Target behavior prediction system based on causal inference and multi-task learning

Abstract

The invention relates to the technical field of target behavior prediction systems based on causal inference and multi-task learning, and particularly discloses a target behavior prediction system based on causal inference and multi-task learning. The system comprises a central coordination server and a plurality of participant clients, constructs a global causal graph through a federated causal discovery algorithm in a collaborative manner, determines a causal feature subset of each prediction task, and carries out multi-task model training through a federated average algorithm under the constraint. In the process, differential privacy and homomorphic encryption technologies are comprehensively applied to protect data privacy. According to the method, more accurate causal discovery and more reliable prediction model training can be realized on the premise of protecting data privacy of all parties.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence and machine learning, and particularly relates to a target behavior prediction system based on causal inference and multi-task learning. BACKGROUND

[0002] Artificial intelligence technology, especially machine learning and deep learning, has been widely applied in social computing, intelligent recommendation, financial risk control and other fields, aiming to mine potential patterns from massive data and achieve accurate prediction. Target behavior prediction, as a key research direction, is committed to analyzing users' historical interactions, environmental states and other multi-source data to predict future possible behaviors or events, thereby providing a basis for decision support.

[0003] The target behavior prediction system based on causal inference and multi-task learning aims to integrate causal reasoning framework and multi-task learning paradigm to improve the accuracy, interpretability and generalization ability of prediction. Its core goal is not only to identify statistical associations in data, but also to reveal causal mechanisms between variables and utilize the potential associations between multi-tasks for knowledge sharing and transfer.

[0004] Existing technologies usually simply concatenate or independently apply causal inference models and multi-task learning models, which has significant limitations. Although causal inference models can construct causal graphs and estimate intervention effects, they are highly dependent on high-quality, unbiased observation data. In real-world scenarios, core feature data used to construct causal relationships are often scattered and stored in different data silos due to privacy and compliance requirements, and cannot be directly aggregated for model training. This makes it difficult for causal models to obtain sufficient and unbiased data support, and the inferred causal effects may be distorted.

[0005] Traditional multi-task learning methods usually need to centralize the original data of each party when conducting cross-institution or cross-domain joint training, which can easily lead to sensitive information leakage and does not comply with increasingly stringent data security regulations. In addition, existing methods fail to effectively solve the fundamental contradiction of how to collaboratively utilize scattered, privacy-sensitive multi-task data for causal discovery and model training without directly sharing original data, which severely restricts the practical application performance in financial, medical and other scenarios with high requirements for privacy and causality. SUMMARY

[0006] The purpose of the present application is to provide a target behavior prediction system based on causal inference and multi-task learning to solve the contradiction that existing technologies cannot effectively collaboratively utilize scattered multi-task data for causal discovery and model training under data privacy constraints.

[0007] The present application provides a target behavior prediction system based on causal inference and multi-task learning, comprising:

[0008] a central coordination server, deployed with a global causal graph learning module, a global multi-task prediction model aggregation module and a secure aggregation protocol engine;

[0009] at least two participant clients, each deployed with a local causal feature extraction module, a local multi-task prediction model, a local differential privacy perturbation unit and a homomorphic encryption calculation unit;

[0010] The global causal graph learning module is configured to execute a federated causal discovery algorithm based on encrypted causal statistics received from each participant client, and construct and iteratively update a global causal graph. The global multi-task prediction model aggregation module is configured to execute a federated averaging algorithm based on encrypted model parameter updates received from each participant client, and aggregate, generate and distribute an updated global multi-task prediction model. The secure aggregation protocol engine is configured to coordinate and execute a secure multi-party computation process between each participant client.

[0011] The local causal feature extraction module is configured to calculate predefined causal statistics based on a locally stored private data set. The local multi-task prediction model is configured to be initialized by model parameters issued by the global multi-task prediction model aggregation module, and trained based on a local private data set to output prediction results for at least two related prediction tasks. The local differential privacy perturbation unit is configured to add Laplace noise or Gaussian noise conforming to a differential privacy definition to the causal statistics calculated by the local causal feature extraction module before the causal statistics are sent to the central coordination server. The homomorphic encryption calculation unit is configured to perform homomorphic encryption operations on model parameter updates generated by the local multi-task prediction model training to generate encrypted model parameter updates, and send the encrypted model parameter updates to the central coordination server.

[0012] Preferably, the execution process of the federated causal discovery algorithm is as follows:

[0013] The global causal graph learning module initializes an empty causal graph containing all shared variables of the participants.

[0014] In each iteration, the central coordination server broadcasts a conditional independence assumption to be verified to all participant clients through the secure aggregation protocol engine.

