Joint modeling method, device, electronic device and storage medium
By performing sample alignment and binning processing in federated learning, combined with a secure aggregation algorithm, the decision tree is built layer by layer, which solves the problem of low accuracy of the joint model when sample features and quantity is missing, and improves the accuracy and robustness of the model .
Patent Information
- Application Number
- CN202210046307.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-14
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2042-01-14
AI Technical Summary
In federated learning, if the participants lack both sample characteristics and sample size, the joint model constructed using existing horizontal or vertical federated learning schemes is not very accurate.
By receiving the homomorphic encryption key and sample gradient ciphertext sent by each second participant, sample alignment processing and binning are performed, target gradient histogram ciphertext is constructed, gradient information is merged using a secure aggregation algorithm to determine the characteristics to be segmented and Split points and build the decision tree layer by layer until the preset depth threshold is reached and the model convergence is judged.
The joint model accuracy in the absence of sample features and sample size is improved, combining the advantages of horizontal and vertical participants, and the robustness and accuracy of the model are enhanced.
Smart Images

Figure CN114422105B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data processing, and in particular to a joint modeling method, device, electronic device and storage medium. Background Art
[0002] In order to eliminate data silos and ensure data security, federated learning has been widely used in joint modeling. In the process of federated learning, each participant does not share data, but uses local data to train the model, and updates the model parameters by exchanging encrypted parameters to complete the modeling.
[0003] Federated learning includes horizontal federated learning and vertical federated learning. Usually, if the participants have the same sample characteristics but the number of samples is insufficient, the horizontal federated learning solution is adopted; if the number of samples of the participants is sufficient but the sample characteristics are missing, the vertical federated learning solution is adopted. However, in the case of lack of both sample characteristics and sample quantity, the accuracy of the constructed joint model is not high regardless of whether the horizontal federated learning solution or the vertical federated learning solution is used. Therefore, there is an urgent need for a joint modeling method to improve the accuracy of the joint model in the absence of sample characteristics and sample quantity. Summary of the invention
[0004] In view of the above, it is necessary to provide a joint modeling method, aiming to improve the accuracy of the joint model.
[0005] The joint modeling method provided by the present invention is applied to any first participant in a joint modeling system, wherein the joint modeling system includes a plurality of first participants and a plurality of second participants in communication connection, each first participant and each second participant contain the same sample objects and different sample features, each second participant contains the same sample features and different sample objects, the samples of the first participants do not contain labels, and the samples of the second participants carry labels, and the method includes:
[0006] Receive a public key in a homomorphic encryption key pair sent by each second participant in the joint modeling system, perform sample alignment processing on the locally stored first sample set and the second sample set of each second participant based on the public key, and split the first sample set into sub-sample sets corresponding to each second participant based on the sample alignment processing result;
[0007] Receiving a first-order gradient ciphertext and a second-order gradient ciphertext of each sample in the second sample set sent by each second participant;
[0008] Assign each sub-sample set to the root node of the corresponding tree to be constructed;
[0009] Bin each feature in the sample on each root node, and construct the target gradient histogram ciphertext of each root node based on the binning results, the first-order gradient ciphertext, and the second-order gradient ciphertext;
[0010] Sending the target gradient histogram ciphertext to the corresponding second participants, and receiving the gradient histogram merged graph processed by the secure aggregation algorithm returned by each second participant;
[0011] Determine the features to be segmented and the points to be segmented according to the gradient histogram merged graph, and construct the next layer of tree nodes according to the features to be segmented and the points to be segmented;
[0012] When the depth of the tree node reaches the preset depth threshold, it is determined whether the constructed tree meets the convergence condition. If so, the joint modeling is completed.
[0013] Optionally, constructing a target gradient histogram ciphertext of each root node based on the binning result, the first-order gradient ciphertext, and the second-order gradient ciphertext includes:
[0014] Select a root node of the tree corresponding to a subsample set, sum the first-order gradient ciphertexts of the samples in each box of each feature on the root node of the selected tree, and obtain the sum of the first-order gradient ciphertexts corresponding to each box of each feature;
[0015] Sum up the second-order gradient ciphertexts of the samples in each box of each feature at the root node of the selected tree to obtain the sum of the second-order gradient ciphertexts corresponding to each box of each feature;
[0016] Draw an initial gradient histogram ciphertext corresponding to each feature on the root node of the selected tree based on the first-order gradient ciphertext and the second-order gradient ciphertext;
[0017] The initial gradient histogram ciphertext corresponding to each feature is summarized to obtain the target gradient histogram ciphertext corresponding to the root node of the selected tree.
[0018] Optionally, determining the features to be segmented and the points to be segmented according to the gradient histogram merged graph includes:
[0019] For each feature in each gradient histogram, the starting point of each box is used as the segmentation point, and the information gain value corresponding to each segmentation point is calculated;
[0020] The segmentation point with the largest information gain value is taken as the point to be segmented, and the feature corresponding to the point to be segmented is taken as the feature to be segmented.
[0021] Optionally, constructing the next layer of tree nodes according to the features to be segmented and the points to be segmented includes:
[0022] If the feature to be segmented belongs to a first participant, the feature to be segmented and its segmentation point are sent to the first participant, so that the first participant can separate the samples on the corresponding root node, and send the separation result to the corresponding second participant, so that the corresponding second participant can complete the construction of the next layer of tree nodes;
[0023] If the feature to be segmented belongs to a second participant, the feature to be segmented and its point to be segmented are sent to each second participant, so that each second participant separates the samples of its root node and completes the construction of the next layer of tree nodes.
[0024] Optionally, determining whether the constructed tree satisfies a convergence condition includes:
[0025] Receive the loss value processed by the security aggregation algorithm and corresponding to the tree constructed by each second participant;
[0026] Aggregate the loss values to obtain a total loss value. If the difference between the obtained total loss value and the total loss value of the last constructed tree is less than the loss threshold, the constructed tree meets the convergence condition.
