A longitudinal federated EBM commodity recommendation method with differential privacy protection

Abstract

The application discloses a longitudinal federated EBM commodity recommendation method with differential privacy protection, and comprises the following steps: 1, a master server uses a consumption feature data set to train an EBM recommendation model; 2, a slave server number is initialized; 3, the slave server uses the consumption feature data set to train the EBM recommendation model; and 4, whether the EBM recommendation model training is ended is determined according to the number of iterations. The application can effectively protect the user consumption feature information, thereby preventing the leakage of the user consumption feature information and providing effective commodity recommendation.

Description

Technical Field

[0001] This invention belongs to the field of vertical federated learning recommendation, specifically a differential privacy-preserving vertical federated EBM product recommendation method. Background Technology

[0002] With the development of the internet, people face massive amounts of diverse data every day. This data might be text, audio, or files from consumer apps, or a dazzling array of products from online shopping malls. However, this data may be stored across multiple companies, and the data from a single company is insufficient to support a highly efficient and accurate recommendation system. This creates the problem of data silos, necessitating an effective architecture to integrate data from multiple companies. Vertical federated learning provides such a solution. It can integrate data from multiple companies to build a product recommendation system without leaking local company data. Compared to product recommendation systems built from a single company, vertical federated learning offers faster and more accurate recommendations, providing users with a better service experience.

[0003] However, in federated learning recommendation algorithms, companies upload their user data to third parties without reservation. This data may include sensitive information such as consumption records, medical records, and educational records. If an untrustworthy third party leaks this data to other institutions or individuals for commercial gain, it will inevitably lead to the leakage of user privacy, reduce user interest in online services, lower user satisfaction, and hinder the development of the internet economy. Furthermore, the establishment and use of existing vertical federated learning recommendation systems require significant communication time. Therefore, achieving efficient and rapid vertical federated learning recommendations without disclosing sensitive personal information becomes particularly important. Summary of the Invention

[0004] This invention aims to address the shortcomings of existing technologies by proposing a differential privacy-preserving vertical federated EBM product recommendation method. This method effectively solves the problem of poor user data privacy in existing federated learning recommendation schemes, thereby better protecting user data privacy and maintaining a balance between the accuracy, diversity, and efficiency of recommendation results.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0006] The present invention discloses a differential privacy-preserving vertical federated EBM product recommendation method, characterized by its application in a network environment consisting of one master server, L-1 slave servers, and an EBM recommendation model. The user consumption feature dataset held by the master server is denoted as... Among them, X A,iLet i be the consumption feature of the i-th user on the main server, Y be the tag value of the recommended product, and n be the value of the user's consumption feature on the main server. A The number of user consumption characteristics on the main server; and: in, For X A,i The t-th consumption feature value, X A,i The dimension is d; Y = {y1, y2, ..., y t ,…,y d}, where y t Let be the t-th value of the recommended product tag Y; the dimension of Y is d; let the l-th user consumption feature dataset from the server be denoted as . X p,l,j For the consumption characteristics of the l-th slave user j-th user, n P Let l be the number of user consumption features of the l-th slave server. in, For X P,l,i The t-th consumption feature value, X P,l,i The dimension is d; 1≤l≤L-1, 1≤i≤n A , 1≤t≤d, 1≤j≤n P The vertical federal EBM product recommendation method is performed according to the following steps:

[0007] Step 1: The main server uses U A Training the EBM recommendation model:

[0008] Step 1.1: Define and initialize the current iteration number inter = 1; define the maximum number of iterations Inter; define the minimum recommendation gain value as Gain. min The maximum depth of a weak recommendation decision tree is depth. max ;

[0009] Define the residual of the product recommendation as C = {c1, c2, ..., c t ,…,c d}; where c t Let C represent the t-th residual value and initialize C = Y;

[0010] Step 1.2: Initialize the number of servers st that complete training in the inter-th iteration. inter =0;

[0011] Step 1.3: Obtain the privacy budget of the master server in the inter-th iteration according to equation (1).

