Multi-agent oriented collaborative system ethical risk prediction method, system, device and medium

Abstract

This application relates to a method, system, device, and medium for predicting ethical risks in multi-agent collaborative systems. The method includes: acquiring real-time interaction logs of multiple agents in the education field, obtaining a dynamic change sequence of interactions through time-series analysis; dividing time stages based on interaction frequencies exceeding a preset threshold, calculating a stage risk evolution index, and quantifying the mean value by combining it with core ethical risk indicators to generate an educational ethical risk early warning vector; based on this vector, training a Bayesian network using Bayesian estimation to obtain a joint probability distribution of risk evolution; generating risk evolution paths from the distribution using the Monte Carlo method, and identifying propagation path types using a machine learning classification model; inputting this type of time-series data into a pre-trained long short-term memory network, and weightedly fusing it with real-time interaction data to generate educational ethical risk prediction parameters. This method can improve the timeliness and accuracy of risk early warning, avoid related ethical risks, and contribute to the standardized and safe development of artificial intelligence education.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence education technology, and in particular relates to methods, systems, devices and media for predicting ethical risks in collaborative systems for multiple agents. Background Technology

[0002] With the rapid development of Artificial Intelligence in Education (AIED), multi-agent collaborative systems are increasingly widely used in educational scenarios such as teaching assistance, learning assessment, and resource recommendation, injecting strong momentum into educational innovation and efficiency improvement. However, the complex interaction process between multiple agents is accompanied by numerous ethical risks, such as data privacy leaks, transmission of algorithmic bias, damage to academic integrity, and misleading information. These risks not only affect educational fairness and students' legitimate rights but may also have a profound negative impact on the healthy development of the educational ecosystem. Existing research mainly focuses on the ethical norms and static review of single-agent systems, lacking dynamic and precise risk prediction schemes for multi-agent collaborative scenarios. This makes it difficult to effectively capture the temporal evolution characteristics and propagation patterns of risks during interaction, resulting in delayed risk warnings and limited control effects. Against this backdrop, constructing an ethical risk prediction method adapted to the characteristics of multi-agent collaboration in the education field, enabling early identification, dynamic tracking, and accurate prediction of risks, has become a key issue that urgently needs to be addressed in the current field of AI education ethics governance. Summary of the Invention

[0003] Therefore, it is necessary to provide a method, system, device, and medium for predicting ethical risks in multi-agent collaborative systems that can effectively avoid the infringement of educational fairness and students' legitimate rights and interests by ethical risks such as data privacy leakage, algorithmic bias transmission, and damage to academic integrity, and to help the standardized and safe development of the field of artificial intelligence education.

[0004] Firstly, this application provides a method for predicting ethical risks in multi-agent cooperative systems, including:

[0005] We acquire real-time data logs of multi-agent interactions in the education field, and use time-series analysis to process the real-time data logs to obtain the dynamic change sequence of the interaction process.

[0006] The dynamic change sequence is divided into time stage boundaries. The differentiated structure of time stages is extracted from the time stage boundaries, and the stage risk evolution index of each time stage is calculated to generate an educational ethics risk early warning vector.

[0007] Based on the early warning vector of educational ethics risks, the dependency relationship between different risk indicators is modeled by Bayesian network to obtain the joint probability distribution of the evolution of educational ethics risks.

[0008] The Monte Carlo method is used to generate multiple risk evolution paths for the joint probability distribution. These risk evolution paths are then input into a machine learning classification model to determine the propagation path type of the current educational ethics risk evolution.

[0009] By inputting the propagation path type into a pre-trained long short-term memory network to construct a temporal risk value sequence, the statistical features of the temporal risk value sequence are extracted and weighted and fused with real-time interaction data of multiple agents in the field of education to generate educational ethics risk prediction parameters.

[0010] In one embodiment, the dynamic change sequence is divided into time stage boundaries, the differentiated structure of the time stages is extracted from the time stage boundaries, and the stage risk evolution index of each time stage is calculated to generate an educational ethics risk early warning vector, including:

[0011] Obtain interaction frequency data at each time point in the dynamic change sequence of multi-agent interactions in the field of education.

[0012] The interaction frequency data is compared with the preset threshold one by one. If the interaction frequency data at the current time point exceeds the preset threshold, the time point is defined as the starting boundary of the new stage.

[0013] By using the established starting boundaries of each stage, the dynamic change sequence is segmented to obtain multiple continuous time stage sequences.

[0014] Interaction pattern features, educational subject participation features, and ethically sensitive word features are extracted for each time stage sequence and integrated to form a differentiated structure corresponding to each time stage.

[0015] From the differentiated structures at different time stages, we extracted the characteristics of risk transmission intensity, risk accumulation trend, frequency of ethical conflicts between subjects, and delay in intervention response.

[0016] Based on the characteristics of risk propagation intensity and risk accumulation trend, the stage risk evolution index for each time period is calculated using a preset algorithm.