[0027] Optionally, the loss value calculation process includes:
[0028] Each second participant calculates a weight value for each leaf node of the tree constructed by it, and calculates a new prediction value for each sample in its second sample set based on the weight value;
[0029] Each second participant calculates the loss value corresponding to the tree it constructs based on the label of each sample in its second sample set and the new prediction value.
[0030] Optionally, after determining whether the constructed tree meets the convergence condition, the method further includes:
[0031] If the judgment is no, receive the new first-order gradient ciphertext and the new second-order gradient ciphertext of each sample in the second sample set sent by each second participant, and construct the next tree based on the new first-order gradient ciphertext and the new second-order gradient ciphertext.
[0032] In order to solve the above problems, the present invention further provides a joint modeling device, the device comprising:
[0033] An alignment module is used to receive a public key in a homomorphic encryption key pair sent by each second participant in the joint modeling system, perform sample alignment processing on a locally stored first sample set and a second sample set of each second participant based on the public key, and split the first sample set into sub-sample sets corresponding to each second participant based on the sample alignment processing result;
[0034] A receiving module, used for receiving a first-order gradient ciphertext and a second-order gradient ciphertext of each sample in the second sample set sent by each second participant;
[0035] An allocation module, used for respectively allocating each sub-sample set to the root node of the corresponding tree to be constructed;
[0036] The binning module is used to bin each feature in the sample on each root node, and construct the target gradient histogram ciphertext of each root node based on the binning result, the first-order gradient ciphertext and the second-order gradient ciphertext;
[0037] A sending module, used to send the target gradient histogram ciphertext to the corresponding second participant, and receive the gradient histogram merged graph processed by the security aggregation algorithm returned by each second participant;
[0038] A determination module, used for determining the features to be segmented and the points to be segmented according to the gradient histogram, and constructing the next layer of tree nodes according to the features to be segmented and the points to be segmented;
[0039] The judgment module is used to judge whether the constructed tree meets the convergence condition when the depth of the tree node reaches a preset depth threshold. If the judgment is yes, the joint modeling is completed.
[0040] In order to solve the above problem, the present invention further provides an electronic device, the electronic device comprising:
[0041] at least one processor; and,
[0042] a memory communicatively connected to the at least one processor; wherein,
[0043] The memory stores a joint modeling program that can be executed by the at least one processor, and the joint modeling program is executed by the at least one processor so that the at least one processor can perform the above-mentioned joint modeling method.
[0044] In order to solve the above problems, the present invention also provides a computer-readable storage medium, on which a joint modeling program is stored. The joint modeling program can be executed by one or more processors to implement the above joint modeling method.
[0045] Compared with the prior art, the present invention first assigns each sub-sample set to the root node of the corresponding tree to be constructed, bins each feature in the sample on each root node, and constructs the target gradient histogram ciphertext of each root node based on the binning results, the first-order gradient ciphertext and the second-order gradient ciphertext; then, the target gradient histogram ciphertext is sent to the corresponding second participant, and the gradient histogram merged graph processed by the secure aggregation algorithm returned by each second participant is received; then, the features to be segmented and their points to be segmented are determined according to the gradient histogram merged graph, so as to construct the next layer of tree nodes; finally, when the depth of the tree node reaches the preset depth threshold, it is determined whether the constructed tree meets the convergence condition, and if it is determined to be yes, the joint modeling is completed. The present invention unites horizontal participants with the same sample features and different sample objects and vertical participants with the same sample objects and different sample features to jointly construct one or more decision trees, thereby improving the accuracy of the joint model. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 A schematic diagram of a process flow of a joint modeling method provided by an embodiment of the present invention;
[0047] Figure 2 A schematic diagram of modules of a joint modeling device provided by an embodiment of the present invention;
[0048] Figure 3 A schematic diagram of the structure of an electronic device for implementing a joint modeling method provided by an embodiment of the present invention;
[0049] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION
[0050] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0051] It should be noted that the descriptions of "first", "second", etc. in the present invention are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in the field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.
[0052] The present invention provides a joint modeling method, which is applied to any first participant in a joint modeling system, wherein the joint modeling system includes a plurality of first participants and a plurality of second participants in communication connection. Figure 1 FIG. 1 is a flow chart of a joint modeling method provided by an embodiment of the present invention. The method can be executed by an electronic device (the electronic device corresponding to the first participant executing the solution), and the electronic device can be implemented by software and / or hardware.
[0053] In this embodiment, each first participant and each second participant include the same sample objects and different sample features, each second participant includes the same sample features and different sample objects, the samples of the first participant do not contain labels, and the samples of the second participant carry labels. The joint modeling method includes:
[0054] S1. Receive the public key in the homomorphic encryption key pair sent by each second participant in the joint modeling system, perform sample alignment processing on the locally stored first sample set and the second sample set of each second participant based on the public key, and split the first sample set into sub-sample sets corresponding to each second participant based on the sample alignment processing result.
[0055] In this embodiment, each second participant generates its own homomorphic encryption key pair and sends the public key to each first participant, where the public key is used for encryption and the private key is used for decryption. Therefore, the first participant cannot decrypt the encrypted data, fully ensuring the security of the data.
[0056] The characteristics of the homomorphic encryption algorithm are: multiple homomorphic encrypted data are calculated using a preset calculation formula to obtain the calculation result, and the calculation result is decrypted. The decrypted result is the same as the result of calculating the unencrypted original data using the same calculation formula. The homomorphic encryption algorithm can ensure the normal operation of data without leaking the original data.
[0057] In this embodiment, each first participant and each second participant contain the same sample objects (by performing sample alignment processing on the corresponding sample sets to determine the same sample objects between the first participant and the second participant) and different sample features, thereby ensuring the sufficiency of the sample features; each second participant contains the same sample features and different sample objects, thereby ensuring the sufficiency of the sample quantity.