[0012]

[0013] In equation (1), ε represents the privacy budget, ε max For maximum privacy budget, ε inc Increased privacy budget for servers;

[0014] Step 1.4: Initialize i = 1;

[0015] Step 1.5: Define the residual of the i-th consumer feature trained in the inter-th iteration of the main server. in, express The k-th residual value, d A Residual The number of, and initialize and d A =d, 1≤k≤d A ;

[0016] Step 1.6, from X A,i Randomly select n s Each consumption characteristic value constitutes consumption characteristic X. A,i The set of segmentation values in, yes The w-th segment value, 1≤w≤n s <d;

[0017] Step 1.7: Define and initialize the depth of the weak recommendation decision tree to depth = 0;

[0018] Step 1.8: Initialize w = 1;

[0019] Step 1.9, regarding the segmentation value master node Storage Assign depth+1 to depth; use the main node right Perform the partitioning to obtain the residual after the partitioning by the main server in the inter-th iteration. in, Residual The left residual after partitioning, and The dimension is express The right residual after partitioning, and The dimension is in, and

[0020] Step 1.10: Let variable b equal left or right, and calculate the left or right residual using equation (2). Recommended gain

[0021]

[0022] In equation (2),

[0023] Step 1.11: Determine if depth = depth max Check if the condition is met. If it is met, proceed to step 1.13; otherwise, proceed to step 1.12.

[0024] Step 1.12, Judgment Check if the condition is met. If it is met, proceed to step 1.13; otherwise, proceed to step 1.14.

[0025] Step 1.13: Use left or right residuals Generate left or right leaf nodes The leaf nodes of the main server are obtained using equation (3). Product recommendation forecast

[0026]

[0027] In equation (3), Lap() is a random function that follows a Laplace distribution, Range is the range of values ​​for the product recommendation tag Y, and η is the learning rate;

[0028] Step 1.14: Assign w+1 to w, and check if w > n. s Check if the condition is true; if true, then a set of leaf nodes has been generated. and with the master node Together they constitute the i-th user consumption feature X of the main server. A,i Weak recommendation decision tree Proceed to step 1.15; otherwise, Assign to Will Assigned to d A Return to step 1.9 and execute sequentially;

[0029] Step 1.15: Update the product recommendation residual value Then, assign i+1 to i, and check if i>n. A If the condition is met, it means the main server has completed the `inter`th round of training on the EBM recommendation model, and the trained EBM recommendation model is denoted as [model name missing]. and st inter +1 is assigned to stinter If yes, proceed to step 2; otherwise, return to step 1.5.

[0030] Step 2: Initialize l=1; the main server will update the product recommendation residual values ​​C and st. inter Transmit to the l-th slave server; As the (l-1)th EBM recommendation model trained from the server,

[0031] Step 3: The l-th slave server uses the product feature dataset U P,l And the received recommended residual values ​​C and st inter right After performing the inter-th round of training, the trained EBM recommendation model is denoted as . and st inter +1 is assigned to st inter ;

[0032] Step 4: Determine if l = L - 1 is true. If true, proceed to step 5; otherwise, assign l + 1 to l, and then transmit the updated recommended residual value C and st. inter Move to the l-th slave server and return to step 3 for sequential execution;

[0033] Step 5: Determine if inter > Inter is true. If true, the final recommendation model is generated. And utilize recommendation models Recommend products based on the input consumption features and generate a product recommendation set; otherwise, assign inter+1 to inter, send the updated recommendation residual value C to the main server, and return to step 1.2 for sequential execution.

[0034] The differential privacy-preserving vertical federated EBM product recommendation method described in this invention is also characterized in that step 3 includes:

[0035] Step 3.1: Obtain the privacy budget of the l-th slave server in the inter-th iteration according to equation (4).

