining the first loss value includes: The similarity for each dimension is defined as follows: ； in, This represents the normalized value of the average logits output along the v-th dimension. Let represent the standardized value of the logits output of the i-th participant in the u-th dimension, as follows: ； in, This represents the logits output of the i-th participant along the u-th dimension. This represents the average value. Indicates standard deviation; The loss function used to calculate the first loss value during the federated training phase is defined as follows: ； in, It is a hyperparameter; the loss function in the federated training phase aims to encourage similarity in the outputs of the logits of each participant along the same dimension. The knowledge distillation stage specifically includes: Each participant uses its own federated model to calculate logits output from the private data. and output based on logits Calculate the second loss value; Logits are calculated using the original model on private data. and combined with logits output With logits output Calculate the third loss value; The fourth loss value is calculated using the second and third loss values, and the federated model is adjusted based on the fourth loss value to obtain the updated model; Each point of interest in the set is described by three features: a unique ID, geographical location, and category. Each rating record in the rating set represents a user's overall rating of a specific point of interest, and consists of rating ID, point of interest ID, user ID, rating time, rating score, and rating text.

6. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory is used to store processor-executable instructions; A processor, when executing instructions stored in memory, implements the spatiotemporal graph hinting method based on federated learning as described in any one of claims 1-4.

Images