[0058] For example, assuming that the first participants include the Social Security Bureau and a shopping platform, and the second participants include Bank 1, Bank 2, and Bank 3, the first sample set of the Social Security Bureau includes sample data with sample IDs of 1 to 10000, and the sample features corresponding to the sample data include social security contribution base, social security contribution company, social security contribution duration, etc.; the first sample set of the shopping platform includes sample data with sample IDs of 1 to 9500, and the sample features corresponding to the sample data include the number of purchases, types of purchases, and amount of purchases made by users in the past six months; the second sample set of Bank 1 includes sample data with sample IDs of 1 to 1000, and the sample data includes the number of deposits, number of loans, and number and types of financial products purchased by users in the past six months; the second sample sets of Bank 2 and Bank 3 include sample data with sample IDs of 1000 to 6000 and 6001 to 9000, respectively, and the sample features of the sample data are the same as the sample features of Bank 1.
[0059] If the first participant to implement the plan is the Social Security Bureau, the Social Security Bureau will use the sample data with sample IDs of 1 to 1000 in its first sample set (the shared sample ID with Bank 1) as the first first sub-sample set, the sample data with sample IDs of 1001 to 6000 in its first sample set (the shared sample ID with Bank 2) as the second first sub-sample set, and the sample data with sample IDs of 6001 to 9000 in its first sample set (the shared sample ID with Bank 3) as the third first sub-sample set. Since the remaining sample data in the first sample set will not be used in this modeling process, this remaining data can be transferred from the first sample set to other places for storage.
[0060] For the Social Security Bureau, the corresponding relationship between the first sub-sample set and the second participant is:
[0061] The first subsample set corresponds to Bank 1;
[0062] The second first subsample set corresponds to Bank 2;
[0063] The 3rd first sub-sample set corresponds to the bank 3 set.
[0064] Using the above method, other first participants also split their local sample sets into sub-sample sets corresponding to each second participant.
[0065] The performing sample alignment processing on the locally stored first sample set and the second sample set of each second participant based on the public key respectively includes:
[0066] A11. Select a second participant, calculate a first hash value for each sample ID in the first sample set, encrypt the first hash value using a public key in a homomorphic encryption key pair corresponding to the selected second participant to obtain a first ciphertext, and establish a mapping relationship between the first ciphertext and the sample ID;
[0067] The sample ID may be a user ID, and the user ID may be a user's mobile phone number, ID card number, or other information that identifies the user.
[0068] The security of the sample ID is fully guaranteed by performing hash operations and encryption on the sample ID.
[0069] A12. Receive a second ciphertext sent by the selected second participant, where the second ciphertext is obtained by encrypting a second hash value of each sample ID in the second sample set by the selected second participant using a public key in the same homomorphic encryption key pair;
[0070] The selected second participant also performs hash operations and encryption on its second sample set.
[0071] A13. Calculate the intersection of the first ciphertext and the second ciphertext to obtain the public sample ID ciphertext, determine the plaintext data of the public sample ID ciphertext based on the mapping relationship, and complete data alignment.
[0072] The above mapping relationship includes a mapping relationship between each sample ID in the first sample set and the corresponding first ciphertext, and the plaintext data of the public sample ID ciphertext can be queried from the mapping relationship.
[0073] The first participant who executes this scheme also sends the public sample ID ciphertext to the selected second participant, so that the selected second participant can obtain the public sample ID.
[0074] S2. Receive the first-order gradient ciphertext and the second-order gradient ciphertext of each sample in the second sample set sent by each second participant.
[0075] In this embodiment, the sample of the first participant does not contain a label, and the sample of the second participant carries a label. Therefore, the first participant needs to use the label of the second participant for joint modeling. To ensure the security of the label, the second participant encrypts the label-related information and transmits it to the first participant.
[0076] The process of determining the first-order gradient ciphertext and the second-order gradient ciphertext includes:
[0077] B11. A second participant initializes the predicted value of each sample in its second sample set;
[0078] The process of determining the first-order gradient ciphertext and the second-order gradient ciphertext of each sample in the second sample set is performed by the corresponding second participant.
[0079] In this embodiment, a prediction value is randomly initialized for each sample in the second sample set.
[0080] B12, determining the prediction error of each sample in the second sample set based on the label of each sample in the second sample set and the predicted value;
[0081] The label of the sample reflects the true value of the sample. The prediction error of each sample in the second sample set can be determined based on the label and the predicted value.
[0082] B13, calculating the first-order gradient and the second-order gradient of each sample in the second sample set based on the prediction error and the prediction value;
[0083] The calculation formula of the first-order gradient is:
[0084]
[0085] Among them, y ij is the label of the jth sample in the second sample set of the i-th second participant, is the predicted value of the jth sample in the second sample set of the i-th second participant at the (t-1)th time, is the prediction error between the predicted value and label of the jth sample in the second sample set of the i-th second participant at the (t-1)th time, is the first-order gradient of the j-th sample in the second sample set of the i-th second participant at the t-th time.
[0086] In this embodiment, at this moment t=1, is the predicted value of the jth sample in the second sample set of the i-th second participant at the 0th time (i.e., at initialization), is the first-order gradient of the jth sample in the second sample set of the i-th second participant at the first moment (i.e., the moment of building the first tree).
[0087] The calculation formula of the second-order gradient is:
[0088]
[0089] Among them, y ij is the label of the jth sample in the second sample set of the i-th second participant, is the predicted value of the jth sample in the second sample set of the i-th second participant at the (t-1)th time, is the prediction error between the predicted value and label of the jth sample in the second sample set of the i-th second participant at the (t-1)th time, is the second-order gradient of the j-th sample in the second sample set of the i-th second participant at the t-th time.
[0090] B14. Use the public key in the homomorphic encryption key pair to encrypt the first-order gradient and the second-order gradient to obtain the first-order gradient ciphertext and the second-order gradient ciphertext of each sample in the second sample set of the second participant.
[0091] S3. Assign each sub-sample set to the root node of the corresponding tree to be constructed.
[0092] This embodiment implements joint modeling of multiple participants based on the XGBoost algorithm, and each participant performs joint modeling by constructing one or more decision trees.
[0093] For example, when jointly constructing the first tree, each sub-sample set of each first participant corresponds to a tree (equivalent to an initial preset model), and each first participant also corresponds to a tree. Each party assigns the corresponding sample set or sub-sample set to the root node of the corresponding tree, combines the sample data of each tree to determine the structure of the tree, and completes the construction of the first tree.
