A personalized data service method

Abstract

The application relates to a personalized data service method, and belongs to the field of Internet of Things. The method comprises the following steps: region division; data collection: an edge server collects and stores historical search request and evaluation feedback data of users in a jurisdictional region; model construction: a personalized search model with personal-global level semantic enhancement is constructed by taking the historical search request record of a user as data input; joint inference: a federal learning framework is cooperatively constructed by the edge and cloud servers to jointly train the personalized search model, and the trained model is sent to each edge server for storage; search initiation: a user initiates a search request to the edge server through a client device; search response: the edge server performs a search process and matches a plurality of related entities, and then calls the personalized search model to reorder the results of the plurality of entities and returns the results to the user. The application can effectively improve the precision of user-oriented data services and provide personalized and high-quality search service experience for users.

Description

Technical Field

[0001] This invention belongs to the field of the Internet of Things and relates to a method for providing personalized data services. Background Technology

[0002] IoT search technology oriented towards entity status information can help users quickly obtain relevant entity data of interest. This technology can narrow the search scope in the vast cyberspace and provide users with convenient, fast, and high-quality data services. It has become one of the indispensable key technologies in the development of the Internet of Things.

[0003] Since the introduction of IoT search technology, a large number of researchers have devoted their research efforts to it, mainly in the following aspects: (1) Prediction mechanism for data search services: For time series composed of historical entity state data, design advanced time series prediction methods and upload the predicted entity state data from the perception layer to the edge server or cloud server in advance, which can quickly estimate relevant results for users and reduce system communication overhead; (2) Caching strategy for data search services: For edge or cloud server cache capacity limited, design high-performance caching methods to maximize cache hit rate, maximize user service experience and improve system service quality; (3) Fast search method in data services: In order to adapt to the high-speed dynamic changes of entity state data, two aspects are considered to ensure the real-time performance of data services. One is the design of service architecture, introducing edge computing and using the advantage of edge servers being close to the local area to quickly respond to user needs. The other is the design of indexing methods, establishing a fast indexing mechanism based on entity characteristics to accelerate search time; (4) Guarantee of search accuracy in data services: Study the correlation matching method between search requests and entity information to provide users with accurate and satisfactory search results, which can improve the data service quality of the system and fully meet user needs. (5) Privacy protection in data services: Based on users’ increasing attention to data privacy, design a data service mechanism in a privacy environment and perform search and entity matching.

[0004] The following are some research works on data services. Zhang P, Li X, Liu Y, Kang X, and Liu Y, in “SDU: State-Based Dual-Mode Sensor Search Mechanism Toward Internet of Things” [in IEEE Access, vol. 7, pp. 147962-147974, 2019], address the limitation of prediction accuracy in shallow learning theory. Based on deep learning theory, they propose a high-precision data prediction method for the current state of the sensor to improve prediction accuracy. Zhang P, Li X, Wu D, and Wang R, in “Edge-Cloud Collaborative EntityState Data Caching Strategy Toward Networking Search Service in CPSs” [in IEEE Transactions on Industrial Informatics, vol. 17, no. 10, pp. 6906-6915, Oct 2021], categorize IoT entity data into transient and slowly changing entity data by mining historical features and cache them on edge servers and cloud servers respectively. This implements a classification caching strategy based on the changing characteristics of IoT data, effectively ensuring that users can obtain entity information in real time. Zhang P, Chui Y, Liu H, Yang Z, Wu D, and Wang R, in “Efficient and Privacy-Preserving Search over Edge-Cloud Collaborative Entity in IoT” [in IEEE Internet of Things Journal, doi:10.1109 / JIOT.2021.3132910.], achieved security and privacy in the data service process using a symmetric searchable encryption mechanism and designed an indexing method based on a clustered balanced binary tree to significantly accelerate the search time.

