Interest point prediction method and system with privacy protection

Abstract

The invention relates to an interest point prediction method and system with privacy protection, and belongs to the technical field of machine learning, and the method comprises the following steps: S1, employing an anonymization technology to collect user data; s2, the participants classify and gather the data according to a preset service field range; s3, the participant performs preprocessing and feature extraction on the gathered data; s4, the participants perform modeling from four key factors of social relation, user similarity, user intention and spatio-temporal context through a multi-factor fusion personalized interest point recommendation method so as to train a local model; s5, the participant disturbs the original gradient by adding paired masks and uploads the encrypted gradient; and S6, the calculation service provider carries out aggregation operation on the received encrypted information, global model updating information is issued to each participant, the participant receives and updates the local model, and the above process is repeated until the model converges.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning and relates to a method and system for predicting points of interest with privacy protection. Background Art

[0002] The combination of recommendation systems and location-based social networks has given rise to point of interest (POI) recommendation systems, which can predict the next point of interest that a user may visit within a specific time based on the available information in the user's check-in sequence. It is widely used in fields such as smart cars, traffic management and service industries, and provides technical guidance for related applications such as route planning, commercial advertising and traffic prediction. Most existing POI recommendations focus on the static features of user-point of interest interactions, and usually only consider a single factor, while ignoring the mining of semantic information and spatiotemporal context information. For example, if a person's daily pattern is identified as "going to work" by the POI recommendation system, the system may recommend places he frequently visits based on his historical location data, but ignores the fact that he may not be able to leave his workplace to visit these places during work; for example, a user may be more inclined to visit entertainment venues on weekends, while he may visit office locations more frequently on weekdays. In addition, most existing recommendation methods are based on centralized architectures for data storage and model training, which have certain limitations. On the one hand, the centralized collection and storage of check-in data has caused users to worry about privacy leakage. Users' real-time location or historical location data is very sensitive and may expose users' private information such as residence and workplace. On the other hand, regulations such as the General Data Protection Regulation impose strict restrictions on referral service providers in collecting, transmitting and processing user data.

[0003] At present, there are some methods for POI prediction using deep learning. Liu et al. proposed a GNN-based model in "GNN-based long and short term preference modeling for next-location prediction" [in Information Sciences], which converts the location into a low-dimensional representation and integrates the user's long-term and short-term preferences to fully express dynamic preferences. Although the data sparsity and dynamics are solved to a certain extent, more comprehensive feature information such as semantic information is not considered, and there is a risk of privacy leakage. On this basis, some privacy-preserving POI prediction methods have also been proposed. Perifanis et al. proposed a federated neural collaborative filtering algorithm in "Federated neural collaborative filtering" [in Knowledge-Based Systems, 2022, 242: 108441], which allows each participant to update the model locally and only upload the calculated weights to the server as a measure to protect user privacy. However, this study did not fully consider the impact of multi-dimensional influencing factors on the performance of the recommendation model, nor did it consider the security issues in the gradient transmission process.

[0004] Although POI prediction methods have been widely studied and applied, these techniques still have many limitations and defects. The following are some of the main challenges:

[0005] (1) Single factor dependence: Current POI prediction methods often rely on only a single factor, such as the user's historical check-in data, geographic location information, social relationships, etc., but ignore the semantic information and spatiotemporal context information of the user's behavior pattern. For example, they do not consider the different travel tendencies of users on weekdays and weekends, and the situation where users are restricted by their current location, which can easily make the recommendation results outdated or irrelevant.

[0006] (2) Insufficient privacy protection: The existing centralized data collection and storage model easily causes users to worry about personal privacy leakage, especially when processing users' real-time location or historical location data. This model fails to fully protect users' sensitive information and increases the risk of data abuse or leakage.