[0094] S4. Bin each feature in the sample on each root node, and construct the target gradient histogram ciphertext of each root node based on the binning results, the first-order gradient ciphertext and the second-order gradient ciphertext.
[0095] In order to make the constructed tree more robust, continuous features need to be binned, and the next layer of tree nodes need to be constructed based on the binning results. For example, the first layer of the tree is the root node. The best segmentation features and their best segmentation points of the samples on the root node need to be determined based on the binning results, and the samples on the root node are divided into two parts accordingly. The two parts of samples are respectively assigned to the two nodes of the second layer of the tree, and then the features of the samples on each node of the second layer are binned separately. According to the binning results, the samples on each node of the second layer are divided into two parts, and respectively assigned to the two nodes of the third layer corresponding to the node (each node will have two child nodes, and the third layer of the tree has a total of 4 nodes), ...
[0096] Binning is to discretize continuous feature values into K integers and divide them into different boxes. Each participant can choose one or more binning methods according to their own characteristics. After determining the binning method, they can determine the corresponding number of bins.
[0097] The binning methods include equal frequency, equal interval, optimal and custom binning methods. Among them, the equal frequency binning method means that the number of samples allocated in each box is consistent; the equal interval binning method means that the segmented areas of each box are consistent; the optimal binning method means that recursive partitioning is used for binning, and the segmentation point is determined according to statistical tests; the custom binning method means that the user defines the segmentation point.
[0098] For example, for the Social Security Bureau, based on the social security contribution base, the samples in each sub-sample set can be divided into multiple boxes according to the equal-frequency binning method; based on the duration of social security payments, the samples in each sub-sample set can be divided into four boxes of 0 to 3 years, 4 to 6 years, 7 to 9 years, and greater than or equal to 10 years according to the custom binning method.
[0099] The target gradient histogram ciphertext of each root node is constructed based on the binning result, the first-order gradient ciphertext and the second-order gradient ciphertext, including:
[0100] C11. Select a root node of a tree corresponding to a subsample set, and sum the first-order gradient ciphertexts of the samples in each box of each feature on the root node of the selected tree to obtain the sum of the first-order gradient ciphertexts corresponding to each box of each feature;
[0101] In this embodiment, for each feature on the root node of the selected tree, the first-order gradient ciphertext sum of each sample in the box is calculated according to the box number.
[0102] C12, sum up the second-order gradient ciphertexts of the samples in each box of each feature on the root node of the selected tree to obtain the sum of the second-order gradient ciphertexts corresponding to each box of each feature;
[0103] Similarly, the second-order gradient ciphertext sum of each sample in the box is calculated according to the box number.
[0104] C13, based on the first-order gradient ciphertext and the second-order gradient ciphertext and the initial gradient histogram ciphertext corresponding to each feature on the root node of the selected tree;
[0105] Each feature corresponds to an initial gradient histogram ciphertext, which includes the sum of the first-order gradient ciphertext and the sum of the second-order gradient ciphertext of each box.
[0106] C14. Summarize the initial gradient histogram ciphertext corresponding to each feature to obtain the target gradient histogram ciphertext corresponding to the root node of the selected tree.
[0107] The initial gradient histogram ciphertext of each feature of the sample on the root node is summarized to obtain the target gradient histogram ciphertext of the root node.
[0108] Similarly, other first participants also construct the target gradient histogram ciphertext of the root node of the tree corresponding to each of their sub-sample sets according to this step; each second participant constructs the target gradient histogram of the plaintext state corresponding to its second sample set.
[0109] S5. Send the target gradient histogram ciphertext to the corresponding second participants, and receive the gradient histogram merged graph processed by the secure aggregation algorithm returned by each second participant.
[0110] In this embodiment, each first participant sends the target gradient histogram ciphertext corresponding to each of its sub-sample sets to the corresponding second participant. For example, the Social Security Bureau and the shopping platform both send the target gradient histogram ciphertext corresponding to the first sub-sample set (samples with sample IDs of 1 to 1000) to Bank 1, and send the target gradient histogram ciphertext corresponding to the second sub-sample set (samples with sample IDs of 1001 to 6000) to Bank 2, and so on.
[0111] After the second participant receives the target gradient histogram ciphertext sent by each first participant, it decrypts it using the private key in the homomorphic encryption key pair, and merges the decrypted gradient histogram with its local target gradient histogram to obtain a gradient histogram merge graph, which includes the gradient information of all features of the corresponding sample ID. The gradient histogram merge graph is processed using a secure aggregation algorithm and sent to the first participant executing this solution.
[0112] The secure aggregation algorithm is a technology that generates addition and subtraction random numbers between different participants through the Difie-Hellman key exchange technology, thereby hiding plain text information from third parties. The third party will offset the random numbers when aggregating, which will not affect the result of the operation. That is, when other participants add or subtract random numbers generated based on the Difie-Hellman key exchange technology in the data they send, the first participant who executes this scheme will offset all random numbers when aggregating (i.e. summing).
[0113] S6. Determine the features to be segmented and the points to be segmented according to the gradient histogram merging graph, and construct the next layer of tree nodes according to the features to be segmented and the points to be segmented.
[0114] The step of determining the features to be segmented and the points to be segmented according to the gradient histogram merged graph includes:
[0115] D11. For each feature in each gradient histogram, use the starting point of each box as the segmentation point and calculate the information gain value corresponding to each segmentation point.
[0116] For example, for the number of borrowing times in the gradient histogram merged graph with sample IDs of 1 to 1000, when binning, the samples are divided into 5 boxes according to 0 to 5, 6 to 10, 11 to 19, and greater than or equal to 20. Then 0, 6, 11, and 20 are used as the split points of the number of borrowing times, and the information gain value of each split point is calculated.
[0117] The calculation formula of the information gain value is:
[0118]
[0119] Among them, Gain ijis the information gain value corresponding to the jth segmentation point of the i-th feature, G Lij G is the sum of the first-order gradients of the samples whose values of the i-th feature in the corresponding sample set are less than the values of the j-th segmentation point. Rij H is the sum of the first-order gradients of the samples whose values of the i-th feature in the corresponding sample set are greater than or equal to the values of the j-th segmentation point. Lij H is the sum of the second-order gradients of the samples whose values of the i-th feature in the corresponding sample set are less than the values of the j-th segmentation point. Rij It is the sum of the second-order gradients of the samples whose values of the i-th feature in the corresponding sample set are greater than or equal to the values of the j-th segmentation point. λ and γ are hyperparameters.
