A privacy-preserving method and system for predicting points of interest

Abstract

This invention relates to a privacy-preserving method and system for predicting points of interest, belonging to the field of machine learning technology, and includes the following steps: S1: Collecting user data using anonymization technology; S2: Participants classify and aggregate the data according to a pre-defined service domain; S3: Participants preprocess and extract features from the aggregated data; S4: Participants train a local model by modeling four key factors—social relationships, user similarity, user intent, and spatiotemporal context—using a multi-factor fusion personalized point of interest recommendation method; S5: Participants perturb the original gradient by adding paired masks and upload the encrypted gradient; S6: The computing service provider aggregates the received encrypted information and distributes the global model update information to each participant. Participants receive and update their local models, and the above process is repeated until the model converges.

Description

Technical Field

[0001] This invention belongs to the field of machine learning technology and relates to a privacy-preserving method and system for predicting points of interest. Background Technology

[0002] The combination of recommender systems and location-based social networks has given rise to Point of Interest (POI) recommender systems. These systems predict a user's next point of interest within a specific timeframe based on available information in their check-in sequence. They are widely used in fields such as smart cars, traffic management, and the service industry, while also providing technical guidance for related applications such as route planning, commercial advertising, and traffic prediction. Existing POI recommendations mostly focus on the static characteristics of user interactions with points of interest and typically consider only a single factor, neglecting the mining of semantic and spatiotemporal contextual information. For example, if a person's daily pattern is identified as "going to work" by a POI recommender system, the system might recommend places they frequently visit based on their historical location data, but it ignores the fact that they may not be able to leave their workplace to go to these places during work hours. Similarly, a user might be more inclined to visit entertainment venues on weekends and more frequently visit their workplace on weekdays. Furthermore, existing recommendation methods are mostly based on centralized architectures for data storage and model training, which has certain limitations. On the one hand, the centralized collection and storage of check-in data raises concerns about privacy leaks, as users' real-time or historical location data is highly sensitive and may expose private information such as their residence and workplace. On the other hand, regulations such as the General Data Protection Regulation (GDPR) impose strict restrictions on referral service providers in collecting, transmitting, and processing user data.

[0003] Currently, several methods for POI prediction employ deep learning. Liu et al., in "GNN-based long and short term preference modeling for next-location prediction" [in InformationSciences], proposed a GNN-based model that transforms location into a low-dimensional representation and integrates both long-term and short-term user preferences to comprehensively express dynamic preferences. While this addresses data sparsity and dynamism to some extent, it doesn't consider more comprehensive feature information such as semantic information and carries the risk of privacy breaches. Building on this, some privacy-preserving POI prediction methods have been proposed. Perifanis et al., in "Federated neural collaborative filtering" [in Knowledge-Based Systems, 2022, 242:108441], proposed a federated neural collaborative filtering algorithm that allows each participant to update the model locally and only uploads the calculated weights to the server as a measure to protect user privacy. However, this research does not fully consider the impact of multidimensional influencing factors on the recommendation model's performance, nor does it address the security issues during gradient transfer.

[0004] Although POI prediction methods have been extensively researched and applied, these techniques still have many limitations and shortcomings. Here are some of the main challenges:

[0005] (1) Dependence on a single factor: Current POI prediction methods often rely on only a single factor, such as the user's historical check-in data, geographical location information, social relationships, etc., but ignore the semantic information and spatiotemporal context information of user behavior patterns. For example, they do not consider the different travel preferences of users on weekdays and weekends, as well as the user's current location restrictions, which can easily make the recommendation results outdated or irrelevant.

[0006] (2) Inadequate privacy protection: The existing centralized data collection and storage model easily raises concerns among users about the leakage of personal privacy, especially when processing users' real-time or historical location data. This model fails to adequately protect users' sensitive information, increasing the risk of data misuse or leakage.

[0007] (3) Computational resource pressure: Centralized model training requires processing and analyzing a large amount of user data, which places extremely high demands on computing resources. This model may not only lead to low processing efficiency, but also increase operating costs and limit the scalability and response speed of the system. Summary of the Invention

[0008] In view of this, the purpose of this invention is to provide a privacy-preserving method and architecture for predicting points of interest. To address the single-factor dependency problem of existing point of interest prediction methods, this invention comprehensively considers four influencing factors: social relationships, user similarity, user intent, and spatiotemporal context. It also incorporates deep learning technology to alleviate the data sparsity problem and achieve high-precision recommendations.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] On the one hand, the present invention provides a privacy-preserving method for predicting points of interest, comprising the following steps:

[0011] S1: Use anonymization techniques to collect user data;

[0012] S2: Participants categorize and aggregate data according to pre-defined service areas;

[0013] S3: Participants preprocess and extract features from the aggregated data;

[0014] S4: Participants use a multi-factor fusion personalized interest recommendation method to model four key factors: social relationships, user similarity, user intent, and spatiotemporal context, thereby training a local model.