[0017] The risk evolution index of each time stage is weighted and fused with the corresponding inter-subject ethical conflict frequency characteristics to generate an educational ethics risk warning vector for each time stage.

[0018] In one embodiment, the stage risk evolution index is calculated using the following formula:

[0019]

[0020] in, This represents the stage risk evolution index, with a value range of [value range missing]. , This represents the quantitative value of the risk transmission intensity characteristic, with a value range of [0, 10]. This represents a quantitative value indicating the cumulative trend of risk, with a value range of [value range missing]. , The weighting coefficient representing the intensity of risk transmission has a range of values. , The coefficient representing the synergistic adjustment of risk propagation and accumulation, with a value range of [value missing]. .

[0021] In one embodiment, based on the educational ethics risk warning vector, a Bayesian network is used to model the dependencies between different risk indicators to obtain the joint probability distribution of the evolution of educational ethics risks, including:

[0022] A Bayesian network structure is constructed based on the risk indicators in the educational ethics risk warning vector.

[0023] Bayesian network structures represent the conditional dependencies between different risk indicators.

[0024] A Bayesian estimation method based on historical sample data of educational ethics risks was used to train the Bayesian network structure and determine the conditional probability table corresponding to each node in the network.

[0025] Conditional probability tables quantify the probabilistic relationships of conditional dependencies between nodes.

[0026] By inputting the early warning vector of educational ethics risk into a Bayesian network with a defined conditional probability table, and calculating the posterior probability distribution corresponding to the early warning vector, the joint probability distribution of the evolution of educational ethics risk is obtained.

[0027] The joint probability distribution reflects the probabilistic laws governing the coordinated evolution of various risk indicators.

[0028] In one embodiment, the Monte Carlo method is used to generate multiple risk evolution paths from the joint probability distribution. These risk evolution paths are then input into a machine learning classification model to determine the propagation path type of the current educational ethics risk evolution, including:

[0029] Based on the joint probability distribution, the Monte Carlo method is used to conduct multiple rounds of random simulation to generate multiple time-series risk evolution paths; each risk evolution path records the activation state sequence of each educational ethics risk node at each moment.

[0030] For each risk evolution path, risk propagation pattern characteristics are extracted; these characteristics include the risk outbreak time point, the set of educational entities with the greatest impact, and the time sequence of risk propagation speed.

[0031] A clustering algorithm based on Euclidean distance was used to perform cluster analysis on the propagation pattern characteristics of all paths, resulting in several typical pattern categories of educational ethics risk propagation.

[0032] A machine learning classification model is constructed based on typical pattern categories as a risk propagation pattern recognition model. The input of the risk propagation pattern recognition model is the activation state vector of the current educational ethics risk node, and the output is the typical pattern category of risk propagation.

[0033] After converting the real-time collected sequence of educational ethics risk events into risk node activation state vectors, the data is input into a risk propagation pattern recognition model for pattern matching to determine the propagation path type to which the current educational ethics risk evolution belongs.

[0034] In one embodiment, the educational ethics risk prediction parameters are obtained through the following steps:

[0035] Obtain time series data corresponding to the propagation path type; the time series data contains risk propagation characteristics of multiple consecutive time stages.

[0036] Time series data is input into a pre-trained long short-term memory network for time series modeling, and the corresponding educational ethics stage risk value is output for each time stage.

[0037] By linking the risk values ​​of each stage in chronological order, a time-series risk value sequence is constructed.

[0038] A sliding window analysis is performed on the time-series risk value sequence using a preset time window length to calculate the fluctuation range of the risk value within each window.

[0039] If the fluctuation range of a consecutive set of positive integer time periods exceeds the preset risk fluctuation threshold, then the statistical characteristics of the current time-series risk value sequence are extracted.

[0040] By weighted and fused statistical features with real-time interaction data of multiple agents in the field of education, updated parameters for predicting educational ethics risks are generated.

[0041] Real-time interaction process data includes interaction logs, feature extraction results, and risk node activation status data.

[0042] Secondly, this application also provides an ethical risk prediction system for multi-agent cooperative systems, the system comprising:

[0043] The log time-series analysis module is used to acquire real-time data logs of multi-agent interaction behavior in the education field. It uses time-series analysis methods to process the real-time data logs and obtain the dynamic change sequence of the interaction process.

[0044] The stage risk early warning module is used to divide the dynamic change sequence into time stage boundaries, extract the differentiated structure of time stages from the time stage boundaries, calculate the stage risk evolution index of each time stage, and generate an educational ethics risk early warning vector.

[0045] The risk path identification module is used to model the dependencies between different risk indicators based on the educational ethics risk warning vector through a Bayesian network to obtain the joint probability distribution of the evolution of educational ethics risks. It is also used to generate multiple risk evolution paths from the joint probability distribution using the Monte Carlo method, and input the risk evolution paths into a machine learning classification model to determine the propagation path type of the current educational ethics risk evolution.