[0007] (3) Computing 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 inefficient processing, 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 the present invention is to provide a POI prediction method and architecture with privacy protection. To address the single-factor dependency problem existing in existing POI prediction methods, four influencing factors, namely social relationships, user similarity, user intentions and spatiotemporal contexts, are comprehensively considered, and deep learning technology is integrated to alleviate the data sparsity problem and achieve high-precision recommendations.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] In one aspect, the present invention provides a method for predicting points of interest with privacy protection, comprising the following steps:

[0011] S1: Collect user data using anonymization technology;

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

[0013] S3: Participants preprocessed and extracted features from the pooled data;

[0014] S4: Participants used a multi-factor fusion personalized POI recommendation method to model four key factors: social relationship, user similarity, user intent, and spatiotemporal context, to train a local model;

[0015] S5: Participants perturb the original gradients by adding pairwise masks and upload the encrypted gradients;

[0016] S6: The computing service provider aggregates the received encrypted information and sends the global model update information to each participant. The participant receives and updates the local model and repeats the above process until the model converges.

[0017] Furthermore, the user data is collected by using anonymization technology in step S1, specifically including:

[0018] Collect data related to POI predictions 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 the prediction of points of interest include the user's 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, the precise coordinates are converted into an area range through fuzzy processing or generalization technology; for identity information, the device identifier and user name are completely removed or irreversibly deformed to ensure that the collected data cannot be traced back to a specific user directly or indirectly.

[0022] Furthermore, the participant in step S3 pre-processes the aggregated data, specifically including: cleaning the aggregated data, including removing outliers and duplicate data;

[0023] The feature extraction is to extract features that are valuable for predicting points of interest, including working or weekends, different time periods, common activity areas, and regional attributes.

[0024] Furthermore, the modeling described in step S4 is based on four key factors: social relationship, user similarity, user intention and spatiotemporal context, specifically including:

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

[0026]

[0027] For user similarity modeling, we first calculate the collaborative similarity S of two users by combining the similarity method. uv ; Given a user's similarity threshold χ u , if and only if S uv > u , user u and user v will be defined as similar users; then the graph attention mechanism is used, for user u, given the similar neighbor node features and the neighbor set N R (u), α uv is the weight from user u to neighbor v, and the weighted sum of neighbors is obtained to get the similar user node representation

[0028] User intent modeling includes:

[0029] 1) First, we introduce the knowledge graph, classify the labels according to the semantic information, and use the graph convolutional network to update the labels. Then, we adopt a multimodal fusion strategy to consider the semantic information of text, image and video modalities for conversion and representation, 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 are the attention weights of text, image and video modalities respectively, F is a multi-layer perceptron, v new is the fused vector;

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

[0034] 3) Use the generative adversarial network to optimize the parameters of the fully connected layer and the activation function. Suppose the input of the generator is a random noise vector z, and the generated vector is v gen , the discriminator is used to identify the real user intention vector v pool and generate vector v gen The outputs are D(v pool ) and D(v gen ); the training goal is to minimize the following loss function:

[0035]

[0036] For spatiotemporal context modeling, the user’s check-in behavior is affected by the specific time point t k 、Geographical location j =(long j ,lat j ), whether it is a holiday Distance from last check-in and the time difference τ kThe influence of the five factors, firstly, the one-hot encoding method is used to convert the five factors affecting each sign-in into a d-dimensional vector, and the temporary embedding vector at each moment is aggregated into in, is a learnable parameter; using the contextual attention mechanism, the user k The embedding vector of the moment check-in record is expressed as:

[0037]

[0038] in, is the embedding vector of the i-th feature factor after the attention mechanism, The embedding vector representation of the i-th factor of the k-th historical check-in, is the corresponding weight, is a learnable parameter; Input into LSTM and get t k Hidden state at the moment Generating weights through temporal attention mechanism Finally, we get the user's spatiotemporal preference representation

[0039] After obtaining the embedding vectors of the four influencing factors, we first aggregate the different embedding vectors into

[0040]

[0041] Calculate its embedding vector with a specific POI The inner product of to predict the user's rating of POI 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 Represents the feature vector of the training sample, y k is the corresponding label.