[0015] S5: Participants perturb the original gradient by adding paired masks and then upload the encrypted gradient;

[0016] S6: The computing service provider aggregates the received encrypted information and distributes the global model update information to each participant. The participants receive and update their local models, and repeat the above process until the model converges.

[0017] Furthermore, the collection of user data using anonymization technology described in step S1 specifically includes:

[0018] Collect data related to point of interest prediction, and use anonymization technology to process the data in real time to remove or obfuscate information that can directly or indirectly identify the user.

[0019] The data related to point of interest prediction includes user location information, identity information, browsing history, search keywords, and application usage frequency;

[0020] The anonymization techniques include k-anonymity, l-diversity, and differential privacy.

[0021] For location information, precise coordinates are converted into a regional range through fuzzing or generalization techniques; for identity information, device identifiers and user names are completely removed or irreversibly deformed to ensure that the collected data cannot be directly or indirectly traced back to a specific user.

[0022] Furthermore, in step S3, the participants preprocess the aggregated data, specifically including cleaning the aggregated data, including removing outliers and duplicate data;

[0023] The feature extraction involves extracting valuable features for predicting points of interest, including work or weekend, different time periods, frequently visited areas, and regional attributes.

[0024] Furthermore, the modeling described in step S4, which involves four key factors—social relationships, user similarity, user intent, and spatiotemporal context—specifically includes:

[0025] For social relationship modeling, a graph attention mechanism is used, given the representation of social neighbor nodes. and neighbor set N S (u), α uj Given the attention weights from user u to neighbor j, sum the weights of the neighbors to obtain the representation of user nodes with social relationships.

[0026]

[0027] For user similarity modeling, the first step is to calculate the collaborative similarity S between two users using similarity methods. uv Given a similarity threshold χ for users u If and only if S uv >χ u Only when user u and user v are defined as similar users; then, a graph attention mechanism is used, for user u, given the features of similar neighbor nodes. and neighbor set N R (u), α uv Given the weights from user u to neighbor v, sum the weights over the neighbors to obtain the representation of similar user nodes.

[0028] User intent modeling specifically includes:

[0029] 1) First, a knowledge graph is introduced to classify tags based on semantic information, and a graph convolutional network is used to update the tags. Then, a multimodal fusion strategy is adopted to transform and represent the semantic information of text, image, and video modalities, as shown below:

[0030] v new =F(A(T) v ,I v ,Sv ),T v ,I v ,S v )

[0031] =RELU(W2·RELU(W1·[a T T v +a I I v +a S S v ]+b1)+b2)

[0032] Where A(T) v ,I v ,S v )=(a T ,a I ,a S ), a T ,a I ,a S Here, represent the attention weights for text, image, and video modalities, respectively; F is the multilayer perceptron; and v... new It is the fused vector;

[0033] 2) Process sequence data using LSTM, let H = [h1, h2, ..., h T ] is the output matrix of LSTM. An adaptive multi-pooling strategy is applied to the output matrix of LSTM. A selection function p(H) is defined, which determines the pooling method based on the eigenvalue distribution of H. Then, the pooled vector is obtained through the selected pooling method.

[0034] 3) Utilize generative adversarial networks (GANs) to optimize the parameters of fully connected layers and activation functions. Let the generator's input be a random noise vector z, and the generated vector be v. gen The discriminator analyzes the true user intent vector v. pool and the generated vector v gen The outputs are D(v) pool ) and D(v gen The training objective is to minimize the following loss function:

[0035]

[0036] For spatiotemporal context modeling, a user's check-in behavior is affected by a specific time point t. k Geographical location j =(long j ,lat j Is it a holiday? Distance from the previous check-in and time difference τ kThe influence of five factors is first addressed using one-hot encoding, transforming the five factors affecting each check-in into a d-dimensional vector. At each time step, the temporarily embedded vectors are aggregated. in, These are learnable parameters; a contextual attention mechanism is used, and the user pays attention to the parameters in t. k The embedding vector representation of the check-in record at any given time is as follows:

[0037]

[0038] in, Let be the embedding vector of the i-th feature factor after passing through the attention mechanism. Let represent the embedding vector representation of the i-th factor in the k-th historical check-in. For the corresponding weights, These are learnable parameters; The input is fed into an LSTM to obtain t. k Hidden state of time Weights are generated using a time attention mechanism. Finally, the user's spatiotemporal preference representation is obtained.