[0046] The temporal parameter generation module is used to input the propagation path type into the pre-trained long short-term memory network to construct a temporal risk value sequence, extract the statistical features of the temporal risk value sequence and weight and fuse it with the real-time interaction process data of multiple agents in the field of education to generate educational ethics risk prediction parameters.

[0047] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0048] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method.

[0049] The aforementioned method, system, computer equipment, and storage medium for predicting ethical risks in multi-agent collaborative systems acquire real-time data logs of the interactive behaviors of multiple agents in the education field. These logs cover multiple agents (such as AI teaching tools, student-side systems, teacher-side platforms, and educational management AI). The system collects key information such as the frequency of interactions between systems (e.g., data transmission content, operation command feedback), and uses time-series analysis to structure the real-time data logs, extracting interaction correlation features along the time dimension to obtain a dynamic change sequence of the interaction process. Based on this dynamic change sequence, time stage boundaries are defined with the interaction frequency exceeding a preset threshold as the starting condition for a new stage. Interaction pattern features, educational subject participation features, and ethically sensitive word features are extracted from the sequence fragments corresponding to each time stage boundary, and integrated to form a differentiated structure for each time stage. Then, a stage risk evolution index for each time stage is calculated using a preset formula. Combined with the quantitative mean of core ethical risk indicators (data privacy leakage, algorithmic bias, damage to academic integrity, ambiguity of responsibility, etc.), an educational ethics risk warning vector is generated. Based on this educational ethics risk warning vector, a Bayesian network is trained using Bayesian estimation to model the conditional dependencies between different risk indicators, determine the conditional probability table corresponding to each node in the network, and proceed with further analysis. The joint probability distribution of the evolution of educational ethics risks is calculated. For this joint probability distribution, a random perturbation coefficient adapted to the educational scenario is incorporated for simulation. A Monte Carlo method is used to generate multiple risk evolution paths. Each path records the activation state sequence of each risk node at each moment. Features such as the risk outbreak time point, the set of educational entities with the greatest impact, and the propagation speed time sequence of each path are extracted. These features are input into a machine learning classification model for clustering and identification to determine the propagation path type of the current educational ethics risk evolution. The time series data corresponding to the propagation path type is input into a pre-trained Long Short-Term Memory network for time-series modeling. The educational ethics stage risk values ​​corresponding to each time stage are output and concatenated in chronological order to construct a time-series risk value sequence. Statistical features such as the mean, variance, trend slope, and peak density of this sequence are extracted. These statistical features are then weighted and fused with real-time interaction data of multiple agents in the education field to generate educational ethics risk prediction parameters.This method focuses on the ethical risk governance needs of multi-agent collaborative scenarios in the education field. Through real-time data collection and time-series analysis, it accurately captures the dynamic evolution characteristics of multi-agent interaction processes, solving the problem that existing studies often focus on static review of single agents and are difficult to adapt to complex collaborative scenarios. By using time stage segmentation and differentiated structure extraction, combined with Bayesian networks and Monte Carlo simulation, it clearly characterizes the dependencies and propagation patterns between risk indicators, overcoming the limitation of traditional methods in effectively tracking the temporal evolution path of risks. Through temporal modeling of long short-term memory networks and real-time data fusion optimization, it achieves accurate prediction of risk values ​​and dynamic parameter updates, improving the timeliness and accuracy of educational ethical risk warnings. It provides education administrators with a practical and dynamic control basis, effectively avoiding the infringement of educational fairness and students' legitimate rights by ethical risks such as data privacy leakage, algorithmic bias transmission, and damage to academic integrity, and contributing to the standardized and safe development of the field of artificial intelligence education. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A flowchart of an ethical risk prediction method for multi-agent cooperative systems provided in an embodiment of the present invention;

[0052] Figure 2 This is a structural block diagram of an ethical risk prediction system for multi-agent collaborative systems provided in an embodiment of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] In one embodiment, such as Figure 1 As shown, this application provides a method for predicting ethical risks in multi-agent cooperative systems, which may include the following steps:

[0055] Step S101: Obtain real-time data logs of the interaction behavior of multiple agents in the field of education, and process the real-time data logs using time series analysis methods to obtain the dynamic change sequence of the interaction process.

[0056] Specifically, multi-agent interaction encompasses various AI systems and related entities involved in the interaction within the education field, including AI teaching assistants, intelligent learning tools for students, AI management platforms for teachers, and educational data processing systems. The acquired real-time interaction data logs must fully record the core information of the interaction between the multi-agents, specifically including the timestamp of the interaction, the identifiers of the participating agents, the type of interaction (e.g., data transmission, command issuance, result feedback, function call), details of the interaction content (e.g., text data, operation parameters, feedback results), and the interaction status (success, failure, anomaly). When processing using time-series analysis, the original logs are first cleaned to remove invalid data (e.g., duplicate logs, logs with format errors) and outliers. Then, the cleaned data is structured and organized according to the time dimension, extracting the interaction correlation features at each time point (e.g., interaction frequency, interaction type ratio, data transmission volume, etc. within the same time window). This ultimately forms a dynamic change sequence with time as the axis, containing key interaction features. This sequence can intuitively reflect the temporal evolution trajectory of the multi-agent interaction process.