[0005] While existing research on IoT search for data services has achieved high performance in areas such as prediction mechanisms, caching strategies, search time, and privacy protection, search accuracy still needs improvement. Traditional relevance-based matching methods can only guarantee consistent feedback results for all user needs, neglecting each user's personalized needs and search intent, thus failing to deliver a significant leap in user experience. Personalized search methods can leverage emerging deep learning models to uncover deep-seated user preferences hidden in historical request records. These preferences help accurately understand the user's current search intent, and combined with ranking models, personalized ranking is achieved, providing the most relevant results to the user. However, this method requires the use of large amounts of historical user data, which undoubtedly poses a significant threat to user privacy. Therefore, it is necessary to design personalized data service methods while ensuring user security and privacy, comprehensively improving the quality of data services. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a personalized data service method. First, it proposes a personalized search model with personal-global semantic enhancement to achieve accurate user interest modeling and preference perception. Then, it introduces an edge-cloud collaborative system model to construct a federated learning training framework. Federated learning technology is used to achieve joint collaborative inference on the personalized search models of all users. The models obtained after joint inference are then deployed to edge servers to ensure personalized data services for all users. This method can improve entity search accuracy while protecting user data privacy, achieving high-quality, secure, and personalized data services for users.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A personalized data service method specifically includes the following steps:

[0009] S1: Region Division: Each edge server is divided into regions to manage based on its coverage capabilities, and manages users and physical entities within the regions;

[0010] S2: Data Collection: Edge servers collect and store historical search request records and evaluation feedback data of users within their jurisdiction;

[0011] S3: Model Building: Construct a personalized search model with personal-global semantic enhancement, including: using users' historical search request records and evaluation feedback data as input to the Transformer model to obtain three implicit features contained in the input data: personal-level semantic disambiguation query features q p Personal-level user interest and preference features p and global-level entity representation e U; Use three implicit features to obtain the semantically enhanced current query And calculate the current query using cosine similarity. Personalized matching score with entity e This allows for personalized result sorting;

[0012] S4: Joint Inference: Edge servers and cloud servers form an edge-cloud collaborative model to perform joint inference on the model. Through the joint inference of edge-cloud collaboration, the personalized search model is deployed on all edge servers.

[0013] S5: Initiate a search: The user initiates a search request for entity status data to the edge server through their client device;

[0014] S6: Search Response: The edge server or cloud server filters out relevant matching entity results. The edge server then reorders the feedback results using a stored personalized search model before returning them to users within the managed area.

[0015] Furthermore, in step S3, the Transformer model consists of multiple multi-head attention mechanisms and a position feedforward network, as shown below:

[0016] Transformer(x) = LN(M) x +D(PF(M x )))

[0017] Among them, M x = LN(x+D(MS(x))), where LN(·) is the batch normalization layer for the output data, and D(·) is the dropout layer to prevent overfitting; the multi-head attention mechanism MS(x) in the model is represented as MS(x) = [head1, ..., head1] i , ..., head h W O Each head i =Att(xW i Q xW i K xW i V Attention is x represents the data input to the model, W O W i Q W i K W i VThis represents the parameters that need to be updated in the model; Q, K, and V are three important pieces of information in the Transformer, represented as the data input vector, the vector of the correlation between the information of interest and other information, and the vector of the information of interest, respectively; the position feedforward network is represented as PF(x) = C2(ReLU(C1(x)). T ))) T Where C2(·) and C1(·) are two convolution operations with different parameters, x T ReLU represents the transpose operation of the input data vector, and ReLU(·) is a non-linear activation function.

[0018] Furthermore, in step S3, the constructed personalized search model with personal-global semantic enhancement specifically includes the following steps:

[0019] S31: Perform personal-level query disambiguation using the Transformer model, specifically including: replacing the input data x with the query q containing the current time. T The query sequence q = {q1, q2, ..., q} t , ..., q T-1 q T}, to obtain personal-level semantic disambiguation query features This indicates the positional embedding of each query word in the query sequence;

[0020] Using the Transformer model to mine personal interests and preferences, specifically including: replacing the input data x with the user's historical search records H = {h1, h2, ..., h...} t , ..., h T-1}, h t ={q t r t}, r = {r1, r2, ..., r t ,...,r T-1} represents the evaluation feedback corresponding to each historical search request of the user, indicating the user service experience. Obtain the personal-level user interest and preference features p = Transformer(H + H) p+ ), H p+ This represents the positional embedding of each record tuple consisting of a search request and its corresponding evaluation feedback in the historical search records.