[0039] After obtaining the embedding vectors of the four influencing factors, the different embedding vectors are first aggregated using an attention mechanism.

[0040]

[0041] Calculate its embedding vector with a specific POI. The inner product is used to predict user ratings for POIs. And calculate the loss function.

[0042] Participants calculate the loss value of the sample Then, iteratively minimize the loss to obtain the optimal model, where x k Let y represent the feature vector of the i-th training sample. k That is the corresponding tag.

[0043] Furthermore, in step S5, the participants perturb the original gradient by adding paired masks and then upload the encrypted gradient, specifically including the following steps:

[0044] S51: Participants calculate local gradients

[0045] S52: Participants mask the local gradient by perturbing the original gradient through the addition of paired masks.

[0046]

[0047] Where, τ i It is a random noise term, α ij It is a coefficient used to control the effect of the gradient of the j-th participant on the mask of the i-th participant.

[0048] Furthermore, step S6 specifically includes the following steps:

[0049] S61: The computation service provider collects and aggregates the transmitted perturbation gradients. Decrypt using the private key to obtain the global model x' agg ;

[0050] S62: The computing service provider computes and updates the global model:

[0051] S63: After receiving the global model update information from the computing service provider, the participant updates the local model.

[0052] S64: Repeat steps S61-S63, that is, the participants continuously train the local model, generate new encrypted update information and send it to the computing service provider. The computing service provider then aggregates and distributes the update information again until the model reaches a convergent state.

[0053] S65: Participants distribute the final trained global model to users to provide personalized POI recommendations.

[0054] On the other hand, the present invention provides a privacy-preserving point of interest prediction system, including a collection layer, a convergence layer, and a calculation layer;

[0055] Collection layer: Uses anonymization technology to collect user data and uploads it to the POI service provider in the aggregation layer, enabling the POI service provider and the computing service provider to jointly train an interest point prediction model;

[0056] Aggregation Layer: Each participant receives anonymized user data from the acquisition layer and aggregates this data within a specific service domain; participants use the aggregated user data to train an interest point prediction model; after the model training is complete, they send encrypted model gradients to the computing service provider.

[0057] Computation layer: In each training cycle, the computing service provider receives the local training results from each participant for that round, performs global aggregation of the uploaded encrypted gradients, decrypts them, and broadcasts the latest global model to each participant.

[0058] The beneficial effects of this invention are as follows: By processing anonymized data locally at the aggregation layer and transmitting only encrypted model updates or gradient information to the computation layer, this invention significantly reduces the risk of sensitive data leakage, thereby strengthening the protection of user privacy. Furthermore, by processing user data at the aggregation layer, this invention effectively reduces the computational burden on the acquisition and computation layers. Through deep learning technology, this invention considers four dimensions—social relationships, user similarity, user intent, and spatiotemporal context—to achieve personalized modeling. This not only improves the accuracy of capturing the dynamics of user behavior but also effectively alleviates the data sparsity problem, enhancing the accuracy of the recommendation system and user satisfaction.

[0059] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0060] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0061] Figure 1 This is a system architecture diagram as described in this invention;

[0062] Figure 2 This invention relates to a multi-factor fusion-based personalized point of interest recommendation method. Detailed Implementation

[0063] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0064] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0065] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0066] This invention provides a privacy-preserving method for predicting points of interest, such as... Figure 1 As shown, the architecture used in this method is a three-layer architecture consisting of a data acquisition layer, a data aggregation layer, and a computation layer.

[0067] Collection layer: Uses anonymization technology to collect user data and uploads it to the POI service provider in the aggregation layer, enabling the POI service provider and the computing service provider to jointly train an accurate point of interest prediction model.

[0068] Aggregation Layer: First, the aggregation layer consists of multiple Points of Interest (POI) service providers, which are the participants in federated learning. Each participant receives anonymized user data from the acquisition layer and aggregates this data within a specific service domain. Second, the participants use the aggregated user data to train an POI prediction model. After the model training is complete, encrypted model gradients are sent to the computing service providers to ensure data confidentiality during transmission.