[0015] The local causal feature extraction module of each participant client calculates a conditional independence test statistic corresponding to the assumption based on its local private data set, and sends the statistic after being added with noise by the local differential privacy perturbation unit and encrypted by the homomorphic encryption calculation unit.

[0016] The secure aggregation protocol engine coordinates the participants to securely sum and decrypt all received encrypted and noisy statistics to obtain aggregated global statistics.

[0017] The global causal graph learning module determines whether to accept or reject the conditional independence assumption according to whether the global statistic is greater than a preset significance threshold, and adds, deletes or orients edges in the causal graph accordingly;

[0018] The process is iterated until the causal graph structure no longer changes or a preset maximum number of iterations is reached.

[0019] Preferably, the global causal graph learning module is further configured to:

[0020] After the completion of the federated causal discovery algorithm, a corresponding causal feature subset is generated for each prediction task in the system based on the constructed global causal graph, the causal feature subset including all direct and indirect cause variables pointing to the target variable of the task identified from the global causal graph;

[0021] The central coordination server distributes the causal feature subset corresponding to each task to each participant client as a mandatory input feature constraint for local multi-task prediction model training.

[0022] Preferably, the local multi-task prediction model uses a hard parameter sharing neural network architecture.

[0023] The hard parameter sharing neural network architecture includes a bottom layer feature encoding network shared by all prediction tasks, and multiple independent task output networks.

[0024] The shared bottom layer feature encoding network strictly limits the input layer to the causal feature subset corresponding to the prediction task to be performed by the client received from the central coordination server;

[0025] Each task output network receives an abstract feature representation from the shared bottom network and outputs a prediction result for the corresponding task.

[0026] Preferably, the federated averaging algorithm is executed as follows:

[0027] In each training round, the global multi-task prediction model aggregation module distributes the current global model parameters to all participant clients;

[0028] The local multi-task prediction model of each participant client is initialized using the distributed parameters and performs batch local training based on the local private dataset and its causal feature subset to calculate the update amount of the model parameters;

[0029] The homomorphic encryption calculation unit encrypts the update amount to generate an encrypted model parameter update amount and uploads it to the central coordination server;

[0030] The secure aggregation protocol engine coordinates a secure weighted average calculation of the encrypted model parameter update uploaded by all participants, and the weight is the proportion of the sample size of the local data set of each participant;

[0031] The calculated encrypted global update is decrypted and used by the global multi-task prediction model aggregation module to update the global model parameters.

[0032] Preferably, the secure aggregation protocol engine uses a secure multi-party summation protocol based on threshold secret sharing;

[0033] The execution process of the secure multi-party summation protocol is as follows: when the participant client needs to upload encrypted data, the homomorphic encryption calculation unit of the participant client first divides the data into multiple shares, and encrypts the shares using the public keys of other participants;

[0034] The encrypted shares are distributed to all other participant clients and the central coordination server, so that any participant or server only holds a part of the shares of the data and cannot restore the original data;

[0035] In the aggregation phase, each participant cooperates with the server to perform summation operation on the shares in the ciphertext state, and finally only outputs the plaintext of the aggregation result, but cannot obtain the specific input value of any single participant.

[0036] Preferably, the noise level added by the local differential privacy perturbation unit is controlled by a dynamic noise scale parameter based on a privacy budget allocation strategy;

[0037] The calculation process of the dynamic noise scale parameter is as follows: first, evaluate the maximum influence of the current causal statistic to be uploaded on the final global causal graph structure, which is quantified by the number of changes in the edge set of the causal graph caused by changing the value of the statistic;

[0038] Then, according to the preset total privacy budget, the consumed privacy budget and the maximum influence value, the optimal noise scale of the current round is solved by a linear programming model, so that the loss of causal discovery accuracy is reduced under the premise of meeting the overall differential privacy constraint.

[0039] Preferably, the calculation process of the conditional independence test statistic is as follows:

[0040] The local causal feature extraction module extracts the observation data of the corresponding variables from the local private data set according to the received conditional independence assumption;

[0041] Based on the extracted data, the partial correlation coefficient between variables is calculated ;

[0042] By taking the square root of the partial correlation coefficient , a statistic statistical values of distribution:

[0043]

[0044] representative statistical values, representative sample size of local data set, representative condition set number of variables in the middle.

[0045] Preferably, the system runs in a causally guided federated multi-task learning framework;

[0046] The causally guided federated multi-task learning framework divides the overall process into a causal discovery phase and a model training phase;

[0047] In the causal discovery phase, the system executes the federated causal discovery algorithm, constructs a global causal graph and determines the causal feature subset of each task, and this phase mainly consumes the privacy budget for protecting causal statistics;

[0048] In the model training phase, the system executes the federated averaging algorithm for multi-task model training under the constraint of the determined causal feature subset, and this phase relies on homomorphic encryption to protect the model parameter update quantity without consuming the differential privacy budget;

[0049] The two phases are sequentially executed and the causal feature subset remains unchanged in the training phase.