[0057] Step S102: Divide the dynamic change sequence into time stage boundaries, extract the differentiated structure of time stages from the time stage boundaries, calculate the stage risk evolution index of each time stage, and generate an educational ethics risk early warning vector.

[0058] Furthermore, when dividing the dynamic change sequence into time stage boundaries, a preset interaction frequency threshold is used as the criterion. When the interaction frequency at a certain time point in the sequence exceeds this threshold, it is defined as the starting boundary of a new stage. Based on this, the dynamic change sequence is divided into multiple continuous time stage sequences. For each time stage sequence, three types of core features are extracted to form a differentiated structure for the time stage: first, interaction mode features (such as the proportion of synchronous / asynchronous interaction, the frequency of one-to-one / many-to-many interaction, etc.); second, educational subject participation features (such as the participation degree of different subjects such as students, teachers, and administrators, the frequency of interaction triggers, etc.); and third, ethically sensitive word features (such as the frequency and distribution of keywords related to discrimination, false information, and privacy). Based on the extracted features, a preset formula is used to calculate the stage risk evolution index for each time stage. Then, combined with the quantitative mean of core ethical risk indicators such as data privacy leakage, algorithmic bias, damage to academic integrity, and ambiguity of responsibility attribution, a complete and quantitatively clear educational ethical risk early warning vector is generated.

[0059] Step S103: Based on the educational ethics risk warning vector, the dependency relationship between different risk indicators is modeled through a Bayesian network to obtain the joint probability distribution of the evolution of educational ethics risks.

[0060] When constructing a Bayesian network based on an educational ethics risk warning vector, each core ethical risk indicator in the warning vector is used as a network node, and the parameters of the Bayesian network are trained using Bayesian estimation. During training, historical sample data of ethical risks in the education field are used as a basis to learn the conditional dependencies between different risk indicators, such as the causal relationship between algorithmic bias and information accuracy deviation, and the logical relationship between data privacy leakage and ambiguity in responsibility attribution. This allows for the determination of the conditional probability table corresponding to each node in the network, which quantifies the specific probability values ​​of the conditional dependencies between different nodes. By inputting the educational ethics risk warning vector into the trained Bayesian network, the posterior probability distribution corresponding to the vector is calculated, ultimately yielding the joint probability distribution of the evolution of educational ethics risks. This distribution clearly reflects the probabilistic laws governing the co-evolution of each core risk indicator during the interaction process.

[0061] Step S104: Use the Monte Carlo method to generate multiple risk evolution paths for the joint probability distribution, and input the risk evolution paths into the machine learning classification model to determine the propagation path type of the current educational ethics risk evolution.

[0062] When processing the joint probability distribution of the evolution of educational ethics risks, the core parameters of the distribution are first clarified: the state transition probability of risk nodes (derived based on historical data and Bayesian networks) and the random perturbation coefficient adapted to the educational scenario (used to simulate uncertainties in actual interactions). A Monte Carlo method is used for multiple rounds of stochastic simulation. Each round generates a time-series risk evolution path based on the joint probability distribution. Each path must fully record the activation state (activated / inactive) and state changes of each risk node at every moment. For the generated multiple risk evolution paths, key propagation pattern features are extracted, including the risk outbreak time point (timestamp of the first risk node activation), the set of educational subject nodes with the greatest impact (core agents and related subjects affected by the risk), and the propagation speed time series (data on the change in the number of activated risk nodes per unit time). These features are input into a machine learning classification model. First, a clustering algorithm is used to classify the features, dividing them into several typical risk propagation pattern categories. Then, the model identifies the propagation path type to which the current educational ethics risk evolution belongs, providing targeted input for subsequent risk value prediction.

[0063] Step S105: Input the propagation path type into the pre-trained long short-term memory network to construct a time-series risk value sequence, extract the statistical features of the time-series risk value sequence and weight and fuse it with the real-time interaction process data of multiple agents in the field of education to generate educational ethics risk prediction parameters.

[0064] The identified propagation path types are used as input data, and their corresponding time series data contains risk propagation characteristics (such as propagation speed, scope of impact, and risk intensity) across multiple consecutive time stages under that propagation type. This time series data is input into a pre-trained Long Short-Term Memory (LSTM) network. Leveraging the network's modeling capabilities for time series data, the time dependencies of risk propagation are processed, and the corresponding educational ethics stage risk values ​​for each time stage are output. These risk values ​​are then concatenated in chronological order to construct a time-series risk value sequence, which accurately represents the evolution trend of risk over time. Statistical features are extracted from the time-series risk value sequence, including key indicators such as the mean, variance, trend slope, and peak density. Simultaneously, real-time interaction data of multiple agents in the education field is collected, encompassing raw interaction log data, extracted feature results, and real-time activation status data of each risk node. The above statistical features are then weighted and fused with the real-time interaction data, assigning corresponding weights based on data importance. Finally, updated educational ethics risk prediction parameters are generated, which can be directly used for subsequent accurate prediction and dynamic management of educational ethics risks.