[0069] Computation layer: In each training cycle, the computing service provider receives the local training results from each participant for that round, performs global aggregation of the uploaded encrypted gradients, decrypts them, and broadcasts the latest global model to each participant.

[0070] Through the collaborative work of the aggregation layer and the computation layer, model knowledge sharing is achieved. In this architecture, users anonymously upload data to the aggregation layer, where data is processed only. Encrypted model updates or gradient information are sent to the computation layer for aggregation. This design not only reduces the risk of data leakage but also effectively alleviates the computational burden on the computation and acquisition layers by performing local model training at the aggregation layer.

[0071] Figure 2 This paper presents a multi-factor fusion-based personalized interest point recommendation method that incorporates deep learning technology. Personalized modeling is implemented for four influencing factors: social relationships, user similarity, user intent, and spatiotemporal context. This not only improves the accuracy of capturing the dynamics of user behavior but also effectively alleviates the data sparsity problem, thereby enhancing the accuracy of the recommendation system and user satisfaction.

[0072] This method includes the following steps:

[0073] S1: Use anonymization techniques to collect user data;

[0074] First, user data is collected, including but not limited to user location information, identification information, browsing history, search keywords, application usage frequency, and other data that may be related to point-of-interest prediction. During data collection, anonymization techniques such as k-anonymity, l-diversity, or differential privacy are used to process the data in real time, removing or obfuscating information that can directly or indirectly identify the user. For location information, precise coordinates are converted into a larger geographical area through obfuscation or generalization techniques; for identification information, device identifiers, user names, and other information are completely removed or irreversibly transformed to ensure that the collected data cannot be directly or indirectly traced back to a specific user.

[0075] S2: Participants categorize and aggregate data according to pre-defined service areas;

[0076] S3: Participants preprocess and extract features from the aggregated data;

[0077] The aggregated data is cleaned, including removing outliers and duplicate data; then features are extracted, including features that are valuable for predicting points of interest, such as work / weekend, different time periods, frequently active areas, and regional attributes.

[0078] S4: Participants use a multi-factor fusion personalized interest recommendation method to model four key factors: social relationships, user similarity, user intent, and spatiotemporal context, thereby training a local model.

[0079] S41: For social relationship modeling, a graph attention mechanism is used, given the representation of social neighbor nodes. and neighbor set N S (u), α uj Given the attention weights from user u to neighbor j, sum the weights of the neighbors to obtain the representation of user nodes with social relationships.

[0080] S42: For modeling similar users, first calculate the collaborative similarity S between the two users using a similarity method. uv Given a similarity threshold χ for users u If and only if S uv >χ u Only when user u and user v are defined as similar users will user u and user v be considered similar users. Then, a graph attention mechanism is used, where, for user u, given the features of similar neighbor nodes... and neighbor set N R (u), α uv Given the weights from user u to neighbor v, sum the weights over the neighbors to obtain the representation of similar user nodes.

[0081] S43: For user intent modeling, a knowledge graph is first introduced to classify tags based on semantic information, and a graph convolutional network is used to update the tags. Then, a multimodal fusion strategy is adopted, which considers the semantic information of text, image, and video modalities for transformation and representation, as shown below:

[0082] v new =F(A(T) v ,I v ,S v ),T v ,I v ,S v )

[0083] =RELU(W2·RELU(W1·[a T T v +a I I v +a S S v ]+b1)+b2)

[0084] Where A(T) v ,I v ,S v )=(a T ,a I ,a S ), a T ,a I ,a S Here, represent the attention weights for text, image, and video modalities, respectively; F is the multilayer perceptron; and v... new It is the fused vector.

[0085] S45: Process sequence data using LSTM, let H = [h1, h2, ..., h T [H] is the output matrix of the LSTM. An adaptive multi-pooling strategy is applied to the output matrix of the LSTM. A selection function p(H) is defined, which determines the pooling method based on the eigenvalue distribution of H. Then, the pooled vector v is obtained through the selected pooling method. pool .

[0086] S46: Utilize generative adversarial networks (GANs) to optimize the parameters of fully connected layers and activation functions. Let the generator's input be a random noise vector z, and the generated vector be v. gen The discriminator analyzes the true user intent vector v. pool and the generated vector v gen The outputs are D(v) pool ) and D(v gen The training objective is to minimize the following loss function:

[0087]

[0088] S47: For spatiotemporal context modeling, a user's check-in behavior is affected by a specific time point t. k Geographical location j =(long j ,lat j Is it a holiday? Distance from the previous check-in and time difference τ k The influence of five factors is first addressed using one-hot encoding, transforming the five factors affecting each check-in into a d-dimensional vector. At each time step, the temporarily embedded vectors are aggregated. in, These are learnable parameters.