[0050] Preferably, the causally guided federated multi-task learning framework further includes a global model parameter storage area and a participant meta-information table;

[0051] The global model parameter storage area is used to store all weight and bias parameters of the current latest version of the global multi-task prediction model;

[0052] The participant meta-information table is used to record the unique identifier of each participant client, the sample size of its local data set, and the specific prediction task list that the client needs to undertake.

[0053] Compared with the prior art, the beneficial effects of the present application are:

[0054] ​1、The present application realizes the collaborative learning of global causal structure under the premise of physically dispersed data and protected privacy by designing a federated causal discovery algorithm and a secure aggregation protocol. Each participant does not need to upload the original data, but only uploads the encrypted causal statistics protected by differential privacy. The center server aggregates and infers through secure multi-party computation. This mechanism fundamentally breaks the constraints of data silos on causal discovery, enabling the system to use a wider range of data sources to build more accurate and unbiased causal graphs, providing a reliable causal feature basis for subsequent prediction.

[0055] 2、The present application requires that the local multi-task prediction model can only use the causal feature subset derived from the global causal graph as input. This design directly and rigidly embeds the conclusions of causal inference into the structure of the prediction model, ensuring that the model learns the true causal mechanism between variables rather than false correlations. This not only significantly improves the interpretability of the model's prediction results, making its decision-making consistent with causal logic, but also improves the model's generalization ability and stability in distribution change scenarios by filtering out a large number of non-causal confounding variables.

[0056] 3、The present application creatively applies differential privacy and homomorphic encryption technology in stages and differentially. In the causal discovery stage, differential privacy is used to protect the low-frequency uploaded causal statistics, trading off controllable precision loss for strong privacy protection. In the model training stage, homomorphic encryption is used to protect the high-frequency uploaded model parameter updates, achieving lossless secure aggregation. This hybrid security framework achieves an optimal balance between privacy protection, model utility, and computational efficiency, enabling the system to meet strict data compliance requirements while completing causal enhanced multi-task learning in a federated environment. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is the overall technical scheme architecture diagram of the target behavior prediction system based on causal inference and multi-task learning proposed by the present application;

[0058] Figure 2 is the core principle framework diagram of the causal-guided federated multi-task learning framework in the present application;

[0059] Figure 3 is the logical flow framework diagram of the federated causal discovery algorithm in the present application;

[0060] Figure 4 is the logical flow framework diagram of the federated average algorithm and multi-task model aggregation in the present application;

[0061] Figure 5 is the multi-level interaction relationship and data flow diagram of the secure aggregation protocol engine in the present application. DETAILED DESCRIPTION

[0062] This embodiment elaborates the specific implementation of the target behavior prediction system based on causal inference and multi-task learning. Please refer to the attached Figures 1 to 5 The overall architecture of the system includes a central coordination server and at least two participant clients, which communicate securely through the network. The central coordination server serves as the coordination and computing hub of the system, deploying a global causal graph learning module, a global multi-task prediction model aggregation module, and a secure aggregation protocol engine.

[0063] Each participant client represents an independent data holder, deploying a local causal feature extraction module, a local multi-task prediction model, a local differential privacy perturbation unit, and a homomorphic encryption calculation unit. All modules work together to achieve global causal discovery and federated multi-task model training without data leaving the local premise.

[0064] The global causal graph learning module of the central coordination server is the core function of the federated causal discovery algorithm, which builds and continuously optimizes the global causal graph describing the causal relationships between all participant shared variables. At initialization, according to the variable name list pre-negotiated and reported by all participant clients, the initial empty causal graph containing these variables as nodes but no edge connection is created. The global causal graph learning module maintains a causal graph structure object internally, which uses an adjacency matrix and adjacency table dual data structure for storage to support efficient edge query, addition, deletion, and direction determination operations.

[0065] The module maintains a queue of conditional independence assumptions to be verified, which is generated based on the causal discovery algorithm, such as using the constraint-based PC algorithm framework, initially containing all variable pairs' independence test assumptions under the empty condition set.

[0066] The global multi-task prediction model aggregation module is responsible for managing the life cycle of the global multi-task prediction model, including model initialization, parameter aggregation, version control, and distribution. The module internally maintains a global model parameter storage area, storing all weight and bias parameters of the current latest version of the model. These parameters are organized in the form of tensors and strictly correspond to the network structure definition of the local multi-task prediction model. The global multi-task prediction model aggregation module also maintains a participant meta-information table, recording the unique identifier of each participant client, the sample size of its local dataset, and the specific prediction task list that the client needs to undertake. The sample size will be used as the weight basis for subsequent federated averaging algorithm aggregation.