[0065] The aforementioned method for predicting ethical risks in multi-agent collaborative systems first acquires real-time logs of interactions, data transmission content, and operational command feedback from multiple agents, including AI teaching tools and student-side systems. Time-series analysis is then used to extract time-related interaction features, resulting in a dynamic interaction change sequence. Time phase boundaries are defined based on interaction frequencies exceeding preset thresholds. Interaction patterns, educational subject participation, and ethically sensitive word features are extracted from the sequence fragments to form a differentiated structure. A phase risk evolution index is calculated using a preset formula. The mean value is quantified by combining core ethical risk indicators such as data privacy leakage and algorithmic bias, generating an educational ethical risk warning vector. Based on this warning vector, a Bayesian network is trained using Bayesian estimation to model the conditional dependencies between risk indicators and determine a conditional probability table, resulting in a joint probability distribution of risk evolution. Random perturbation coefficients from the educational scenario are incorporated, and a Monte Carlo method is used to generate multiple risk evolution paths. Path features are extracted and identified using a machine learning classification model. The time-series data corresponding to this type is input into a pre-trained long short-term memory network, outputting phase risk values ​​and constructing a time-series risk value sequence. Statistical features are extracted and weighted with real-time interaction data to generate educational ethical risk prediction parameters. This method addresses the shortcomings of existing research, which focuses on static review of single agents and is difficult to adapt to complex collaborative scenarios. It clearly characterizes the dependencies and propagation patterns of risk indicators, overcomes the limitations of traditional methods in tracking the temporal evolution path of risks, improves the timeliness and accuracy of risk warnings, provides educational administrators with a basis for dynamic management, avoids the infringement of educational equity and students' rights by related ethical risks, and promotes the standardized and safe development of artificial intelligence education.

[0066] In one embodiment, dividing the dynamic change sequence into time stage boundaries, extracting the differentiated structure of time stages from the time stage boundaries, calculating the stage risk evolution index of each time stage, and generating an educational ethics risk early warning vector may include the following steps:

[0067] Step S201: Obtain the interaction frequency data at each time point in the dynamic change sequence of multi-agent interactions in the education field.

[0068] Step S202: Compare the interaction frequency data with the preset threshold one by one. If the interaction frequency data at the current time point exceeds the preset threshold, then define the time point as the starting boundary of the new stage.

[0069] Step S203: Using the determined starting boundaries of each stage, the dynamic change sequence is segmented to obtain multiple continuous time stage sequences.

[0070] Step S204: Extract interaction pattern features, educational subject participation features, and ethically sensitive word features from each time stage sequence, and integrate them to form a differentiated structure corresponding to each time stage.

[0071] Step S205: Extract the characteristics of risk transmission intensity, risk accumulation trend, frequency of ethical conflict between subjects, and delay in intervention response from the differentiated structures of each time stage.

[0072] Step S206: Based on the characteristics of risk propagation intensity and risk accumulation trend, the stage risk evolution index for each time period is calculated using a preset algorithm.

[0073] Step S207: The stage risk evolution index of each time stage is weighted and fused with the corresponding inter-subject ethical conflict frequency characteristics to generate an educational ethics risk warning vector for each time stage.

[0074] Preferably, the two types of core data to be fused both originate from the extracted time-phase differentiated structures: the first is the phase risk evolution index for each time phase, which comprehensively reflects the co-evolutionary trend of risk propagation and accumulation within the phase; the second is the corresponding inter-subject ethical conflict frequency characteristics, that is, the frequency and density quantification value of conflicts between subjects (students, teachers, agent developers, etc.) caused by ethical issues such as data privacy, algorithmic bias, and academic integrity during multi-agent interaction within the time phase, with a value range of [0,10]. Before fusion, the two types of data need to be normalized to eliminate the influence of dimensional differences on the fusion results. After normalization, the data values ​​are uniformly mapped to the [0,1] interval. The weights for weighted fusion are set based on the ethical risk priorities of multi-agent collaborative scenarios in the education field, combined with the principles of "fairness" and "responsibility" in technological ethics and the core requirements of utilitarian ethics. These weights are jointly determined by industry ethics experts and education administrators. For example, if the impact of ethical conflicts between agents on educational equity is more direct, the weight of the frequency characteristic of ethical conflicts between agents can be set to 0.5, and the weight of the stage risk evolution index can be set to 0.5, satisfying α + β = 1 (α is the weight of the stage risk evolution index, β is the weight of the frequency characteristic of ethical conflicts between agents, α, β ∈ (0, 1)). This is achieved through the formula... (in For the fused single-dimensional risk value, For the normalized stage risk evolution index, After calculating the single-dimensional fusion result (for the normalized intersubjective ethical conflict frequency characteristics), and combining the quantitative mean of core ethical risk indicators such as data privacy leakage, algorithmic bias, damage to academic integrity, and ambiguity of responsibility attribution, a multi-dimensional structured educational ethical risk early warning vector is constructed. Each dimension in this vector corresponds to a core ethical risk, and the quantitative value intuitively reflects the severity of the risk at the corresponding time stage.