[0036]

[0037] Step 3.2: Initialize j = 1;

[0038] Step 3.3: Define the residual of the l-th iteration of training the j-th consumption feature from the server. in, express Find the z-th residual value and initialize it. and d P,l=d, 1≤z≤d P,l ;

[0039] Step 3.4, from X P,l,j Randomly select n s Each consumption characteristic value constitutes consumption characteristic X. P,l,j Set of split values in, yes The w-th segment value, 1≤w≤n s <d;

[0040] Step 3.5: Define and initialize the depth of the weak recommendation decision tree to depth = 0;

[0041] Step 3.6: Initialize w = 1;

[0042] Step 3.7, regarding the segmentation value master node Storage Assign depth+1 to depth; use the main node right Perform the segmentation to obtain the l-th residual after the inter-th iteration from the server. in, express The left residual, and The dimension is express The right residual, and The dimension is in, and

[0043] Step 3.8: Let variable b equal left or right, and calculate the time passed by the l-th slave server in the inter-th iteration using equation (5). Left or right residual after segmentation Recommended gain

[0044]

[0045] In equation (5),

[0046] Step 3.9: Determine if depth = depth max Check if the condition is met. If it is met, proceed to step 3.11; otherwise, proceed to step 3.10.

[0047] Step 3.10, Judgment Check if the condition is met. If it is met, proceed to step 3.11; otherwise, proceed to step 3.12.

[0048] Step 3.11: Utilize the left or right residual Generate leaf nodes The leaf node of the l-th slave server is obtained through equation (6). Product recommendation forecast

[0049]

[0050] In equation (3),

[0051] Step 3.12: Assign w+1 to w, and check if w > n. s Check if the condition is true; if true, generate a set of leaf nodes. and with the master node Together they constitute the consumption feature X of the j-th user on the l-th slave server. P,l,j Weak recommendation decision tree Proceed to step 3.13; otherwise, Assign to Will Assigning d P,l Return to step 3.7 and execute sequentially;

[0052] Step 3.13: Update the product recommendation residual value Then, assign j+1 to j, and check if j>n. P If the condition is met, it means that the l-th slave has completed the inter-th round of training on the EBM recommendation model from the server; otherwise, return to step 3.3 and execute sequentially.

[0053] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the vertical federated EBM product recommendation method, and the processor is configured to execute the program stored in the memory.

[0054] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of the vertical federated EBM product recommendation method.

[0055] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0056] 1. Compared with the traditional method of combining vertical federated learning with decision tree integrated product recommendation model, this invention uses EBM recommendation algorithm combined with vertical federated learning scenario to realize product recommendation function. The biggest difference between it and the decision tree integrated product recommendation model is that its weak recommendation decision tree is determined by a single consumption feature and the split point set is randomly generated, thereby improving the efficiency of product recommendation and providing better product recommendation solutions based on user needs.

[0057] 2. This invention uses differential privacy to protect consumption features. Differential privacy protects privacy by adding noise to the predicted values ​​of the leaf nodes of the weak recommendation decision tree. Through this noise-adding mechanism, the level of privacy protection for product recommendations can be controlled, solving the problem of user privacy leakage. The resulting product recommendation scheme not only maintains the diversity of product recommendations but also improves the security of product recommendations.

[0058] 3. Unlike the vertical federated learning combined with decision tree ensemble recommendation model, the EBM recommendation model only needs to transmit residual values ​​and the number of trained servers to other servers during training. In the case of multi-round recommendation model training, it will save a lot of communication time, thereby reducing the latency waiting time for users to use product recommendation services and reducing the communication consumption generated by product recommendation. Attached Figure Description

[0059] Figure 1 This is a flowchart of a vertically federated EBM product recommendation method with differential privacy protection according to the present invention. Detailed Implementation