[0043] Furthermore, in step S5, the participant perturbs the original gradient by adding paired masks and uploads the encrypted gradient, which specifically includes the following steps:

[0044] S51: Participants compute local gradients

[0045] S52: Participants mask the local gradient and perturb the original gradient by adding paired masks:

[0046]

[0047] Among them, τ i is a random noise term, α ij is a coefficient that controls the influence of the gradient of the jth participant on the mask of the ith participant.

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

[0049] S61: The computing service provider collects and aggregates the disturbance gradients sent Decrypt using the private key to obtain the global model x' agg ;

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

[0051] S63: After receiving the global model update information sent by the computing service provider, the participant updates the local model;

[0052] S64: repeating steps S61-S63, i.e. the participant continuously trains the local model, generates new encrypted update information and sends it to the computing service provider, and the computing service provider aggregates and sends the update information again until the model reaches a convergence state;

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

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

[0055] Collection layer: Use anonymization technology to collect user data and upload the user data to the POI service provider in the aggregation layer, so that the POI service provider and the computing service provider can jointly train a point of interest prediction model;

[0056] Aggregation layer: Each participant receives anonymized user data from the collection layer and aggregates the data within a specific service area. The participants use the aggregated user data to train the POI prediction model. After the model training is completed, the encrypted model gradient is sent to the computing service provider.

[0057] Computing layer: In each training cycle, the computing service provider receives the local training results of each participant in the current round, and globally aggregates the uploaded encrypted gradients. After decryption, the aggregated latest global model is broadcast to each participant.

[0058] The beneficial effects of the present invention are as follows: the present invention significantly reduces the risk of sensitive data leakage by locally processing anonymous data at the aggregation layer and only transmitting encrypted model updates or gradient information to the computing layer, thereby strengthening the protection of user privacy. In addition, by processing user data at the aggregation layer, the present invention effectively reduces the computational burden of the collection layer and the computing layer. Through deep learning technology, the present invention considers four dimensions, namely social relationships, user similarities, user intentions, and spatiotemporal context, to achieve personalized modeling, which not only improves the ability to accurately capture the dynamics of user behavior, but also effectively alleviates the problem of data sparsity, and improves the accuracy of the recommendation system and user satisfaction.

[0059] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0061] Figure 1 This is a system architecture diagram of the present invention;

[0062] Figure 2 A personalized POI recommendation method based on multi-factor fusion designed for the present invention. DETAILED DESCRIPTION

[0063] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways 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 only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0064] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0065] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present invention difficult to understand.

[0066] The present invention provides a method for predicting points of interest with privacy protection. Figure 1 As shown, the architecture adopted by this method is a three-layer architecture of collection layer-aggregation layer-computing layer.

[0067] Collection layer: Use anonymization technology to collect user data and upload the user data to the POI service provider in the aggregation layer, so that the POI service provider and the computing service provider can jointly train an accurate POI prediction model.

[0068] Aggregation layer: First, the aggregation layer has multiple POI service providers, that is, participants in federated learning. Each participant receives anonymized user data from the collection layer and aggregates this data within a specific service area. Secondly, participants use the aggregated user data to train the POI prediction model. After the model training is completed, the encrypted model gradient is sent to the computing service provider to ensure the confidentiality of the data during transmission.

[0069] Computing layer: In each training cycle, the computing service provider receives the local training results of each participant in the current round, and globally aggregates the uploaded encrypted gradients. After decryption, the aggregated latest global model is broadcast to each participant.

[0070] The sharing of model knowledge is achieved through the collaboration of the aggregation layer and the computing layer. In this architecture, users anonymously upload data to the aggregation layer, and the data is only processed in the aggregation layer, while the encrypted model update or gradient information is sent to the computing layer for aggregation. This design not only reduces the risk of data leakage, but also effectively reduces the computational burden of the computing layer and the acquisition layer by performing local model training in the aggregation layer.

[0071] Figure 2 This is a personalized POI recommendation method that integrates multiple factors and combines deep learning technology. Personalized modeling is implemented for the data characteristics of the four influencing factors, namely social relationships, user similarity, user intent, and spatiotemporal context. This not only improves the ability to accurately capture the dynamics of user behavior, but also effectively alleviates the problem of data sparsity, improving the accuracy of the recommendation system and user satisfaction.