[0075] Specifically, the interaction frequency data at each time point in the dynamic change sequence of multi-agent interactions in the education field is obtained. This interaction frequency data is compared one by one with a preset threshold. If the interaction frequency data at the current time point exceeds the preset threshold, that time point is defined as the starting boundary of a new stage. Using the determined starting boundaries of each stage, the dynamic change sequence is segmented to obtain multiple continuous time stage sequences. For each time stage sequence, interaction mode features (such as the proportion of synchronous / asynchronous interactions, frequency of one-to-one / many-to-many interactions, etc.), educational subject participation features (such as the participation level of students, teachers, administrators, etc., and interaction trigger frequency, etc.), and ethically sensitive word features (such as the frequency and distribution of keywords involving privacy, discrimination, and false information) are extracted and integrated to form a differentiated structure corresponding to each time stage. From the differentiated structure of each time stage, risk propagation intensity features, risk accumulation trend features, inter-subject ethical conflict frequency features, and intervention response delay features are extracted. Based on the risk propagation intensity features and risk accumulation trend features, a preset algorithm (such as...) is used to... ,in Quantifying the intensity of risk transmission Quantify the cumulative trend of risk. For weighting coefficients, The stage risk evolution index for each time stage is calculated using the coordination adjustment coefficient. The stage risk evolution index for each time stage is then weighted and fused with the corresponding inter-subject ethical conflict frequency characteristics according to preset weights to generate an educational ethics risk early warning vector for each time stage. The vector fully covers the quantitative characteristics and correlations of the core ethical risks in that stage.

[0076] This embodiment is based on dynamic data of multi-agent interactions. It achieves precise definition of time stages through time-series threshold division, ensuring the relevance and independence of risk feature extraction at each stage. By extracting interaction features, differentiated structures, and risk-related features in a hierarchical manner, it comprehensively covers the key influencing factors of ethical risks in multi-agent collaborative scenarios, avoiding feature omissions or redundancy. The generated educational ethics risk early warning vector realizes the quantitative integration and structured presentation of risk-related features, providing high-quality input data for subsequent Bayesian network modeling, risk evolution path analysis, and other steps. It effectively solves the problem that traditional methods are difficult to capture staged risk features in dynamic interactions, laying a solid foundation for the early identification and accurate assessment of ethical risks in multi-agent collaborative systems in the education field, and helping to improve the scientific and forward-looking nature of ethical risk governance.

[0077] In one embodiment, the stage risk evolution index can be calculated using the following formula:

[0078]

[0079] in, This represents the stage risk evolution index, with a value range of [value range missing]. , This represents the quantitative value of the risk transmission intensity characteristic, with a value range of [0, 10]. This represents a quantitative value indicating the cumulative trend of risk, with a value range of [value range missing]. , The weighting coefficient representing the intensity of risk transmission has a range of values. , The coefficient representing the synergistic adjustment of risk propagation and accumulation, with a value range of [value missing]. .

[0080] Preferably, The quantitative value of risk propagation intensity is defined as the normalized result of the frequency and scope of the spread of ethical risks among multiple agents (such as AI teaching assistants, student-side intelligent tools, teacher-side AI systems, and education management AI platforms) during this time period.

[0081] Example: If, within a given period, "the risk of algorithmic bias propagates from the AI ​​assessment system to the AI ​​recommendation system, and affects more than three types of educational stakeholders (students, teachers, and administrators), then..." Take 8-10; if it only propagates within a single intelligent body... Take 1-3.

[0082] The quantitative value of risk accumulation trend characteristics is defined as: the ratio of the cumulative value of risk characteristic values ​​(such as the frequency of occurrence of ethically sensitive words and the number of subject conflicts) to the duration of the period, and the normalized result of the growth rate relative to the previous period.

[0083] Example: If the cumulative risk value in this stage increases by 50% compared to the previous stage, then Take 7; if there is no growth, then... Take 2; negative growth (risk mitigation) then Take 0.

[0084] The weighting coefficients for the intensity of risk transmission satisfy + (1- = 1, dynamically adjusted according to the characteristics of the educational scenario, such as when there is dense interaction among multiple agents (e.g., in a real-time collaborative teaching scenario in the classroom). Take 0.6-0.7 (focusing on the risk of transmission); for non-real-time interactive scenarios (such as after-school AI homework correction + AI tutoring collaboration). Take 0.3-0.4 (focusing on cumulative risk).