[0060] In this embodiment, a differential privacy-preserving vertical federated EBM product recommendation method is applicable to protecting user privacy during the vertical federated recommendation process. First, the master server uses its local dataset to train the EBM recommendation model. Then, it transmits residual data to the next slave server. After the last slave server completes its EBM recommendation model training, it begins the next round of EBM recommendation model building until all rounds are completed. Specifically, it is applied to a network environment consisting of one master server, L-1 slave servers, and one EBM recommendation model. The user consumption feature dataset held by the master server is denoted as... Among them, X A,i Let i be the consumption feature of the i-th user on the main server, Y be the tag value of the recommended product, and n be the value of the user's consumption feature on the main server. A The number of user consumption characteristics on the main server; and: in, For X A,i The t-th consumption feature value, X A,i The dimension is d; Y = {y1, y2, ..., yt ,…,y d}, where y t Let be the t-th value of the recommended product tag Y; the dimension of Y is d; let the l-th user consumption feature dataset from the server be denoted as . X p,l,j For the consumption characteristics of the l-th slave user j-th user, n P Let l be the number of user consumption features of the l-th slave server. in, For X P,l,i The t-th consumption feature value, X P,l,i The dimension is d; 1≤l≤L-1, 1≤i≤n A , 1≤t≤d, 1≤j≤n P ;like Figure 1 As shown, the method is performed according to the following steps:

[0061] Step 1: The main server uses U A Training the EBM recommendation model:

[0062] Step 1.1: Define and initialize the current iteration number inter = 1; define the maximum number of iterations Inter; define the minimum recommendation gain value as Gain. min The maximum depth of a weak recommendation decision tree is depth. max ;

[0063] Define the residual of the product recommendation as C = {c1, c2, ..., c t ,…,c d}; where c t Let C represent the t-th residual value and initialize C = Y;

[0064] Step 1.2: Initialize the number of servers st that complete training in the inter-th iteration. inter =0;

[0065] Step 1.3: Obtain the privacy budget of the master server in the inter-th iteration according to equation (1). The smaller the privacy budget, the higher the level of privacy protection. By controlling the privacy budget, the accuracy of product recommendations and the degree of privacy protection can be achieved.

[0066]

[0067] In equation (1), ε represents the privacy budget, ε max For maximum privacy budget, ε inc Increased privacy budget for servers;

[0068] Step 1.4: Initialize i = 1;

[0069] Step 1.5: Define the residual of the i-th consumption feature in the inter-th iteration training of the master server Wherein, denotes the k-th residual value of, d A denotes the residual and initialize and d A = d, 1 ≤ k ≤ d A ;

[0070] Step 1.6: Randomly select n A,i consumption feature values from X s to form the set of splitting values of the consumption feature X A,i Wherein, is the w-th splitting value in, 1 ≤ w ≤ n < d, randomly selecting the set of splitting points can speed up the construction speed of the commodity recommendation model without affecting the recommendation accuracy of the model; s

[0071] Step 1.7: Define and initialize the depth of the weak recommendation decision tree as depth = 0;

[0072] Step 1.8: Initialize w = 1;

[0073] Step 1.9: Store in the main node about the splitting value and assign depth + 1 to depth; use the main node to perform splitting to obtain the residual after splitting by the master server in the inter-th iteration Wherein, denotes the left residual after splitting of the residual and the dimension of is denotes the right residual after splitting of, and the dimension of is Wherein, and

[0074] Step 1.10: Let the variable b be equal to left or right, and calculate the recommendation gain of the left or right residual through formula (2)

[0075]

[0076] In equation (2),

[0077] Step 1.11: Determine if depth = depth max Check if the condition is met. If it is, proceed to step 1.13; otherwise, execute step 1.12 to set a depth threshold for the weak recommendation decision tree. max It can limit the height of weak recommendation decision trees, thereby improving the accuracy of recommendations;

[0078] Step 1.12, Judgment If the condition is met, proceed to step 1.13; otherwise, proceed to step 1.14. A smaller gain indicates a better segmentation effect. The gain threshold (Gain) is set accordingly. min This can prevent weak recommendation decision trees from overfitting;

[0079] Step 1.13: Use left or right residuals Generate left or right leaf nodes The leaf nodes of the main server are obtained using equation (3). Product recommendation forecast

[0080]