[0089] S48: Because the degree to which various factors influence user behavior varies, a contextual attention mechanism is employed. The user in t k The embedding vector representation of the check-in record at any given time is as follows:

[0090]

[0091] in, Let be the embedding vector of the i-th feature factor after passing through the attention mechanism. Let represent the embedding vector representation of the i-th factor in the k-th historical check-in. For the corresponding weights, These are learnable parameters. The input is fed into an LSTM to obtain t. k Hidden state of time Weights are generated using a time attention mechanism. Finally, the user's spatiotemporal preference representation is obtained.

[0092]

[0093] S49: After obtaining the embedding vectors of the four influencing factors, the different embedding vectors are first aggregated using an attention mechanism. Calculate its embedding vector with a specific POI. The inner product is used to predict user ratings for POIs. And calculate the loss function.

[0094] Participants calculate the loss value of the sample Then, iteratively minimize the loss to obtain the optimal model. Where x k Let y represent the feature vector of the i-th training sample. k That is the corresponding tag.

[0095] S5: Participants perturb the original gradient by adding paired masks and then upload the encrypted gradient;

[0096] Participants first calculate the local gradient The participants then masked the local gradient by perturbing the original gradient through the addition of paired masks. Where, τ i It is a random noise term, α ij It is a coefficient used to control the effect of the gradient of the j-th participant on the mask of the i-th participant.

[0097] S6: The computing service provider aggregates the received encrypted information and distributes the global model update information to each participant. The participants receive and update their local models, and repeat the above process until the model converges.

[0098] S61: The computation service provider collects and aggregates the transmitted perturbation gradients. Decrypt using the private key to obtain the global model x' agg .

[0099] S62: The computing service provider computes and updates the global model:

[0100] S63: After receiving the global model update information from the computing service provider, the participant updates the local model.

[0101] S64: The above process is repeated continuously, that is, the participants continue to train the local model, generate new encrypted update information and send it to the computing service provider, the computing service provider aggregates and sends the update information again, until the model reaches a convergence state.

[0102] S65: Participants distribute the final trained global model to users to provide personalized POI recommendations.

[0103] In the above embodiments, the reference to "this embodiment" in the specification indicates that a specific feature, structure, or characteristic described in connection with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple appearances of "this embodiment" do not necessarily refer to the same embodiment.

[0104] In the above embodiments, although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory structures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed. The embodiments of the invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0105] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the methods in this embodiment.

[0106] This embodiment also provides an electronic terminal, including: a processor and a memory;

[0107] The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory to cause the terminal to perform any of the methods in this embodiment.

[0108] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0109] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0110] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0111] 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.