[0120] For example, for the second split point of the number of borrowings (value 6), G Lij is the sum of the first-order gradients of the samples in the first box, G Rij is the sum of the first-order gradients of the samples in the last three boxes, H Lij is the sum of the second-order gradients of the samples in the first box, G Rij is the sum of the second-order gradients of the samples in the last three boxes.
[0121] D12. The segmentation point with the largest information gain value is used as the point to be segmented, and the feature corresponding to the point to be segmented is used as the feature to be segmented.
[0122] The information gain value of each segmentation point of each feature in each gradient histogram is calculated separately to determine the optimal segmentation point of the optimal segmentation feature.
[0123] The step of constructing the next layer of tree nodes according to the features to be segmented and the points to be segmented includes:
[0124] E11. If the feature to be segmented belongs to a first participant, the feature to be segmented and the point to be segmented are sent to the first participant, so that the first participant can separate the samples on the corresponding root node, and the separation result is sent to the corresponding second participant, and the corresponding second participant completes the construction of the next layer of tree nodes;
[0125] In this embodiment, the structure of the tree is maintained by the second participant. If the segmentation point is on the first participant, the first participant separates the sample on the root node of the corresponding tree into two parts according to the features to be segmented and the segmentation point, and sends the IDs of the two separated sample parts to the corresponding second participant. The corresponding second participant places the two parts of the sample on two tree nodes of the second layer respectively.
[0126] E12. If the feature to be segmented belongs to a second participant, the feature to be segmented and its point to be segmented are sent to each second participant, so that each second participant separates the samples of its root node and completes the construction of the next layer of tree nodes.
[0127] Since the second participants include the same sample features, if the feature to be segmented belongs to a second participant, the feature to be segmented and the point to be segmented are sent to each second participant, so that each second participant can separate the samples on its root node according to the feature to be segmented and the point to be segmented. For example, if the feature to be segmented is the number of loans, the segmentation points of Bank 1 are 0, 6, 11, and 20, the segmentation points of Bank 2 are 0, 8, 15, and 30, and the segmentation points of Bank 3 are 0, 12, 25, and 50, respectively. The point to be segmented (optimal segmentation point) is 8, then Bank 1, Bank 2, and Bank 3 all perform sample separation according to 8 as the point to be segmented, and construct the corresponding next-level tree nodes.
[0128] S7. When the depth of the tree node reaches a preset depth threshold, determine whether the constructed tree meets the convergence condition. If so, complete the joint modeling.
[0129] According to the above steps S4-S7, the next layer of tree nodes can be constructed. When the depth of the tree nodes reaches the depth threshold (for example, 10 layers), it is necessary to determine whether the constructed tree converges. If it converges, the joint modeling is completed.
[0130] The step of judging whether the constructed tree satisfies the convergence condition comprises:
[0131] F11, receiving the loss value processed by the security aggregation algorithm corresponding to the tree constructed by each second participant;
[0132] The following steps G1-G2 describe in detail the process of the second party calculating the loss value.
[0133] F12. Aggregate the loss values to obtain a total loss value. If the difference between the obtained total loss value and the total loss value of the last constructed tree is less than the loss threshold, the constructed tree meets the convergence condition.
[0134] If the first tree is constructed at this time, the initial total loss value is calculated based on the initial predicted value and label. If the difference between the obtained total loss value and the initial total loss value is less than the loss threshold (for example, 0.01%), the constructed tree converges.
[0135] The calculation process of the loss value includes:
[0136] G11. Each second participant calculates a weight value for each leaf node of the tree it constructs, and calculates a new prediction value for each sample in its second sample set based on the weight value;
[0137] The calculation formula of the weight value is:
[0138]
[0139] in, is the weight value of the jth leaf node in the tree constructed by the i-th second participant at the t-th time (constructing the t-th tree), is the sum of the first-order gradients of all samples on the j-th leaf node in the tree constructed by the i-th second participant at the t-th time, It is the sum of the second-order gradients of all samples on the j-th leaf node in the tree constructed by the i-th second participant at the t-th time, and λ is a hyperparameter.
[0140] The calculation formula of the new predicted value is:
[0141]
[0142] in, is the new predicted value of the a-th sample on the j-th leaf node in the tree constructed by the i-th second participant at the t-th time (i.e., the t-th tree constructed), is the predicted value of the a-th sample on the j-th leaf node in the tree constructed by the i-th second participant at the t-1th time (i.e., the t-1th tree constructed), is the weight value of the jth leaf node in the tree constructed by the i-th second participant at the t-th time, and η is the learning rate.
[0143] G12. Each second participant calculates the loss value corresponding to the tree it constructs based on the label and new prediction value of each sample in its second sample set.
[0144] Substituting the new predicted value and label of each sample into the predetermined loss function, the loss value corresponding to each second participant can be obtained.
[0145] After determining whether the constructed tree meets the convergence condition, the method further includes:
[0146] If the judgment is no, receive the new first-order gradient ciphertext and the new second-order gradient ciphertext of each sample in the second sample set sent by each second participant, and construct the next tree based on the new first-order gradient ciphertext and the new second-order gradient ciphertext.
[0147] If the currently constructed tree does not meet the convergence condition, the next tree is constructed. At this time, the prediction value corresponding to each sample (i.e., the new prediction value) has been updated, and the new first-order gradient ciphertext and new second-order gradient ciphertext of each sample can be calculated. The next tree can be constructed according to the above steps S2 to S7.