[0067] The secure aggregation protocol engine is the core component that ensures the privacy and security of the entire federated learning process. Please refer to the attached Figure 5The engine implements a secure multi-party computation protocol based on threshold secret sharing, specifically a secure multi-party summation protocol. The secure aggregation protocol engine includes a protocol scheduler, a share management submodule, and a ciphertext computation coordination submodule.

[0068] The protocol scheduler is responsible for initiating and coordinating the complete secure aggregation process according to different aggregation requirements during the causal discovery phase and the model training phase.

[0069] The share management submodule is responsible for distributing temporary session public key pairs for secret sharing to each participant before uploading data on the participant client, and destroying the relevant key materials after the aggregation is completed.

[0070] The ciphertext computation coordination submodule is responsible for coordinating the central coordination server and other participant clients during the aggregation process to perform specified computation operations in ciphertext state, such as addition and number multiplication, and finally organizing a sufficient number of share holders to collaborate decryption to obtain the plaintext of the aggregation result. The entire process ensures that any single entity cannot recover the original input of any single participant.

[0071] The local causal feature extraction module of the participant client directly operates on the locally stored private data set. The private data set is usually stored in table form, with each row representing a sample and each column representing a feature variable or target variable.

[0072] The local causal feature extraction module is preloaded with a list of variable names consistent with the central coordination server. Its core task is to calculate various causal statistics required by the federated causal discovery algorithm. For conditional independence testing, the module extracts the corresponding data columns from the local data set according to the specific hypothesis received, such as testing the independence of variable and variable under the given variable set.

[0073] The calculation process uses specific statistical test methods, such as partial correlation coefficient test for continuous variables or chi-square test for discrete variables. The module internally implements corresponding statistical quantity calculation functions, with input as the corresponding data vector or matrix and output as the real number value of the statistical quantity and its corresponding degree of freedom or sample size information.

[0074] In addition to the conditional independence test statistics, the module can also calculate local causal direction scores, such as scores based on the additive noise model ANM, to assist in determining the direction of the edge after determining the dependency between variables.

[0075] The local multi-task prediction model is a specific neural network model instance. Its architecture uses a hard parameter sharing mode. The architecture includes a shared bottom feature encoding network and multiple independent task output networks.

[0076] The shared bottom feature encoding network can be a multi-layer perceptron, whose input layer has exactly as many neurons as the total number of variables in the causal feature subset corresponding to the prediction task that the client is required to perform, as issued by the central coordination server. Assuming the causal feature subset contains variables, the input layer has neurons. The network can subsequently contain several fully connected layers, using activation functions for non-linear transformation, and finally output an abstract feature representation vector at the high layer. Each task output network receives this shared feature vector as input, goes through one or more task-specific fully connected layers, and finally outputs the prediction result for the corresponding task.

[0077] For regression tasks, the output layer is usually linear neurons; for classification tasks, the output layer is a Softmax layer. All parameters of the model, including the weights and biases of the shared network and all task-specific networks, are stored in the form of floating-point tensors.

[0078] The local differential privacy perturbation unit is integrated into the causal statistic upload path. The unit receives the original causal statistic value calculated by the local causal feature extraction module. It implements a noise injection function internally. The function first calls a dynamic noise scale calculation subroutine to determine the scale parameter of the noise to be added this time.

[0079] The execution logic of the dynamic noise scale calculation subroutine is as follows:

[0080] The subroutine receives the causal statistic identifier to be uploaded currently, such as the variable pair and condition set information it is testing. According to the pre-set causal graph structure sensitivity model, it evaluates the maximum possible number of edges in the global causal graph edge set that can be changed due to a unit change in the statistic value, denoted as the maximum impact value L.

[0081] The subroutine reads the total privacy budget initially allocated by the system from the local persistent storage , and the privacy budget that has been accumulated and consumed so far . Based on the remaining privacy budget and the importance weight factor of the current iteration round in the overall causal discovery algorithm , the optimal noise scale is obtained by solving a linear programming problem. The objective of this linear programming is to minimize the loss of the discriminative ability of the statistic caused by the introduction of noise, while satisfying the constraint

[0082] . After solving , the noise injection function samples noise values from a Laplace distribution or a Gaussian distribution. The density function of the Laplace distribution is ,The noise value is added to the sampled noise value to obtain a noisy statistic. This unit ensures that for any two adjacent datasets differing only in a single sample, the output distribution satisfies the approximate indistinguishability required by the differential privacy definition.