[0081] In equation (3), Lap() is a random function that follows a Laplace distribution, Range is the range of values ​​for the product recommendation tag Y, and η is the learning rate;

[0082] Step 1.14: Assign w+1 to w, and check if w > n. s Check if the condition is true; if true, then a set of leaf nodes has been generated. and with the master node Together they constitute the i-th user consumption feature X of the main server. A,i Weak recommendation decision tree Proceed to step 1.15; otherwise, Assign to Will Assigned to d A Return to step 1.9 and execute sequentially;

[0083] Step 1.15: Update the product recommendation residual value Then, assign i+1 to i, and check if i>n. AIf the condition is met, it means the main server has completed the `inter`th round of training on the EBM recommendation model, and the trained EBM recommendation model is denoted as [model name missing]. and st inter +1 is assigned to st inter If yes, proceed to step 2; otherwise, return to step 1.5.

[0084] Step 2: Initialize l = 1;

[0085] Step 3: The main server updates the product recommendation residual values ​​C and st. inter The data is transmitted to the l-th slave server. Between servers, only the product recommendation residual value C and st need to be exchanged. inter This can greatly reduce communication losses and improve product recommendation efficiency; As the (l-1)th EBM recommendation model trained from the server,

[0086] Step 3.0: The l-th slave server uses the product feature dataset U P,l And the received recommended residual values ​​C and st inter right Perform the inter round of training:

[0087] Step 3.1: Obtain the privacy budget of the l-th slave server in the inter-th iteration according to equation (4).

[0088]

[0089] Step 3.2: Initialize j = 1;

[0090] Step 3.3: Define the residual of the l-th iteration of training the j-th consumption feature from the server. in, express Find the z-th residual value and initialize it. and d P,l =d, 1≤z≤d P,l ;

[0091] Step 3.4, from X P,l,j Randomly select n s Each consumption characteristic value constitutes consumption characteristic X. P,l,j Set of split values in, yes The w-th segment value, 1≤w≤n s <d;

[0092] Step 3.5: Define and initialize the depth of the weak recommendation decision tree to depth = 0;

[0093] Step 3.6: Initialize w = 1;

[0094] Step 3.7, regarding the segmentation value master node Storage Assign depth+1 to depth; use the main node right Perform the segmentation to obtain the l-th residual after the inter-th iteration from the server. in, express The left residual, and The dimension is express The right residual, and The dimension is in, and

[0095] Step 3.8: Let variable b equal left or right, and calculate the time passed by the l-th slave server in the inter-th iteration using equation (5). Left or right residual after segmentation Recommended gain

[0096]

[0097] In equation (5),

[0098] Step 3.9: Determine if depth = depth max Check if the condition is met. If it is met, proceed to step 3.11; otherwise, proceed to step 3.10.

[0099] Step 3.10, Judgment Check if the condition is met. If it is met, proceed to step 3.11; otherwise, proceed to step 3.12.

[0100] Step 3.11: Utilize the left or right residual Generate leaf nodes The leaf node of the l-th slave server is obtained through equation (6). Product recommendation forecast

[0101]

[0102] In equation (3),

[0103] Step 3.12: Assign w+1 to w, and check if w > n. s Check if the condition is true; if true, generate a set of leaf nodes. and with the master node Together they constitute the consumption feature X of the j-th user on the l-th slave server. P,l,j Weak recommendation decision tree Proceed to step 3.13; otherwise, Assign to Will Assigning d P,l Return to step 3.7 and execute sequentially;

[0104] Step 3.13: Update the product recommendation residual value Then, assign j+1 to j, and check if j>n. P If the condition is met, it means that the l-th slave server has completed the inter-th round of training on the EBM recommendation model, and the trained EBM recommendation model is denoted as . and st inter +1 is assigned to st inter Otherwise, return to step 3.3 and execute sequentially.