[0085] The synergistic moderating coefficient of risk propagation and accumulation reflects the interactive amplification effect of "propagation intensity" and "accumulation trend" (e.g., when high propagation intensity is superimposed with a high accumulation trend, the risk evolution speed will increase non-linearly). The default value is 0.3, which can be adjusted according to the type of educational ethics risk (e.g., data privacy risks). A score of 0.4 is assigned to academic integrity-related risks. (Take 0.2) for fine-tuning.

[0086] This embodiment, through a design combining linear weighting and synergistic terms, not only independently reflects the core impact of risk propagation intensity and cumulative trend, but also accurately captures the amplifying effect of their interaction on risk evolution, overcoming the limitation of traditional single-weighting methods that ignore the interactive influence of risk factors. All parameters in the formula are derived from quantifiable data such as multi-agent interaction logs and stage-differentiated structures, with clear value ranges and logical calculations, ensuring the objectivity and repeatability of the stage-based risk evolution index. The quantified EI value provides a standardized core indicator for the subsequent generation of educational ethics risk early warning vectors, effectively supporting the stage-based assessment and early identification of risks, and providing a precise and operable quantitative basis for the dynamic management of ethical risks in multi-agent collaborative systems.

[0087] In one embodiment, the joint probability distribution of the evolution of educational ethics risks is obtained by modeling the dependencies between different risk indicators using a Bayesian network based on the educational ethics risk warning vector. This may include the following steps:

[0088] Step S301: Construct a Bayesian network structure based on the risk indicators in the educational ethics risk warning vector.

[0089] Among them, the Bayesian network structure represents the conditional dependencies between different risk indicators.

[0090] Step S302: The Bayesian network structure is trained using Bayesian estimation based on historical sample data of educational ethics risks, and the conditional probability table corresponding to each node in the network is determined.

[0091] Among them, the conditional probability table quantifies the probabilistic relationship of conditional dependence between nodes.

[0092] Step S303: Input the educational ethics risk warning vector into the Bayesian network with the determined conditional probability table, and obtain the joint probability distribution of the evolution of educational ethics risk by calculating the posterior probability distribution corresponding to the warning vector.

[0093] Among them, the joint probability distribution reflects the probabilistic law of the coordinated evolution of various risk indicators.

[0094] Specifically, based on the core educational ethics risk indicators (including data privacy leakage, algorithmic bias, damage to academic integrity, ambiguity of responsibility attribution, and information accuracy deviation) contained in the educational ethics risk warning vector, a Bayesian network structure is constructed. This Bayesian network structure uses each core risk indicator as an independent network node. Through the connection relationship between nodes, it intuitively and accurately represents the conditional dependency relationship between different risk indicators. For example, the causal dependency relationship between algorithmic bias and information accuracy deviation, and the logical dependency relationship between data privacy leakage and ambiguity of responsibility attribution, clarify the association attribute of each risk indicator as a parent node or child node. A Bayesian estimation method based on historical sample data of educational ethics risks is used to train the parameters of the constructed Bayesian network structure. The historical sample data of educational ethics risks includes risk warning vectors generated in the past multi-agent interactions in the field of education, the occurrence of risks in the corresponding time period, quantitative data of each risk indicator, and records of correlation between indicators. During the training process, statistical analysis and probabilistic learning of the historical sample data are used to explore the conditional probability rules of each network node (risk indicator) under different combinations of parent node states. Then, the conditional probability table corresponding to each risk indicator node in the network is determined. The conditional probability table clearly presents the conditional probability of each child node in different states that it may be in all its parent nodes in the form of quantitative probability values, so as to realize the quantitative representation of the conditional dependency relationship between nodes. The generated educational ethics risk warning vector is used as input data and imported into a Bayesian network with a defined conditional probability table. Using the probabilistic inference mechanism of the Bayesian network and the conditional probability table of each node, the posterior probability distribution corresponding to the warning vector is calculated. Finally, the joint probability distribution of the evolution of educational ethics risks is obtained. This joint probability distribution comprehensively reflects the probability of each core risk indicator in the educational scenario being in different state combinations during the multi-agent interaction process, and clearly presents the probabilistic laws of the collaborative evolution and mutual influence of each risk indicator.