[0112] This invention can be used in a wide range of general-purpose or special-purpose computing system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0113] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A privacy-preserving method for predicting points of interest, characterized in that: Includes the following steps: S1: Use anonymization techniques to collect user data; S2: Participants categorize and aggregate data according to pre-defined service areas; S3: Participants preprocess and extract features from the aggregated data; S4: Participants trained a local model using a multi-factor fusion personalized interest recommendation method, modeling four key factors: social relationships, user similarity, user intent, and spatiotemporal context; specifically including: For social relationship modeling, a graph attention mechanism is used, given the representation of social neighbor nodes. Gather with neighbors , For users To the neighbor The attention weights are then used to sum the values ​​over their neighbors to obtain a representation of user nodes with social relationships. ; For user similarity modeling, the first step is to combine similarity methods to calculate the collaborative similarity between two users. Given a similarity threshold for users If and only if At that time, the user and users Only then will they be defined as similar users; then a graph attention mechanism is used for users. Given the features of similar neighbor nodes Gather with neighbors , For users To the neighbor The weights are used to sum the values ​​of the neighbors, resulting in a representation of similar user nodes. ; User intent modeling specifically includes: 1) First, a knowledge graph is introduced to classify tags based on semantic information, and a graph convolutional network is used to update the tags. Then, a multimodal fusion strategy is adopted to transform and represent the semantic information of text, image, and video modalities, as shown below: in , These are the attention weights for text, image, and video modalities, respectively. It is a multilayer perceptron. It is the fused vector; 2) Process sequence data using LSTM, assuming... This is the output matrix of the LSTM. An adaptive multi-pooling strategy is applied to the output matrix of the LSTM, and a selection function is defined. It is based on The pooling method is determined by the distribution of eigenvalues, and then the pooled vector is obtained by selecting the pooling method. 3) Utilize generative adversarial networks to optimize the parameters of fully connected layers and activation functions; assume the generator's input is a random noise vector. The generated vector is The discriminator analyzes the true user intent vector. and generated vectors The outputs are respectively and The training objective is to minimize the following loss function: For spatiotemporal context modeling, user check-in behavior is affected by specific time points. Geographical location Is it a holiday? Distance from the previous check-in and time difference The influence of five factors is first addressed using one-hot coding to convert the five factors affecting each check-in into... A dimensional vector, with temporary embedding vectors aggregated at each time step. ,in, , These are learnable parameters; a context-based attention mechanism is employed, allowing the user to... The embedding vector representation of the check-in record at any given time is as follows: in, For the first time after passing the attention mechanism Embedding vectors of each feature factor Indicates the first The first time I've checked in historically. Embedded vector representation of each factor For the corresponding weights, , These are learnable parameters; Input into LSTM, get Hidden state of time Weights are generated through a time attention mechanism. Ultimately, the user's spatiotemporal preference representation is obtained. ; After obtaining the embedding vectors of the four influencing factors, the different embedding vectors are first aggregated using an attention mechanism. Calculate its embedding vector with a specific POI. The inner product is used to predict user ratings for POIs. And calculate the loss function. ; Participants calculate the loss value of the sample Then, iteratively minimize the loss to obtain the optimal model, where Let represent the feature vector of the i-th training sample. These are the corresponding tags; S5: Participants perturb the original gradient by adding paired masks and then upload the encrypted gradient, specifically including the following steps: S51: Participants calculate local gradients ; S52: Participants mask the local gradient by perturbing the original gradient through the addition of paired masks. in, It is a random noise term. It is a coefficient used to control the first The gradient of the i-th participant with respect to the j-th The effect of the mask on each participant; S6: The computing service provider aggregates the received encrypted information and distributes the global model update information to each participant. The participants receive and update their local models, and repeat the above process until the model converges.

2. The privacy-preserving point-of-interest prediction method according to claim 1, characterized in that: Step S1, which involves collecting user data using anonymization techniques, specifically includes: Collect data related to point of interest prediction, and use anonymization technology to process the data in real time to remove or obfuscate information that can directly or indirectly identify the user. The data related to point of interest prediction includes user location information, identity information, browsing history, search keywords, and application usage frequency; The anonymization technology includes selection anonymous, Diversity and differential privacy; For location information, precise coordinates are converted into a regional range through fuzzing or generalization techniques; for identity information, device identifiers and user names are completely removed or irreversibly deformed to ensure that the collected data cannot be directly or indirectly traced back to a specific user.

3. The privacy-preserving point-of-interest prediction method according to claim 1, characterized in that: Step S3 involves the participants preprocessing the aggregated data, specifically including cleaning the aggregated data, including removing outliers and duplicate data; The feature extraction involves extracting valuable features for predicting points of interest, including work or weekend, different time periods, frequently visited areas, and regional attributes.

4. The privacy-preserving point-of-interest prediction method according to claim 1, characterized in that: Step S6 specifically includes the following steps: S61: The computation service provider collects and aggregates the transmitted perturbation gradients. The global model is obtained by decryption using the private key. ; S62: The computing service provider computes and updates the global model: ; S63: After receiving the global model update information from the computing service provider, the participant updates the local model. S64: Repeat steps S61-S63, that is, the participants continuously train the local model, generate new encrypted update information and send it to the computing service provider. The computing service provider then aggregates and distributes the update information again until the model reaches a convergent state. S65: Participants distribute the final trained global model to users to provide personalized POI recommendations.

5. A privacy-preserving point-of-interest prediction system, characterized in that: The privacy-preserving point of interest prediction method according to claim 1 includes a collection layer, a convergence layer, and a calculation layer; Collection layer: Uses anonymization technology to collect user data and uploads it to the POI service provider in the aggregation layer, enabling the POI service provider and the computing service provider to jointly train an interest point prediction model; Aggregation Layer: Each participant receives anonymized user data from the acquisition layer and aggregates this data within a specific service domain; participants use the aggregated user data to train an interest point prediction model; after the model training is complete, they send encrypted model gradients to the computing service provider. Computation layer: In each training cycle, the computing service provider receives the local training results from each participant for that round, performs global aggregation of the uploaded encrypted gradients, decrypts them, and broadcasts the latest global model to each participant.

Images