[0021] Global entity representation is achieved using the Transformer model, replacing the input data x with the rating feedback R = {r1, r2, ..., r} of all users for a given entity. u ,...,r U}, obtain the global entity representation e U e U =Transformer(R+R)p+ ), R p+ This represents the location embedding corresponding to all user reviews and feedback for an entity;

[0022] S32: Utilize the three hidden features q obtained in step S31 p p, e U Enhance the current query q T The semantically enhanced current query representation obtained is MLP(·) represents a multilayer perceptron model, which is responsible for fusing these feature vectors together;

[0023] S33: Calculate user-enhanced current query using cosine similarity. Personalized matching score with entity e This completes the construction of personalized search to achieve personalized result ranking, represented as:

[0024] Furthermore, step S4 specifically includes the following steps:

[0025] S41: Joint inference of the edge-cloud collaborative model requires multiple rounds of iterative updates to the personalized search model. During each round of training, each edge server completes the training of each model locally.

[0026] Local training on edge servers: The specific process is represented as follows M t Let t represent the personalized search model parameters of the edge server during the training round at time t, x represent the input data required for the extraction of the three latent features in step S3, and SGD(·) represent the stochastic gradient descent method used for model iterative updates. This represents the parameters of the personalized search model trained on the k-th edge server at time t+1;

[0027] S42: Edge Server Parameter Upload: The model parameters will be updated for each edge server. Uploaded to the cloud server;

[0028] S43: Cloud Server Parameter Aggregation: In each round of joint inference, the cloud server aggregates the model parameters uploaded by each edge server. Compared to the model parameters M in the previous round t The aggregation model parameters at time t+1 are expressed as follows: K represents the total number of edge servers;

[0029] S44: Cloud server parameter distribution: The cloud server will distribute the new round of aggregated model parameters M t+1 Send to each edge server;

[0030] S45: Edge-Cloud Collaborative Joint Inference: In the joint inference and training model, the loss function Loss is defined as:

[0031]

[0032] in, This represents the relevance score between the search request and the target entity. This represents the relevance score of the search request to other candidate entities, and the optimization objective of joint inference is to minimize the objective function: Where Θ k This represents all parameters of the personalized model in the k-th edge server;

[0033] S46: Model Acquisition: Repeat steps S41-S44 until the objective function in step S45 is minimized and stabilized. At this point, each edge server can obtain the optimal personalized search model.

[0034] Furthermore, step S6 specifically includes the following steps:

[0035] S61: The edge server determines the cache location of the entity state data required by the user. If the required entity state data is cached on the edge server, the edge server performs a search based on the traditional retrieval matching method, filters out relevant entity results, and then runs the trained personalized search model to reorder the entity results in a personalized way so that the entity results most satisfactory to the user appear at the top of the sort list. Finally, the personalized entity result list is fed back to the user.

[0036] S62: If the required entity state data is not cached on the edge server, the edge server will encrypt the search request and upload it to the cloud server. The cloud server will perform a general search request and entity relevance matching in the encrypted environment based on the symmetric searchable encryption mechanism, and then feed back the encrypted entity result list to the corresponding edge server. The edge server will decrypt the entity result list and run the trained personalized search model to reorder the entity results in a personalized way, and can also return personalized sorting results to the user.

[0037] The beneficial effects of this invention are as follows: First, it proposes a personalized search model with personal-global semantic enhancement to accurately model user interests, thereby enabling personalized data service functions. Furthermore, to ensure user privacy during the use of user data, an edge-cloud collaborative system model is introduced to construct a federated learning training framework. Based on federated learning, joint collaborative inference is achieved on the personalized search model, enabling personalized data services while protecting user privacy, comprehensively improving data service quality, and maximizing the user service experience.