[0083] A homomorphic encryption computing unit is integrated in the model parameter update amount upload path. This unit implements an additive homomorphic encryption algorithm, such as the Paillier encryption system. The unit internally manages a pair of public and private keys, the public key for encryption and the private key for decryption. Before the start of each round of model training, the unit obtains the temporary homomorphic encryption public key for the current round session from the secure aggregation protocol engine.

[0084] When the local multi-task prediction model completes local training, the model parameter update amount is calculated The update amount is a floating-point tensor. The homomorphic encryption computing unit first normalizes each scalar parameter in , such as scaling and converting to an integer, to meet the plaintext space requirements of the encryption algorithm. Then, using the obtained public key, each parameter value after processing is independently encrypted. The encrypted parameter values constitute the encrypted model parameter update amount, which has the same data structure as the original update amount tensor, but each element is a ciphertext. The unit packages the entire encrypted tensor and attaches the necessary metadata, and sends it to the central coordination server.

[0085] The system runs in accordance with the causal-guided federated multi-task learning framework. Please refer to the attached Figure 2 This framework clearly divides the overall process into two sequentially executed stages: the causal discovery stage and the model training stage.

[0086] The core goal of the causal discovery stage is to construct a global causal graph and determine the causal feature subsets of each prediction task. Please refer to the attached Figure 3 This stage is dominated by the global causal graph learning module and heavily relies on the secure aggregation protocol engine. At the beginning of the stage, the global causal graph learning module initializes an empty causal graph and a hypothesis queue. Then it enters an iterative loop. In each iteration, the global causal graph learning module takes a conditional independence hypothesis to be tested from the hypothesis queue, such as "variable is independent of variable given the variable set ".

[0087] The central coordination server broadcasts this hypothesis to all participant clients through the protocol scheduler of the secure aggregation protocol engine. Upon receiving the hypothesis, the local causal feature extraction module of each participant client immediately extracts variables , and Collect all the corresponding observations of variables. The module calls the internal conditional independence test function to calculate. Take the partial correlation coefficient test as an example, the function first calculates the correlation coefficient of the variables and the partial correlation coefficient after controlling , and then converts it into a statistical value that obeys distribution by formula. The formula describes the conversion process:

[0088] ;

[0089] represent the sample size of the local data set, represent the number of variables in the conditional set . The calculated statistical value is the original causal statistic to be uploaded. This value is sent to the local differential privacy perturbation unit. The perturbation unit samples noise from the Laplace distribution according to the dynamic noise scale parameter of the current round, and generates the noisy statistical value . Then, the homomorphic encryption calculation unit uses the temporary public key distributed by the secure aggregation protocol engine to encrypt , obtaining the encrypted noisy statistical value . The ciphertext is sent back to the central coordination server.

[0090] The ciphertext calculation coordination submodule of the secure aggregation protocol engine, after collecting all the encrypted noisy statistical values uploaded by the participating client , for the statistics of the client , starts the secure multi-party summation protocol. Specifically, the homomorphic encryption calculation unit of each participating client splits into multiple shares using secret sharing technology. The number of shares is equal to the total number of participants. Each share is encrypted using the public keys of other participants and the server. The encrypted share is sent to the designated recipient.

[0091] At this time, each participant and server only holds encrypted shares of other participants' data and cannot decrypt any complete . Under the aggregation instruction, each participant and the server perform addition operation on all the shares they hold that belong to the same statistic in the ciphertext state. Due to the nature of homomorphic encryption, the addition operation on the ciphertext shares corresponds to the addition of the plaintext. After multiple rounds of interactive calculation, the ciphertext share combination of the aggregation result is finally obtained. Only when more than a preset threshold of share holders collaborate to decrypt can the plaintext result be recovered, i.e. the sum of all participating noisy statistical values , Also known as global statistics. This decryption process is coordinated by the secure aggregation protocol engine to ensure the aggregation result. It is safely output to the global causal graph learning module, and no participant can know the results of others' work. value.

[0092] The global cause-effect graph learning module receives the aggregated global statistics. Then, it needs to be converted into a global test statistic that can be used to determine independence. Since the sample sizes of each participant are different, and the introduction of noise affects the original... The module performs the correction and merging process for the statistical distribution. The module knows the total sample size. Sample size for each participant The module estimates the sum based on the central limit theorem and the known distribution characteristics of differential privacy noise. An approximate distribution under the null hypothesis. Then calculate the combined values. . This Compared with the preset significance threshold Compare. If Value greater than If the independence assumption is accepted, the global causal graph learning module will remove the variable from the causal graph. and The edge between, or refuse to add the edge. If Value less than or equal to Then the independence assumption is rejected, and it is believed that... and In the given Add relevant elements and retain or add borders.