[0072] The method comprises the following steps:

[0073] S1: Collect user data using anonymization technology;

[0074] First, collect user data, including but not limited to user location information, identity information, browsing history, search keywords, application usage frequency, and other data that may be related to POI prediction. When collecting data, use anonymization technologies such as k-anonymity, l-diversity, or differential privacy to process the data in real time to remove or blur information that can directly or indirectly identify the user. For location information, use blurring or generalization technology to convert precise coordinates into a larger geographic area; for identity information, completely remove or irreversibly deform device identifiers, user names, and other information to ensure that the collected data cannot be directly or indirectly traced back to specific users.

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

[0076] S3: Participants preprocessed and extracted features from the pooled data;

[0077] The aggregated data is cleaned, including removing outliers and duplicate data; then feature extraction is performed, including features that are valuable for POI prediction, such as working / weekends, different time periods, common activity areas, and regional attributes.

[0078] S4: Participants used a multi-factor fusion personalized POI recommendation method to model four key factors, including social relationships, user similarity, user intent, and spatiotemporal context, to train a local model;

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

[0080] S42: For similar user modeling, firstly, the collaborative similarity S of two users is calculated by combining the similarity method. uv Given a user similarity threshold χ u , if and only if S uv > u , user u and user v will be defined as similar users. Then the graph attention mechanism is used, for user u, given the similar neighbor node features and the neighbor set N R (u), α uv is the weight from user u to neighbor v, and the weighted sum of neighbors is obtained to get the similar user node representation

[0081] S43: For user intent modeling, we first introduce the knowledge graph, classify the labels according to the semantic information, and use the graph convolutional network to update the labels. Then, we adopt a multimodal fusion strategy to consider the semantic information of text, image and video modalities for conversion 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 are the attention weights of text, image and video modalities respectively, F is a multi-layer perceptron, v new is the fused vector.

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

[0086] S46: Use generative adversarial networks to optimize the parameters of the fully connected layer and activation function. Suppose the input of the generator is a random noise vector z, and the generated vector is v gen , the discriminator is used to identify the real user intention vector v pool and generate vector v gen The outputs are D(v pool ) and D(v gen ). The training goal is to minimize the following loss function:

[0087]

[0088] S47: For spatiotemporal context modeling, the user’s check-in behavior is affected by the specific time point t k 、Geographical location j =(long j ,lat j ), whether it is a holiday Distance from last check-in and the time difference τ k The influence of the five factors, firstly, the one-hot encoding method is used to convert the five factors affecting each sign-in into a d-dimensional vector, and the temporary embedding vector at each moment is aggregated into in, are learnable parameters.

[0089] S48: Since the degree to which each factor affects user behavior varies, a contextual attention mechanism is used. k The embedding vector of the moment check-in record is expressed as:

[0090]

[0091] in, is the embedding vector of the i-th feature factor after the attention mechanism, The embedding vector representation of the i-th factor of the k-th historical check-in, is the corresponding weight, is a learnable parameter. Input into LSTM and get t k Hidden state at the moment Generating weights through temporal attention mechanism Finally, we get the user's spatiotemporal preference representation

[0092]

[0093] S49: After obtaining the embedding vectors of the four influencing factors, the different embedding vectors are first aggregated into Calculate its embedding vector with a specific POI The inner product of to predict the user's rating of POI And calculate the loss function

[0094] Participants calculate the loss value of the sample Then iteratively minimize the loss to obtain the optimal model. k Represents the feature vector of the training sample, y k is the corresponding label.

[0095] S5: Participants perturb the original gradients by adding pairwise masks and upload the encrypted gradients;

[0096] Participants first compute local gradients Then the participants masked the local gradients and perturbed the original gradients by adding paired masks. Among them, τ i is a random noise term, α ij is a coefficient that controls the influence of the gradient of the jth participant on the mask of the ith participant.

[0097] S6: The computing service provider aggregates the received encrypted information and sends the global model update information to each participant. The participant receives and updates the local model and repeats the above process until the model converges.