[0038] 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

[0039] 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:

[0040] Figure 1 This is a block diagram of the personalized data service system of the present invention;

[0041] Figure 2 A schematic diagram of the personalized search model with personal-global semantic enhancement designed for this invention;

[0042] Figure 3 The flowchart of the joint inference process of the edge-cloud collaborative model based on federated learning designed for this invention is shown. Detailed Implementation

[0043] 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.

[0044] Please see Figures 1-3 , Figure 1 The diagram shown is a block diagram of the personalized data service system involved in this invention. The system comprises the following modules:

[0045] Users: Ignoring user mobility, each user has a fixed geographical location. Users are assigned to a single edge server for management, and submit search requests to the edge server through their client. Due to factors such as user background, interests, occupation, and surrounding environment, even when sending the same search request for entity status data, each user's search intent will be different. This phenomenon can be called personalized user needs.

[0046] Entities: Located in the sensing layer of the Internet of Things (IoT), these consist of physical entities and their attached sensing devices. The sensing devices collect status data of the entities (such as temperature, humidity, and air quality) to provide a data source for data services.

[0047] Edge servers: Edge servers are devices with strong communication, storage, and computing capabilities. They receive entity data uploaded by sensing devices, manage user and entity information, execute data search services, and respond to user requests. They also support computationally intensive model building and participate in joint model inference, train personalized models, and enable personalized data services.

[0048] Cloud servers: Cloud servers are powerful central devices with strong communication, storage, and computing capabilities. They are responsible for storing encrypted entity state data, accepting encrypted search requests, and performing general data services in an encrypted environment. Furthermore, they also serve as central servers, working together with edge servers for joint model training.

[0049] This invention provides a personalized data service method, specifically including the following steps:

[0050] Step 1: Region Division: Each edge server is assigned a region to manage based on its coverage capabilities, and manages the users and physical entities within that region.

[0051] Step 2: Data Collection: Each edge server stores the historical search records H = {h1, h2, ..., h...} for each user it manages. t , ..., h T-1}, h t ={q t r t}, where q = {q1, q2, ..., q} t , ..., q T-1}, r = {r1, r2, ..., r t ,...,r T-1} represent search request records and evaluation feedback data sorted by time, respectively, using q T This represents the user's search request at the current moment, and this data will serve as input data for the model's inference in the next step.

[0052] Step 3: Model Building: The edge server collects all users' historical search request records and evaluation feedback H = {h1, h2, ..., h...} t , ..., h T-1}, h t ={q t r t As input, three implicit features are obtained from the input data through a personalized search model with personal-global semantic enhancement: personal-level semantic disambiguation query feature q. p Personal-level interests and preferences (p), global-level entity representation (e) U .like Figure 2As shown, the specific steps include:

[0053] Step 3.1: Three hidden features q p p and e U All are extracted by the deep learning model Transformer, which consists of multiple multi-head attention mechanisms and position feedforward networks, and is represented as Transformer(x) = LN(M). x +D(PF(M x ))), M x = LN(x+D(MS(x))). Where LN(·) is the batch normalization layer for the output data, and D(·) is the dropout layer to prevent overfitting. The multi-head attention mechanism MS(x) in the model is represented as MS(x) = [head1, ..., head2]. i , ..., head h W O Each head i =Att(xW i Q xW i K xW i V Attention is x represents the data input to the model, W O W i Q W i K W i V This represents the parameters that need to be updated in the model. Q, K, and V are three important pieces of information in the Transformer, represented as the data input vector, the vector representing the correlation between the information of interest and other information, and the vector representing the information of interest. The positional feedforward network is represented as PF(x) = C2(ReLU(C1(x)). T ))) T C2(·) and C1(·) are two convolution operations with different parameters, x T ReLU represents the transpose operation of the input data vector, and ReLU(·) is a non-linear activation function.