[0093] For cases where edges exist but their directions are undetermined, the module may initiate a new round of local causal direction score aggregation. The process is similar: each participant calculates a direction score, adds noise, encrypts, and uploads it. After secure aggregation, the global module determines the edge direction based on the score ranking. Based on the results of this test, the global causal graph learning module updates the causal graph structure and generates new conditional independence hypotheses based on the algorithm rules, adding them to the queue. This iterative process continues until the hypothesis queue is empty, the causal graph structure remains unchanged for multiple consecutive rounds, or the preset maximum number of iterations is reached.

[0094] After the causal discovery phase, the global causal graph learning module obtains a stable global causal graph. Then, the module performs the causal feature subset derivation task. For each prediction task in the system, its target variable corresponds to a node in the causal graph. The global causal graph learning module starts from the target node and traverses all the directed edges pointing to it. All the upstream nodes visited during the traversal form the causal feature subset for the task. These node variables include the direct parents of the target variable and more indirect ancestor nodes. The central coordination server sends the causal feature subset, i.e., the list of variable names, corresponding to each task to all the participant clients. The participant clients receive and store the corresponding causal feature subset according to the list of tasks they need to perform. This list will serve as the mandatory input feature constraint in the model training phase.

[0095] After the causal discovery phase, the system enters the model training phase. This phase trains the federated multi-task prediction model under the constraint of the fixed causal feature subset. Please refer to the attached Figure 4 This phase is dominated by the global multi-task prediction model aggregation module. At the beginning of the phase, the global multi-task prediction model aggregation module initializes the global model parameters, which can be randomly initialized or pre-trained parameters, and sends the initial parameters to all participant clients. After receiving the parameters, the local multi-task prediction models of each participant client complete the model loading. The input layer of the shared underlying feature encoding network is reconfigured to only retain the input connections corresponding to the causal feature subset of the local task, and the input connection weights of irrelevant features are fixed to zero or masked.

[0096] The training is carried out in rounds. In each training round, the global multi-task prediction model aggregation module broadcasts the current global model parameters. Each participant client updates the local model using the parameters. Subsequently, each participant client performs batch training on its local private data set. The training process strictly follows the causal feature constraint: during each forward propagation, only the column data specified by the causal feature subset is extracted from the local data set and input to the model. The loss function is the sum of the multi-task losses, and the loss of each task is calculated according to its type, such as mean square error for regression and cross-entropy for classification. Through the backpropagation algorithm, the update amount of the model parameters is calculated . The update amount is a tensor that contains the parameter changes of the shared network and all local related task-specific networks.

[0097] The locally calculated is sent to the homomorphic encryption computing unit. This unit uses the homomorphic encryption public key of the current training round to encrypt each parameter value in . The encryption process ensures that it satisfies the additive homomorphism. After encryption, the encrypted model parameter update amount The ciphertext data is uploaded to the central coordination server.

[0098] The security aggregation protocol engine intervenes again to coordinate the secure aggregation of the encrypted model parameter update amount. Due to the large number of model parameters, directly performing independent secure multi-party summation on each parameter has huge overhead. Therefore, an efficient batch processing method is adopted here. Each participant client uploads the encrypted model parameter update amount , which is already a ciphertext tensor. The ciphertext calculation coordination submodule of the update amount secure aggregation protocol engine for each participant client coordinates all participants and servers to directly perform weighted average operation on the ciphertext tensor level. The weight of the weighted average is the proportion of the sample amount of each participant in the total sample amount . Since homomorphic encryption supports plaintext multiplication by a constant and ciphertext addition operations, the weighted average operation can be completed in the ciphertext state. Specifically, it is equivalent to performing a "multiply by " scalar multiplication operation on each participant's encrypted update amount , and then adding all the results. These operations are implemented through secure multi-party computation protocols, ensuring that during the calculation process, any participant cannot obtain the plaintext or intermediate results. Finally, the protocol outputs the encrypted global average update amount . This ciphertext is securely decrypted to obtain the plaintext global average update amount .

[0099] After receiving the global multi-task prediction model aggregation module , it is applied to the current global model parameters , and the new round of global model parameters is updated. The updated global model parameters are stored and used as the starting point for the next round of training. The above training round is repeated until the global model converges in performance on the validation set or reaches the preset total training round.

[0100] During the entire system operation process, the secure multi-party computation protocol provided by the security aggregation protocol engine is the cornerstone of data interaction. Please refer again to the attached Figure 5The workflow of the engine is embodied as multi-level interaction. In the application layer, the causal discovery and model training module initiates an aggregation request. In the protocol layer, the protocol scheduler parses the request, determines the data type of aggregation, the list of participants, and the aggregation function. In the cryptography layer, the share management submodule is responsible for the distribution and recovery of key materials, and the ciphertext calculation coordination submodule guides the local execution of specific ciphertext operations by each participant client through a series of standardized message exchange protocols, and specifies the routing of intermediate results between participants. The network layer is responsible for the reliable transmission of all messages. This layered design enables the secure aggregation protocol engine to provide strong privacy protection for different computing tasks in the upper layer in a unified manner.