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

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

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

[0101] S64: The above process is repeated continuously, that is, the participant continues to train the local model, generates new encrypted update information and sends it to the computing service provider, and 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 for them.

[0103] In the above embodiments, the description's reference to "this embodiment" indicates that a particular feature, structure, or characteristic described in conjunction with the embodiment is included in at least some embodiments, but not necessarily all embodiments. Multiple occurrences of "this embodiment" do not necessarily all refer to the same embodiment.

[0104] In the above-described embodiments, although the invention has been described in conjunction with specific embodiments of the invention, many substitutions, modifications, and variations of these embodiments will be apparent to those of ordinary skill in the art based on the foregoing description. For example, other storage structures (e.g., dynamic RAM (DRAM)) may use the embodiments discussed. Embodiments of the invention are intended to encompass all such substitutions, modifications, and variations that fall within the broad scope of the appended claims.

[0105] This embodiment further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, any one of the methods in this embodiment is implemented.

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

[0107] The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the terminal executes any one of the methods in this embodiment.

[0108] The computer-readable storage medium in this embodiment can be understood by ordinary technicians in this field: all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to the computer program. The aforementioned computer program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk and other media that can store program codes.

[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 with each other. The memory is used to store computer programs, the communication interface is used to communicate, and the processor and the transceiver are used to run computer programs so that the electronic terminal executes each step of the above method.

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

[0111] The above-mentioned processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0112] The present invention can be used in many general or special computing system environments or configurations, such as personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and the like.

[0113] The present invention may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located 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 solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A privacy-preserving point of interest prediction method, characterized by: The following steps are involved: S1: Collect user data using anonymization technology; S2: Participants categorize and aggregate data according to pre-defined service areas; S3: Participants preprocessed and extracted features from the pooled data; S4: Participants used a multi-factor fusion personalized POI recommendation method to model four key factors: social relationship, user similarity, user intent, and spatiotemporal context, to train a local model; S5: Participants perturb the original gradients by adding pairwise masks and upload the encrypted gradients; S6: The computing service provider aggregates the received encrypted information and sends the global model update information to each participant. The participant receives and updates the local model and repeats the above process until the model converges.

2. The method for predicting points of interest with privacy protection according to claim 1, characterized in that: The use of anonymization technology to collect user data in step S1 specifically includes: Collect data related to POI predictions 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 the prediction of points of interest include the user's location information, identity information, browsing history, search keywords, and application usage frequency; The anonymization techniques include k-anonymity, l-diversity and differential privacy; For location information, the precise coordinates are converted into an area range through fuzzy processing or generalization technology; for identity information, the device identifier and user name are completely removed or irreversibly deformed to ensure that the collected data cannot be traced back to a specific user directly or indirectly.

3. The method for predicting points of interest with privacy protection according to claim 1, characterized in that: In step S3, the participants pre-process the aggregated data, specifically including: cleaning the aggregated data, including removing outliers and duplicate data; The feature extraction is to extract features that are valuable for predicting points of interest, including working or weekends, different time periods, common activity areas, and regional attributes.