[0054] Step 3.2: Due to different user search contexts, the same query can also have semantic ambiguity. When the edge server performs model training, it uses the Transformer model to perform personal-level query disambiguation, replacing the input data x with the query q containing the current time. T The query sequence q = {q1, q2, ..., q} t , ..., q T-1 qT}, to obtain personal-level semantic disambiguation query features This indicates the positional embedding of each query word in the query sequence.

[0055] Step 3.3: Since the personal-level query disambiguation method in Step 3.2 can only capture explicit keywords and cannot understand implicit or incomplete search requests, it is necessary to further mine user interests and preferences based on the user's search history. Similarly, the Transformer model is used to complete the personal-level interest and preference mining, replacing the input data x with the user's historical search records H = {h1, h2, ..., h...}. t , ..., h T-1}, h t ={q t e t}, obtain personal user interest and preference features p = Transformer(H+H p+ ), H p+ This represents the positional embedding of each record tuple consisting of a search request and its corresponding evaluation feedback in the historical search records.

[0056] Step 3.4: In steps 3.2 and 3.3, only individual-level query disambiguation and interest preference mining were performed from the perspective of a single user. Due to the limited understanding of entities by individuals and the evolution of personal interests over time, extracting features only from the individual level cannot fully represent a globally fair knowledge representation, ultimately leading to poor personalization performance. Therefore, enhanced feature representations should be performed for all physical entities to fully obtain global entity descriptions and adapt to changes in user interests. Similarly, a global-level entity representation is achieved using a Transformer, replacing the input data x with the evaluation feedback R = {r1, r2, ..., r...} of all users for an entity. u ,...,r U}, obtain the global entity representation e U e U =Transformer(R+R) p+ ), R p+ This represents the location embedding corresponding to all user reviews and feedback for an entity.

[0057] Step 3.5: Utilize the three latent features q obtained in steps 3.2, 3.3, and 3.4 p p, e U Enhance the current query q T The semantically enhanced query representation obtained is MLP(·) represents a multilayer perceptron model, which is responsible for fusing these feature vectors together.

[0058] Step 3.6: Calculate the user-enhanced current query using cosine similarity. The personalized matching score with entity e is used to complete the personalized search construction and achieve personalized result ranking, as shown below:

[0059]

[0060] Step 4: Joint Inference: Since the personalized model construction in Step 3 requires a large amount of user historical data, it poses a risk of user privacy leakage. Cloud servers are generally considered untrusted, so the model cannot be trained directly in the cloud. Therefore, edge servers and cloud servers are combined to form an edge-cloud collaborative model for joint inference. This method allows user historical data to be trained on local edge servers, with only the model parameters uploaded to the cloud server, effectively protecting user privacy while enabling personalized data services. The personalized search model is deployed on all edge servers through edge-cloud collaborative joint inference to achieve personalized search functionality. For example... Figure 3 As shown, the specific steps include:

[0061] Step 4.1: Joint inference for edge-cloud collaboration requires multiple rounds of iterative updates to the personalized search model. During each round of training, each edge server completes the training of the model locally.

[0062] Local training on edge servers: The specific process is represented as follows M t Let t represent the personalized search model parameters of the edge server during the training round at time t, x represent the input data required for the three feature extractions in step 3, and SGD(.) represent the stochastic gradient descent method used for model iterative updates. This represents the parameters of the personalized search model trained on the k-th edge server at time t+1.

[0063] Step 4.2: Edge Server Parameter Upload: The updated model parameters will be uploaded to each edge server. Uploaded to the cloud server.

[0064] Step 4.3: Cloud Server Parameter Aggregation: In each round of joint inference, the cloud server aggregates the model parameters uploaded by each edge server. Compared to the model parameters M in the previous round t The aggregation model parameters at time t+1 are expressed as follows: K represents the total number of edge servers.

[0065] Step 4.4: Cloud Server Parameter Distribution: The cloud server will distribute the new round of aggregated model parameters M. t+1 Send to each edge server.