[0101] The dynamic noise scale calculation of the local differential privacy perturbation unit is a key optimization link. The specific formalization of the linear programming model is as follows: let the maximum influence value of the current statistical quantity to be uploaded be , the remaining privacy budget be , and the weight of the current round in the algorithm be . The goal is to find the optimal noise scale to minimize the expected decision error caused by noise. This error can be modeled as a function related to . The constraint condition is that the privacy budget consumed by the current operation is less than the upper limit of the budget allocated to the current round, and is inversely proportional to . By solving this optimization problem, the system can adaptively allocate appropriate noise intensity to each round of causal verification under the hard constraint of the overall privacy budget, thereby achieving a dynamic balance between privacy protection and causal discovery accuracy.

[0102] In summary, the above-mentioned detailed module design, data flow description, algorithm flow, and secure protocol interaction are used to realize a system that can perform causal discovery and train robust multi-task prediction targets on multi-party data under the premise of privacy protection. The system strictly follows the federated multi-task learning framework guided by causality, first constructs an unbiased causal graph using secure aggregation of differentially private protected causal statistics, then updates the model using homomorphic encryption secure aggregation based on the feature subset constrained by the causal graph, and finally obtains a high-performance, interpretable, and privacy-safe prediction model.

[0103] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and illustrative figures, it should be apparent that the scope of the present application is not limited to these specific embodiments.

[0104] While the embodiments of the application have been shown and described herein, it is to be understood that the scope of the application, jointly pointed out in the appended claims, is not to be limited to the above-described embodiments but can be otherwise variously changed, modified, replaced, and altered within the principles and spirit of the present application.

Claims

1. A target behavior prediction system based on causal inference and multi-task learning, characterized in that, The method comprises the following steps: a central coordination server is deployed with a global causal graph learning module, a global multi-task prediction model aggregation module and a secure aggregation protocol engine; at least two participant clients, each of which is deployed with a local causal feature extraction module, a local multi-task prediction model, a local differential privacy perturbation unit and a homomorphic encryption calculation unit; the global causal graph learning module is configured to execute a federated causal discovery algorithm based on encrypted causal statistics received from each participant client, and to build and iteratively update a global causal graph; the global multi-task prediction model aggregation module is configured to execute a federated averaging algorithm based on encrypted model parameter updates received from each participant client, and to aggregate and generate an updated global multi-task prediction model and distribute it; the secure aggregation protocol engine is configured to coordinate and execute a secure multi-party computation process between the participant clients; 2.The target behavior prediction system based on causal inference and multi-task learning according to claim 1, wherein, the local causal feature extraction module is configured to calculate predefined causal statistics based on a locally stored private data set, the local multi-task prediction model is configured to be initialized by model parameters issued by the global multi-task prediction model aggregation module and trained based on a local private data set to output prediction results for at least two related prediction tasks, the local differential privacy perturbation unit is configured to add Laplace noise or Gaussian noise conforming to the definition of differential privacy to the causal statistics calculated by the local causal feature extraction module before the causal statistics are sent to the central coordination server, and the homomorphic encryption calculation unit is configured to perform homomorphic encryption operation on model parameter updates generated by the local multi-task prediction model training to generate encrypted model parameter updates and send them to the central coordination server. The execution process of the federated causal discovery algorithm is as follows: the global causal graph learning module initializes an empty causal graph containing all shared variables of the participants; in each iteration, the central coordination server broadcasts the conditional independence assumption to be verified to all participant clients through the secure aggregation protocol engine; the local causal feature extraction module of each participant client calculates the conditional independence test statistic corresponding to the assumption based on its local private data set, adds noise through the local differential privacy perturbation unit, and then encrypts and sends the statistic through the homomorphic encryption calculation unit; the secure aggregation protocol engine coordinates the participants to securely sum and decrypt all received encrypted and noisy statistics to obtain aggregated global statistics; the global causal graph learning module determines whether to accept or reject the conditional independence assumption based on whether the global statistics are greater than a preset significance threshold, and accordingly adds, deletes or orients edges in the causal graph; 3.The target behavior prediction system based on causal inference and multi-task learning according to claim 2, wherein, the process is iterated until the causal graph structure no longer changes or a preset maximum number of iterations is reached. The global causal graph learning module is further configured to: after the federated causal discovery algorithm is completed, generate a corresponding causal feature subset for each prediction task in the system based on the built global causal graph, wherein the causal feature subset contains all direct cause variables and indirect cause variables pointing to the target variable of the task identified from the global causal graph; The central coordination server distributes the causal feature subsets corresponding to each task to each participant client as mandatory input feature constraints during local multi-task prediction model training. 4.The target behavior prediction system based on causal inference and multi-task learning according to claim 3, wherein, The local multi-task prediction model adopts a hard parameter sharing neural network architecture; The hard parameter sharing neural network architecture includes a shared bottom feature encoding network shared by all prediction tasks, and multiple independent task output networks; The shared bottom feature encoding network strictly limits the input layer to the causal feature subset corresponding to the prediction task to be performed by the client, which is received from the central coordination server; Each task output network receives an abstract feature representation from the shared bottom network and outputs the prediction result of the corresponding task. 5.The target behavior prediction system based on causal inference and multi-task learning according to claim 1, wherein, The execution process of the federated averaging algorithm is as follows: At each training round, the global multi-task prediction model aggregation module distributes the current global model parameters to all participant clients; The local multi-task prediction model of each participant client initializes using the distributed parameters and performs batch local training based on the local private dataset and its causal feature subset to calculate the update amount of the model parameters; The homomorphic encryption calculation unit encrypts the update amount to generate an encrypted model parameter update amount and uploads it to the central coordination server; The secure aggregation protocol engine coordinates the secure weighted average calculation of all participant uploaded encrypted model parameter update amounts, with the weight being the sample size proportion of each participant's local dataset; The calculated encrypted global update amount is decrypted and used by the global multi-task prediction model aggregation module to update the global model parameters. 6.The target behavior prediction system based on causal inference and multi-task learning according to claim 1, wherein, The secure aggregation protocol engine uses a secure multi-party sum protocol based on threshold secret sharing; The execution process of the secure multi-party sum protocol is as follows: when a participant client needs to upload encrypted data, its homomorphic encryption calculation unit first divides the data into multiple shares and encrypts these shares using the public keys of other participants; The encrypted shares are distributed to all other participant clients and the central coordination server, so that any participant or server only holds a part of the shares and cannot restore the original data; In the aggregation phase, each participant and the server cooperate to perform sum operation on the shares in ciphertext state, and finally only output the plaintext of the aggregation result, but cannot obtain the specific input value of any single participant.