[0105] Step 3.14: Determine if l = L - 1 is true. If true, proceed to step 4; otherwise, assign l + 1 to l, and then transmit the updated recommended residual value C and st. inter Move to the l-th slave server and return to step 3.0 for sequential execution;

[0106] Step 4: Determine if inter > Inter is true. If true, the final recommendation model is generated. And utilize recommendation models Recommendations are made based on the input consumption features to generate a product recommendation set. The accuracy of the recommendation scheme provided by the EBM recommendation model is improved through Inter rounds of training; otherwise, after assigning inter+1 to inter, the updated recommendation residual value C is given to the main server, and the process returns to step 1.2 for sequential execution.

[0107] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0108] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A differential privacy-preserving vertical federated EBM product recommendation method, characterized by: In a network environment consisting of one master server, L-1 slave servers, and an EBM recommendation model, the user consumption feature dataset held by the master server is denoted as... Among them, X A,i Let i be the consumption feature of the i-th user on the main server, Y be the tag value of the recommended product, and n be the value of the user's consumption feature on the main server. A The number of user consumption characteristics on the main server; and: in, For X A,i The t-th consumption feature value, X A,i The dimension is d; Y = {y1, y2, ..., y t ,…,y d }, where y t Let be the t-th value of the recommended product tag Y; the dimension of Y is d; let the l-th user consumption feature dataset from the server be denoted as . X p,l,j For the consumption characteristics of the l-th slave user j-th user, n P Let l be the number of user consumption features of the l-th slave server. in, For X P,l,i The t-th consumption feature value, X P,l,i The dimension is d; 1≤l≤L-1, 1≤i≤n A , 1≤t≤d, 1≤j≤n P The vertical federal EBM product recommendation method is performed according to the following steps: Step 1: The main server uses U A Training the EBM recommendation model: Step 1.1: Define and initialize the current iteration number inter = 1; define the maximum number of iterations Inter; define the minimum recommendation gain value as Gain. min The maximum depth of a weak recommendation decision tree is depth. max ; Define the residual of the product recommendation as C = {c1, c2, ..., c t ,…,c d }; where c t Let C represent the t-th residual value and initialize C = Y; Step 1.2: Initialize the number of servers st that complete training in the inter-th iteration. inter =0; Step 1.3: Obtain the privacy budget of the master server in the inter-th iteration according to equation (1). In equation (1), ε represents the privacy budget, ε max For maximum privacy budget, ε inc Increased privacy budget for servers; Step 1.4: Initialize i = 1; Step 1.5: Define the residual of the i-th consumer feature trained in the inter-th iteration of the main server. in, express The k-th residual value, d A Residual The number of, and initialize and d A =d, 1≤k≤d A ; Step 1.6, from X A,i Randomly select n s Each consumption characteristic value constitutes consumption characteristic X. A,i The set of segmentation values in, yes The w-th segment value, 1≤w≤n s <d; Step 1.7: Define and initialize the depth of the weak recommendation decision tree to depth = 0; Step 1.8: Initialize w = 1; Step 1.9, regarding the segmentation value master node Storage Assign depth+1 to depth; use the main node right Perform the partitioning to obtain the residual after the partitioning by the main server in the inter-th iteration. in, Residual The left residual after partitioning, and The dimension is express The right residual after partitioning, and The dimension is in, and Step 1.10: Let variable b equal left or right, and calculate the left or right residual using equation (2). Recommended gain In equation (2), Step 1.11: Determine if depth = depth max Check if the condition is met. If it is met, proceed to step 1.13; otherwise, proceed to step 1.

12. Step 1.12, Judgment Check if the condition is met. If it is met, proceed to step 1.13; otherwise, proceed to step 1.

14. Step 1.13: Use left or right residuals Generate left or right leaf nodes The leaf nodes of the main server are obtained using equation (3). Product recommendation forecast In equation (3), Lap() is a random function that follows a Laplace distribution, Range is the range of values ​​for the product recommendation tag Y, and η is the learning rate; Step 1.14: Assign w+1 to w, and check if w > n. s Check if the condition is true; if true, then a set of leaf nodes has been generated. and with the master node Together they constitute the i-th user consumption feature X of the main server. A,i Weak recommendation decision tree Proceed to step 1.15; otherwise, Assign to Will Assigned to d A Return to step 1.9 and execute sequentially; Step 1.15: Update the product recommendation residual value Then, assign i+1 to i, and check if i>n. A If the condition is met, it means the main server has completed the `inter`th round of training on the EBM recommendation model, and the trained EBM recommendation model is denoted as [model name missing]. and st inter +1 is assigned to st inter If yes, proceed to step 2; otherwise, return to step 1.