[0066] Step 4.5: Edge-Cloud Collaborative Joint Inference: In joint inference and model training, the loss function Loss is defined as:

[0067]

[0068] in, This represents the relevance score between the search request and the target entity. This represents the relevance score of the search request to other candidate entities, and the optimization objective of joint inference is to minimize the objective function: Where Θ k This represents all parameters of the personalized model in the k-th edge server.

[0069] Step 4.6: Model Acquisition: Repeat steps 4.1-4.4 until the objective function in step 4.5 is minimized and stabilized. At this point, each edge server can obtain the optimal personalized search model.

[0070] Step 5: Initiate a search: The user initiates a search request for entity status data to the edge server through their client device.

[0071] Step 6: Search Response: The edge server or cloud server filters relevant matching entity results based on traditional retrieval and matching methods. The edge server then reorders the feedback results using a stored personalized search model before returning them to users within the managed area. This includes the following steps:

[0072] Step 6.1: The edge server determines the cache location of the entity state data required by the user. If the required entity state data is cached on the edge server, the edge server performs a search based on the traditional retrieval matching method, filters out relevant entity results, and then runs the trained personalized search model to reorder the entity results in a personalized way so that the entity results most satisfactory to the user appear at the top of the sort list. Finally, the personalized entity result list is fed back to the user.

[0073] Step 6.2: If the required entity state data is not cached on the edge server, the edge server encrypts the search request and uploads it to the cloud server. The cloud server performs a general search request and entity relevance matching in the encrypted environment based on a symmetric searchable encryption mechanism, and then feeds back the encrypted entity result list to the corresponding edge server. The edge server decrypts the entity result list and runs the trained personalized search model to reorder the entity results in a personalized manner, thus returning personalized sorting results to the user.

[0074] 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 personalized data service method, characterized in that, The method specifically includes the following steps: S1: Region Division: Each edge server is divided into regions to manage based on its coverage capabilities, and manages users and physical entities within the regions; S2: Data Collection: Edge servers collect and store historical search request records and evaluation feedback data of users within their jurisdiction; S3: Model Building: Construct a personalized search model with personal-global semantic enhancement, including: using users' historical search request records and evaluation feedback data as input to the Transformer model to obtain three implicit features contained in the input data: personal-level semantic disambiguation query features q p Personal-level user interest and preference features p and global-level entity representation e U ; Use three implicit features to obtain the semantically enhanced current query And calculate the current query using cosine similarity. Personalized matching score with entity e This allows for personalized result sorting; S4: Joint Inference: Edge servers and cloud servers form a federated learning-based edge-cloud collaborative model to perform joint inference. Through joint inference in edge-cloud collaboration, the personalized search model is deployed on all edge servers. S5: Initiate a search: The user initiates a search request for entity status data to the edge server through their client device; S6: Search Response: The edge server or cloud server filters out relevant matching entity results. The edge server then reorders the feedback results using a stored personalized search model before returning them to users within the managed area.

2. The personalized data service method according to claim 1, characterized in that, In step S3, the Transformer model consists of multiple multi-head attention mechanisms and a position feedforward network, as shown below: Transformer(x)=LN(M x +D(PF(M x ))) Among them, M x = LN(x+D(MS(x))), where LN(·) is the batch normalization layer for the output data, and D(·) is the dropout layer to prevent overfitting; the multi-head attention mechanism MS(x) in the model is represented as MS(x) = [head1,...,head1] i ,...,head h W O Each head i =Att(xW i Q ,xW i K ,xW i V Attention is x represents the data input to the model, W O W i Q W i K W i V This represents the parameters that need to be updated in the model; Q, K, and V are three important pieces of information in the Transformer, represented as the data input vector, the vector of the correlation between the information of interest and other information, and the vector of the information of interest, respectively; the position feedforward network is represented as PF(x) = C2(ReLU(C1(x)). T ))) T Where C2(·) and C1(·) are two convolution operations with different parameters, x T Represents the transpose operation of the input data vector, and ReLU(·) is a non-linear activation function.