7. The goal behavior prediction system based on causal inference and multi-task learning according to claim 1, characterized in that, The noise level added by the local differential privacy perturbation unit is controlled by a dynamic noise scale parameter based on a privacy budget allocation strategy; The calculation process of the dynamic noise scale parameter is as follows: first, evaluate the maximum impact of the current causal statistic to be uploaded on the final global causal graph structure, which is quantified by the number of changes in the causal graph edge set caused by changing the value of the statistic; Then, according to the preset total privacy budget, the consumed privacy budget, and the maximum impact value, solve the optimal noise scale for the current round through a linear programming model, so that the loss of causal discovery accuracy is reduced under the premise of meeting the overall differential privacy constraint. 8.The target behavior prediction system based on causal inference and multi-task learning of claim 2, wherein, The calculation process of the conditional independence test statistic is as follows: The local causal feature extraction module extracts observation data of corresponding variables from the local private data set according to the received conditional independence assumption; Based on the extracted data, a partial correlation coefficient between variables is calculated ; By taking the partial correlation coefficient statistical values obeying distribution: ； a representative statistical value, a representative sample size of the local dataset, a representative condition set a representative number of variables. 9.The goal behavior prediction system based on causal inference and multi-task learning of claim 1, wherein, The system runs in a causal-guided federated multi-task learning framework; The causal-guided federated multi-task learning framework divides the whole process into a causal discovery phase and a model training phase; In the causal discovery phase, the system executes the federated causal discovery algorithm, constructs a global causal graph and determines the causal feature subset of each task. This phase mainly consumes a privacy budget for protecting causal statistics; In the model training phase, the system executes the federated averaging algorithm for multi-task model training under the constraint of the determined causal feature subset. This phase relies on homomorphic encryption to protect the model parameter update quantity without consuming the differential privacy budget; The two phases are sequentially executed and the causal feature subset remains unchanged in the training phase.

10. The goal behavior prediction system based on causal inference and multi-task learning according to claim 9, characterized in that, The causal-guided federated multi-task learning framework further includes a global model parameter storage area and a participant meta-information table; The global model parameter storage area is used to store all weight and bias parameters of the current latest version of the global multi-task prediction model; The participant meta-information table is used to record the unique identifier of each participant client, the sample quantity of its local data set and the specific prediction task list that the client needs to undertake.

Images