5. Step 2: Initialize l=1; the main server will update the product recommendation residual values ​​C and st. inter Transmit to the l-th slave server; As the (l-1)th EBM recommendation model trained from the server, Step 3: The l-th slave server uses the product feature dataset U P,l And the received recommended residual values ​​C and st inter right After performing the inter-th round of training, the trained EBM recommendation model is denoted as . and st inter +1 is assigned to st inter ; Step 4: Determine if l = L - 1 is true. If true, proceed to step 5; otherwise, assign l + 1 to l, and then transmit the updated recommended residual value C and st. inter Move to the l-th slave server and return to step 3 for sequential execution; Step 5: Determine if inter > Inter is true. If true, the final recommendation model is generated. And utilize recommendation models Recommend products based on the input consumption features and generate a product recommendation set; otherwise, assign inter+1 to inter, send the updated recommendation residual value C to the main server, and return to step 1.2 for sequential execution.

2. The differential privacy-preserving vertical federated EBM product recommendation method according to claim 1, characterized in that, Step 3 includes: Step 3.1: Obtain the privacy budget of the l-th slave server in the inter-th iteration according to equation (4). Step 3.2: Initialize j = 1; Step 3.3: Define the residual of the l-th iteration of training the j-th consumption feature from the server. in, express Find the z-th residual value and initialize it. and d P,l =d, 1≤z≤d P,l ; Step 3.4, from X P,l,j Randomly select n s Each consumption characteristic value constitutes consumption characteristic X. P,l,j Set of split values in, yes The w-th segment value, 1≤w≤n s <d; Step 3.5: Define and initialize the depth of the weak recommendation decision tree to depth = 0; Step 3.6: Initialize w = 1; Step 3.7, regarding the segmentation value master node Storage Assign depth+1 to depth; use the main node right Perform the segmentation to obtain the l-th residual after the inter-th iteration from the server. in, express The left residual, and The dimension is express The right residual, and The dimension is in, and Step 3.8: Let variable b equal left or right, and calculate the time passed by the l-th slave server in the inter-th iteration using equation (5). Left or right residual after segmentation Recommended gain In formula (5) Step 3.9: Determine if depth = depth max Check if the condition is met. If it is met, proceed to step 3.11; otherwise, proceed to step 3.

10. Step 3.10, Judgment Check if the condition is met. If it is met, proceed to step 3.11; otherwise, proceed to step 3.

12. Step 3.11: Utilize the left or right residual Generate leaf nodes The leaf node of the l-th slave server is obtained through equation (6). Product recommendation forecast In equation (3), Step 3.12: Assign w+1 to w, and check if w > n. s Check if the condition is true; if true, generate a set of leaf nodes. and with the master node Together they constitute the consumption feature X of the j-th user on the l-th slave server. P,l,j Weak recommendation decision tree Proceed to step 3.13; otherwise, Assign to Will Assigning d P,l Return to step 3.7 and execute sequentially; Step 3.13: Update the product recommendation residual value Then, assign j+1 to j, and check if j>n. P If the condition is met, it means that the l-th slave has completed the inter-th round of training on the EBM recommendation model from the server; otherwise, return to step 3.3 and execute sequentially.

3. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing the vertical federated EBM product recommendation method of claim 1 or 2, the processor being configured to execute the programs stored in the memory.

4. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when run by a processor, performs the steps of the vertical federated EBM product recommendation method as described in claim 1 or 2.

Images