4. The method for predicting points of interest with privacy protection according to claim 1, characterized in that: Step S4 describes modeling based on four key factors: social relationship, user similarity, user intention, and spatiotemporal context, specifically including: For social relationship modeling, a graph attention mechanism is used. Given a social neighbor node representation and the neighbor set N S (u), α uj is the attention weight from user u to neighbor j, and the weighted sum of neighbors is obtained to obtain the user node representation with social relationship For user similarity modeling, we first calculate the collaborative similarity S of two users by combining the similarity method. uv ; Given a user's similarity threshold χ u , if and only if S uv > u , user u and user v will be defined as similar users; then the graph attention mechanism is used, for user u, given the similar neighbor node features and the neighbor set N R (u), α uv is the weight from user u to neighbor v, and the weighted sum of neighbors is obtained to get the similar user node representation User intent modeling includes: 1) First, we introduce the knowledge graph, classify the labels according to the semantic information, and use the graph convolutional network to update the labels. Then, we adopt a multimodal fusion strategy to consider the semantic information of text, image and video modalities for conversion and representation, as shown below: v new =F(A(T v ,I v ,S v ),T v ,I v ,S v ) =RELU(W2·RELU(W1·[a T T v +a I I v +a S S v ]+b1)+b2) Where A(T v ,I v ,S v )=(a T ,a I ,a S ), a T ,a I ,a S are the attention weights of text, image and video modalities respectively, F is a multi-layer perceptron, v new is the fused vector; 2) Process the sequence data through LSTM, let H = [h1,h2,…,h T ] is the output matrix of LSTM. An adaptive multi-pooling strategy is performed on the output matrix of LSTM. A selection function p(H) is defined, which determines the pooling method according to the eigenvalue distribution of H, and then the pooled vector is obtained through the selected pooling method; 3) Use the generative adversarial network to optimize the parameters of the fully connected layer and the activation function; let the input of the generator be the random noise vector z, and the generated vector be v gen , the discriminator is used to identify the real user intention vector v pool and generate vector v gen The outputs are D(v pool ) and D(v gen ); the training goal is to minimize the following loss function: For spatiotemporal context modeling, the user’s check-in behavior is affected by the specific time point t k 、Geographical location j =(long j ,lat j ), whether it is a holiday Distance from last check-in and the time difference τ k The influence of the five factors, firstly, the one-hot encoding method is used to convert the five factors affecting each sign-in into a d-dimensional vector, and the temporary embedding vector at each moment is aggregated into in, is a learnable parameter; using the contextual attention mechanism, the user k The embedding vector of the moment check-in record is expressed as: in, is the embedding vector of the i-th feature factor after the attention mechanism, The embedding vector representation of the i-th factor of the k-th historical check-in, is the corresponding weight, is a learnable parameter; Input into LSTM and get t k Hidden state at the moment Generating weights through temporal attention mechanism Finally, we get the user's spatiotemporal preference representation After obtaining the embedding vectors of the four influencing factors, we first aggregate the different embedding vectors into Calculate its embedding vector with a specific POI The inner product of to predict the user's rating of POI And calculate the loss function Participants calculate the loss value of the sample Then iteratively minimize the loss to obtain the optimal model, where x k Represents the feature vector of the training sample, y k is the corresponding label.

5. The method for predicting points of interest with privacy protection according to claim 1, characterized in that: In step S5, the participant perturbs the original gradient by adding paired masks and uploads the encrypted gradient, which specifically includes the following steps: S51: Participants compute local gradients S52: Participants mask the local gradient and perturb the original gradient by adding paired masks: Among them, τ i is a random noise term, α ij is a coefficient that controls the influence of the gradient of the jth participant on the mask of the ith participant.

6. The method for predicting points of interest with privacy protection according to claim 1, characterized in that: Step S6 specifically includes the following steps: S61: The computing service provider collects and aggregates the disturbance gradients sent Decrypt using the private key to obtain the global model x' agg ; S62: The computing service provider calculates and updates the global model: S63: After receiving the global model update information sent by the computing service provider, the participant updates the local model; S64: repeating steps S61-S63, i.e. the participant continuously trains the local model, generates new encrypted update information and sends it to the computing service provider, and the computing service provider aggregates and sends the update information again until the model reaches a convergence state; S65: Participants distribute the final trained global model to users to provide personalized POI recommendations for them.

7. A privacy-preserving point of interest prediction system, characterized by: Includes collection layer, aggregation layer, and computing layer; Collection layer: Use anonymization technology to collect user data and upload the user data to the POI service provider in the aggregation layer, so that the POI service provider and the computing service provider can jointly train a point of interest prediction model; Aggregation layer: Each participant receives anonymized user data from the collection layer and aggregates the data within a specific service area. The participants use the aggregated user data to train the POI prediction model. After the model training is completed, the encrypted model gradient is sent to the computing service provider. Computing layer: In each training cycle, the computing service provider receives the local training results of each participant in the current round, and globally aggregates the uploaded encrypted gradients. After decryption, the aggregated latest global model is broadcast to each participant.

Images