3. The personalized data service method according to claim 2, characterized in that, In step S3, the constructed personalized search model with personal-global semantic enhancement specifically includes the following steps: S31: Perform personal-level query disambiguation using the Transformer model, specifically including: replacing the input data x with the query q containing the current time. T The query sequence q = {q1, q2, ..., q t ,...,q T -1,q T }, to obtain personal-level semantic disambiguation query features This indicates the positional embedding of each query word in the query sequence; Using the Transformer model to mine personal interests and preferences, specifically including: replacing the input data x with the user's historical search records H = {h1, h2, ..., h...} t ,...,h T-1 }, h t ={q t ,r t }, r={r1,r2,...,r t ,...,r T-1 } represents the user's service experience by providing feedback for each historical search request; it also retrieves the individual user interest and preference features p = Transformer(H + H). p+ ), H p+ This represents the positional embedding of each record tuple consisting of a search request and its corresponding evaluation feedback in the historical search records; Global entity representation is achieved using the Transformer model, replacing the input data x with the rating feedback R = {r1, r2, ..., r} of all users for a given entity. u ,...,r U }, obtain the global entity representation e U e U =Transformer(R+R) p+ ), R p+ This represents the location embedding corresponding to all user reviews and feedback for an entity; S32: Utilize the three hidden features q obtained in step S31 p p, e U Enhance the current query q T The semantically enhanced current query representation obtained is MLP(·) represents a multilayer perceptron model, which is responsible for fusing these feature vectors together; S33: Calculate user-enhanced current query using cosine similarity. Personalized matching score with entity e This completes the construction of personalized search to achieve personalized result ranking, represented as:

4. The personalized data service method according to claim 3, characterized in that, Step S4 specifically includes the following steps: S41: Joint inference of the edge-cloud collaborative model requires multiple rounds of iterative updates to the personalized search model. During each round of training, each edge server completes the training of each model locally. Local training on edge servers: The specific process is represented as follows M t Let t represent the personalized search model parameters of the edge server during the training round at time t, x represent the input data required for the extraction of the three latent features in step S3, and SGD(·) represent the stochastic gradient descent method used for model iterative updates. This represents the parameters of the personalized search model trained on the k-th edge server at time t+1; S42: Edge Server Parameter Upload: The model parameters will be updated for each edge server. Uploaded to the cloud server; S43: Cloud Server Parameter Aggregation: In each round of joint inference, the cloud server aggregates the model parameters uploaded by each edge server. Compared to the model parameters M in the previous round t The aggregation model parameters at time t+1 are expressed as follows: K represents the total number of edge servers; S44: Cloud server parameter distribution: The cloud server will distribute the new round of aggregated model parameters M t+1 Send to each edge server; S45: Edge-Cloud Collaborative Joint Inference: In the joint inference and training model, the loss function Loss is defined as: in, This represents the relevance score between the search request and the target entity. This represents the relevance score of the search request to other candidate entities, and the optimization objective of joint inference is to minimize the objective function: Where Θ k This represents all parameters of the personalized model in the k-th edge server; S46: Model Acquisition: Repeat steps S41-S44 until the objective function in step S45 is minimized and stabilized. At this point, each edge server can obtain the optimal personalized search model.

5. The personalized data service method according to claim 4, characterized in that, Step S6 specifically includes the following steps: S61: The edge server determines the cache location of the entity state data required by the user. If the required entity state data is cached on the edge server, the edge server performs a search based on the traditional retrieval matching method, filters out relevant entity results, and then runs the trained personalized search model to reorder the entity results in a personalized way so that the entity results most satisfactory to the user appear at the top of the sort list. Finally, the personalized entity result list is fed back to the user. S62: If the required entity state data is not cached on the edge server, the edge server will encrypt the search request and upload it to the cloud server. The cloud server will perform a general search request and entity relevance matching in the encrypted environment based on the symmetric searchable encryption mechanism, and then feed back the encrypted entity result list to the corresponding edge server. The edge server will decrypt the entity result list and run the trained personalized search model to reorder the entity results in a personalized way, and can also return personalized sorting results to the user.

Images