[0148] It can be seen from the above embodiments that the joint modeling method proposed in the present invention firstly assigns each sub-sample set to the root node of the corresponding tree to be constructed, bins each feature in the sample on each root node, and constructs the target gradient histogram ciphertext of each root node based on the binning result, the first-order gradient ciphertext and the second-order gradient ciphertext; then, the target gradient histogram ciphertext is sent to the corresponding second participant, and the gradient histogram merged graph processed by the secure aggregation algorithm returned by each second participant is received; then, the features to be segmented and their points to be segmented are determined according to the gradient histogram merged graph, so as to construct the next layer of tree nodes; finally, when the depth of the tree node reaches the preset depth threshold, it is determined whether the constructed tree meets the convergence condition, and if it is determined to be yes, the joint modeling is completed. The present invention unites horizontal participants with the same sample features and different sample objects and vertical participants with the same sample objects and different sample features to jointly construct one or more decision trees, thereby improving the accuracy of the joint model.
[0149] like Figure 2 , which is a schematic diagram of modules of a joint modeling device provided in one embodiment of the present invention.
[0150] The joint modeling device 100 of the present invention can be installed in an electronic device. According to the functions to be implemented, the joint modeling device 100 may include an alignment module 110, a receiving module 120, an allocation module 130, a binning module 140, a sending module 150, a determination module 160 and a judgment module 170. The module of the present invention may also be referred to as a unit, which refers to a series of computer program segments that can be executed by a processor of an electronic device and can complete fixed functions, which are stored in the memory of the electronic device.
[0151] In this embodiment, the functions of each module / unit are as follows:
[0152] The alignment module 110 is used to receive the public key in the homomorphic encryption key pair sent by each second participant in the joint modeling system, perform sample alignment processing on the locally stored first sample set and the second sample set of each second participant based on the public key, and split the first sample set into sub-sample sets corresponding to each second participant based on the sample alignment processing result.
[0153] The performing sample alignment processing on the locally stored first sample set and the second sample set of each second participant based on the public key respectively includes:
[0154] A21. Select a second participant, calculate a first hash value for each sample ID in the first sample set, encrypt the first hash value using a public key in a homomorphic encryption key pair corresponding to the selected second participant to obtain a first ciphertext, and establish a mapping relationship between the first ciphertext and the sample ID;
[0155] A22. Receive a second ciphertext sent by the selected second participant, where the second ciphertext is obtained by encrypting a second hash value of each sample ID in the second sample set by the selected second participant using a public key in the same homomorphic encryption key pair;
[0156] A23. Calculate the intersection of the first ciphertext and the second ciphertext to obtain the public sample ID ciphertext, determine the plaintext data of the public sample ID ciphertext based on the mapping relationship, and complete data alignment.
[0157] The receiving module 120 is used to receive the first-order gradient ciphertext and the second-order gradient ciphertext of each sample in the second sample set sent by each second participant.
[0158] The process of determining the first-order gradient ciphertext and the second-order gradient ciphertext includes:
[0159] B21. A second participant initializes the predicted value of each sample in its second sample set;
[0160] B22, determining the prediction error of each sample in the second sample set based on the label of each sample in the second sample set and the predicted value;
[0161] B23, calculating the first-order gradient and the second-order gradient of each sample in the second sample set based on the prediction error and the prediction value;
[0162] B24. Use the public key in the homomorphic encryption key pair to encrypt the first-order gradient and the second-order gradient to obtain the first-order gradient ciphertext and the second-order gradient ciphertext of each sample in the second sample set of the second participant.
[0163] The allocation module 130 is used to allocate each sub-sample set to the root node of the corresponding tree to be constructed.
[0164] The binning module 140 is used to bin each feature in the sample on each root node, and construct the target gradient histogram ciphertext of each root node based on the binning result, the first-order gradient ciphertext and the second-order gradient ciphertext.
[0165] The target gradient histogram ciphertext of each root node is constructed based on the binning result, the first-order gradient ciphertext and the second-order gradient ciphertext, including:
[0166] C21. Select a root node of a tree corresponding to a subsample set, add up the first-order gradient ciphertexts of the samples in each box of each feature on the root node of the selected tree, and obtain the sum of the first-order gradient ciphertexts corresponding to each box of each feature;
[0167] C22, sum up the second-order gradient ciphertexts of the samples in each box of each feature on the root node of the selected tree to obtain the sum of the second-order gradient ciphertexts corresponding to each box of each feature;
[0168] C23, based on the first-order gradient ciphertext and the second-order gradient ciphertext and the initial gradient histogram ciphertext corresponding to each feature on the root node of the selected tree;
[0169] C24. Summarize the initial gradient histogram ciphertext corresponding to each feature to obtain the target gradient histogram ciphertext corresponding to the root node of the selected tree.
[0170] The sending module 150 is used to send the target gradient histogram ciphertext to the corresponding second participants, and receive the gradient histogram merged graph processed by the security aggregation algorithm returned by each second participant.
[0171] The determination module 160 is used to determine the features to be segmented and the points to be segmented according to the gradient histogram merging diagram, and construct the next layer of tree nodes according to the features to be segmented and the points to be segmented.
[0172] The step of determining the features to be segmented and the points to be segmented according to the gradient histogram merged graph includes:
[0173] D21. For each feature in each gradient histogram, use the starting point of each box as the segmentation point and calculate the information gain value corresponding to each segmentation point.
[0174] D22. The segmentation point with the largest information gain value is used as the point to be segmented, and the feature corresponding to the point to be segmented is used as the feature to be segmented.
[0175] The step of constructing the next layer of tree nodes according to the features to be segmented and the points to be segmented includes:
[0176] E21. If the feature to be segmented belongs to a first participant, the feature to be segmented and the point to be segmented are sent to the first participant, so that the first participant can separate the samples on the corresponding root node, and the separation result is sent to the corresponding second participant, and the corresponding second participant completes the construction of the next layer of tree nodes;
[0177] E22. If the feature to be segmented belongs to a second participant, the feature to be segmented and its point to be segmented are sent to each second participant, so that each second participant separates the samples of its root node and completes the construction of the next layer of tree nodes.
[0178] The judgment module 170 is used to judge whether the constructed tree meets the convergence condition when the depth of the tree node reaches a preset depth threshold. If it is judged to be yes, the joint modeling is completed.
[0179] The step of judging whether the constructed tree satisfies the convergence condition comprises:
[0180] F21, receiving the loss value processed by the security aggregation algorithm corresponding to the tree constructed by each second participant;
[0181] F22. Aggregate the loss values to obtain a total loss value. If the difference between the obtained total loss value and the total loss value of the last constructed tree is less than the loss threshold, the constructed tree meets the convergence condition.
[0182] The calculation process of the loss value includes:
[0183] G21. Each second participant calculates a weight value for each leaf node of the tree it constructs, and calculates a new prediction value for each sample in its second sample set based on the weight value;
[0184] G22. Each second participant calculates the loss value corresponding to the tree it constructs based on the label and new prediction value of each sample in its second sample set.
[0185] After determining whether the constructed tree meets the convergence condition, the determination module 170 is further used to:
[0186] If the judgment is no, receive the new first-order gradient ciphertext and the new second-order gradient ciphertext of each sample in the second sample set sent by each second participant, and construct the next tree based on the new first-order gradient ciphertext and the new second-order gradient ciphertext.
[0187] like Figure 3 , which is a schematic diagram of the structure of an electronic device for implementing a joint modeling method provided by an embodiment of the present invention.
[0188] The electronic device 1 is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions. The electronic device 1 can be a computer, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of hosts or network servers, wherein cloud computing is a type of distributed computing, a super virtual computer composed of a group of loosely coupled computer sets.
[0189] In this embodiment, the electronic device 1 includes, but is not limited to, a memory 11, a processor 12, and a network interface 13, which can be interconnected and communicated via a system bus. The memory 11 stores a joint modeling program 10, and the joint modeling program 10 can be executed by the processor 12. Figure 3 Only the electronic device 1 having the components 11-13 and the joint modeling program 10 is shown, and it can be understood by those skilled in the art that Figure 3 The structure shown does not constitute a limitation on the electronic device 1 , and may include fewer or more components than shown in the figure, or combine certain components, or arrange the components differently.
[0190] Among them, the memory 11 includes a memory and at least one type of readable storage medium. The memory provides a cache for the operation of the electronic device 1; the readable storage medium may be a non-volatile storage medium such as a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a disk, an optical disk, etc. In some embodiments, the readable storage medium may be an internal storage unit of the electronic device 1, such as a hard disk of the electronic device 1; in other embodiments, the non-volatile storage medium may also be an external storage device of the electronic device 1, such as a plug-in hard disk equipped on the electronic device 1, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. In this embodiment, the readable storage medium of the memory 11 is generally used to store an operating system and various application software installed on the electronic device 1, such as storing the code of the joint modeling program 10 in an embodiment of the present invention, etc. In addition, the memory 11 may also be used to temporarily store various types of data that have been output or are to be output.
[0191] In some embodiments, the processor 12 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips. The processor 12 is generally used to control the overall operation of the electronic device 1, such as performing control and processing related to data interaction or communication with other devices. In this embodiment, the processor 12 is used to run the program code stored in the memory 11 or process data, such as running the joint modeling program 10.
[0192] The network interface 13 may include a wireless network interface or a wired network interface, and the network interface 13 is used to establish a communication connection between the electronic device 1 and a client (not shown in the figure).
[0193] Optionally, the electronic device 1 may further include a user interface, which may include a display (Display), an input unit such as a keyboard (Keyboard), and the optional user interface may also include a standard wired interface and a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, and an OLED (Organic Light-Emitting Diode) touch device, etc. Among them, the display may also be appropriately referred to as a display screen or a display unit, which is used to display information processed in the electronic device 1 and to display a visual user interface.
[0194] It should be understood that the embodiment is for illustration only and the scope of the patent application is not limited to this structure.
[0195] The joint modeling program 10 stored in the memory 11 in the electronic device 1 is a combination of multiple instructions, and when executed in the processor 12, the steps in the above-mentioned joint modeling method can be implemented.
[0196] Specifically, the specific implementation method of the processor 12 for the above-mentioned joint modeling program 10 can refer to Figure 1 The description of the relevant steps in the corresponding embodiments will not be repeated here.
[0197] Furthermore, if the module / unit integrated in the electronic device 1 is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. The computer-readable medium can be non-volatile or non-volatile. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a mobile hard disk, a magnetic disk, an optical disk, a computer memory, and a read-only memory (ROM).
[0198] The computer-readable storage medium stores a joint modeling program 10, and the joint modeling program 10 can be executed by one or more processors to implement the steps in the above-mentioned joint modeling method.
[0199] In the several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative, for example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation.
[0200] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0201] In addition, each functional module in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of hardware plus software functional modules.
[0202] Therefore, no matter from which point of view, the embodiments should be regarded as illustrative and non-restrictive, and the scope of the present invention is limited by the appended claims rather than the above description, so it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any attached figure mark in the claims should not be regarded as limiting the claims involved.
[0203] The blockchain referred to in this invention is a new application mode of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm, etc. Blockchain is essentially a decentralized database, a string of data blocks generated by cryptographic methods. Each data block contains a batch of network transaction information, which is used to verify the validity of its information (anti-counterfeiting) and generate the next block. Blockchain can include the underlying blockchain platform, platform product service layer, and application service layer.
[0204] In addition, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices stated in the system claim can also be implemented by one unit or device through software or hardware. The second and other words are used to indicate names, but not to indicate any particular order.
[0205] 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 spirit and scope of the technical solution of the present invention.
Claims
1. A joint modeling method, applied to any first participant in a joint modeling system, wherein the joint modeling system includes a plurality of first participants and a plurality of second participants in communication connection, It is characterized in that Each first participant and each second participant contain the same sample object and different sample features, each second participant contains the same sample features and different sample objects, the samples of the first participant do not contain labels, and the samples of the second participant carry labels, and the method includes: Receive a public key in a homomorphic encryption key pair sent by each second participant in the joint modeling system, perform sample alignment processing on the locally stored first sample set and the second sample set of each second participant based on the public key, and split the first sample set into sub-sample sets corresponding to each second participant based on the sample alignment processing result; Receiving a first-order gradient ciphertext and a second-order gradient ciphertext of each sample in the second sample set sent by each second participant; Assign each sub-sample set to the root node of the corresponding tree to be constructed; Bin each feature in the sample on each root node, and construct the target gradient histogram ciphertext of each root node based on the binning results, the first-order gradient ciphertext, and the second-order gradient ciphertext; Sending the target gradient histogram ciphertext to the corresponding second participants, and receiving the gradient histogram merged graph processed by the secure aggregation algorithm returned by each second participant; Determine the features to be segmented and the points to be segmented according to the gradient histogram merged graph, and construct the next layer of tree nodes according to the features to be segmented and the points to be segmented; When the depth of the tree node reaches the preset depth threshold, it is determined whether the constructed tree meets the convergence condition. If so, the joint modeling is completed.
2. The joint modeling method according to claim 1, It is characterized in that The target gradient histogram ciphertext of each root node is constructed based on the binning result, the first-order gradient ciphertext and the second-order gradient ciphertext, including: Select a root node of the tree corresponding to a subsample set, sum the first-order gradient ciphertexts of the samples in each box of each feature on the root node of the selected tree, and obtain the sum of the first-order gradient ciphertexts corresponding to each box of each feature; Sum up the second-order gradient ciphertexts of the samples in each box of each feature at the root node of the selected tree to obtain the sum of the second-order gradient ciphertexts corresponding to each box of each feature; Draw an initial gradient histogram ciphertext corresponding to each feature on the root node of the selected tree based on the first-order gradient ciphertext and the second-order gradient ciphertext; The initial gradient histogram ciphertext corresponding to each feature is summarized to obtain the target gradient histogram ciphertext corresponding to the root node of the selected tree.
3. The joint modeling method according to claim 1, It is characterized in that The step of determining the features to be segmented and the points to be segmented according to the gradient histogram merged graph includes: For each feature in each gradient histogram, the starting point of each box is used as the segmentation point, and the information gain value corresponding to each segmentation point is calculated; The segmentation point with the largest information gain value is taken as the point to be segmented, and the feature corresponding to the point to be segmented is taken as the feature to be segmented.
4. The joint modeling method according to claim 1, It is characterized in that The step of constructing the next layer of tree nodes according to the features to be segmented and the points to be segmented includes: If the feature to be segmented belongs to a first participant, the feature to be segmented and its segmentation point are sent to the first participant, so that the first participant can separate the samples on the corresponding root node, and send the separation result to the corresponding second participant, so that the corresponding second participant can complete the construction of the next layer of tree nodes; If the feature to be segmented belongs to a second participant, the feature to be segmented and its point to be segmented are sent to each second participant, so that each second participant separates the samples of its root node and completes the construction of the next layer of tree nodes.
5. The joint modeling method according to claim 1, It is characterized in that The step of judging whether the constructed tree satisfies the convergence condition comprises: Receive the loss value processed by the security aggregation algorithm and corresponding to the tree constructed by each second participant; Aggregate the loss values to obtain a total loss value. If the difference between the obtained total loss value and the total loss value of the last constructed tree is less than the loss threshold, the constructed tree meets the convergence condition.
6. The joint modeling method according to claim 5, It is characterized in that The calculation process of the loss value includes: Each second participant calculates a weight value for each leaf node of the tree constructed by it, and calculates a new prediction value for each sample in its second sample set based on the weight value; Each second participant calculates the loss value corresponding to the tree it constructs based on the label of each sample in its second sample set and the new prediction value.
7. The joint modeling method according to claim 1, It is characterized in that After determining whether the constructed tree meets the convergence condition, the method further includes: If the judgment is no, receive the new first-order gradient ciphertext and the new second-order gradient ciphertext of each sample in the second sample set sent by each second participant, and construct the next tree based on the new first-order gradient ciphertext and the new second-order gradient ciphertext.
8. A joint modeling device, applied to any first participant in a joint modeling system, wherein the joint modeling system includes a plurality of first participants and a plurality of second participants in communication connection, It is characterized in that Each first participant and each second participant contain the same sample object and different sample features, each second participant contains the same sample features and different sample objects, the samples of the first participant do not contain labels, and the samples of the second participant carry labels, and the device includes: An alignment module is used to receive a public key in a homomorphic encryption key pair sent by each second participant in the joint modeling system, perform sample alignment processing on a locally stored first sample set and a second sample set of each second participant based on the public key, and split the first sample set into sub-sample sets corresponding to each second participant based on the sample alignment processing result; A receiving module, used for receiving a first-order gradient ciphertext and a second-order gradient ciphertext of each sample in the second sample set sent by each second participant; An allocation module, used for respectively allocating each sub-sample set to the root node of the corresponding tree to be constructed; The binning module is used to bin each feature in the sample on each root node, and construct the target gradient histogram ciphertext of each root node based on the binning result, the first-order gradient ciphertext and the second-order gradient ciphertext; A sending module, used to send the target gradient histogram ciphertext to the corresponding second participant, and receive the gradient histogram merged graph processed by the security aggregation algorithm returned by each second participant; A determination module, used for determining the features to be segmented and the points to be segmented according to the gradient histogram, and constructing the next layer of tree nodes according to the features to be segmented and the points to be segmented; The judgment module is used to judge whether the constructed tree meets the convergence condition when the depth of the tree node reaches a preset depth threshold. If the judgment is yes, the joint modeling is completed.
9. An electronic device, It is characterized in that The electronic device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores a joint modeling program executable by the at least one processor, and the joint modeling program is executed by the at least one processor so that the at least one processor can perform the joint modeling method according to any one of claims 1 to 7.
10. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores a joint modeling program, which can be executed by one or more processors to implement the joint modeling method according to any one of claims 1 to 7.
Citation Information
Patent Citations
Federation modeling method based on migration learning, device and readable storage medium
CN109255444A
Multi-party joint modeling method and device based on federated learning and medium
CN111368901A
Multi-party joint decision tree construction method and device and readable storage medium
CN111695697A
Decision tree construction method and device based on federated learning system and electronic equipment
CN113408668A
Hybrid federated learning method and architecture
WO2021022707A1