[0095] This embodiment constructs a Bayesian network structure based on actual risk indicators in the educational ethics risk early warning vector. This ensures that the network structure is highly adapted to the actual ethical risks in multi-agent collaborative scenarios in the education field, accurately capturing the inherent dependencies between various risk indicators and avoiding the blindness and detachment from real-world scenarios in network modeling. Combining historical sample data of educational ethics risks, the network is trained using Bayesian estimation, providing solid data support for the conditional probability tables corresponding to network nodes. Compared to traditional parameter learning methods, this effectively utilizes historical experience data to mine probabilistic relationships between indicators, reducing errors caused by subjective settings and ensuring the objectivity, accuracy, and repeatability of the quantitative results of dependencies between risk indicators. The joint probability distribution of educational ethics risk evolution obtained through probabilistic inference of the Bayesian network breaks through the limitations of independent analysis of traditional single risk indicators, comprehensively and accurately capturing the collaborative evolution characteristics and probabilistic laws among multiple risk indicators, and clarifying the evolutionary trend under the interaction of various risk indicators. Meanwhile, the joint probability distribution output by this process provides scientific and quantitative core parameters for generating multiple risk evolution paths using the Monte Carlo method, improving the scientific rigor and reliability of the entire risk prediction method, and providing important probabilistic basis for the precise management of ethical risks in multi-agent collaborative systems in the education field.

[0096] In one embodiment, the Monte Carlo method is used to generate multiple risk evolution paths from the joint probability distribution. These risk evolution paths are then input into a machine learning classification model to determine the propagation path type of the current educational ethics risk evolution. This may include the following steps:

[0097] Step S401: Based on the joint probability distribution, the Monte Carlo method is used to conduct multiple rounds of random simulation to generate multiple time-series risk evolution paths; each risk evolution path records the activation state sequence of each educational ethics risk node at each moment.

[0098] Step S402: Extract risk propagation pattern features for each risk evolution path; the risk propagation pattern features include the risk outbreak time point, the set of educational subject nodes with the greatest impact, and the time sequence of risk propagation speed.

[0099] Step S403: Use a clustering algorithm based on Euclidean distance to perform cluster analysis on the propagation pattern characteristics of all paths to obtain several typical pattern categories of educational ethics risk propagation.

[0100] Step S404: Construct a machine learning classification model based on typical pattern categories as a risk propagation pattern recognition model; the input of the risk propagation pattern recognition model is the activation state vector of the current educational ethics risk node, and the output is the typical pattern category of risk propagation.

[0101] Step S405: After converting the real-time collected sequence of educational ethics risk events into risk node activation state vectors, input them into the risk propagation pattern recognition model for pattern matching to determine the propagation path type to which the current educational ethics risk evolution belongs.

[0102] Specifically, based on the obtained joint probability distribution of educational ethics risk evolution (including the state transition probability of risk nodes and random perturbation coefficients adapted to educational scenarios), a Monte Carlo method is used to conduct multiple rounds of stochastic simulation. The number of simulation rounds is preset according to the risk complexity of the educational scenario, and each round of simulation follows the probability law of the joint probability distribution to generate a time-series risk evolution path. Each risk evolution path is based on time and fully records the activation state sequence of each educational ethics risk node (corresponding to core indicators such as data privacy leakage, algorithmic bias, and damage to academic integrity) at each moment. The activation state is divided into two categories: active (risk occurred) and inactive (risk did not occur), ensuring the temporality and completeness of the path. Risk propagation mode features are extracted for each generated risk evolution path. These features specifically include: the risk outbreak time point (the timestamp of the first risk node in the path changing from inactive to active), the set of educational subject nodes with the greatest impact (the set of risk nodes associated with all affected educational subjects when the risk propagation range is the widest, such as privacy leakage nodes corresponding to student systems and algorithmic bias nodes corresponding to teacher platforms), and the risk propagation speed time-series sequence (the number of newly activated risk nodes per unit time, arranged in chronological order). A clustering algorithm based on Euclidean distance is employed to calculate the Euclidean distance between the features of all path propagation patterns. Paths with similar features are grouped together based on their distance. A pre-defined number of clusters is used, incorporating historical risk cases, to obtain several typical patterns of educational ethics risk propagation after cluster analysis. Each typical pattern corresponds to a risk evolution scenario with similar propagation characteristics. Using these typical pattern categories as training samples, a machine learning classification model (such as support vector machines or decision trees, adapted to small sample scenarios) is constructed as a risk propagation pattern recognition model. The input of this model is the activation state vector of each educational ethics risk node at the current moment (a quantified combination of the activation states of each node), and the output is the typical risk propagation pattern category corresponding to this vector. Educational ethics risk event sequences during multi-agent interaction are collected in real time. The sequences are split at pre-defined time intervals and converted into risk node activation state vectors at the corresponding moments, ensuring that the vector dimension is consistent with the model input dimension. This activation state vector is input into the trained risk propagation pattern recognition model for pattern matching. Through model classification reasoning, the propagation path type to which the current educational ethics risk evolution belongs is determined.

[0103] This embodiment employs multiple rounds of Monte Carlo simulation to comprehensively cover all possible risk evolution scenarios in educational settings, avoiding the limitations of single-path simulations and ensuring the comprehensiveness and representativeness of risk propagation pattern feature extraction. By accurately defining and extracting three core propagation pattern features, it clarifies the key time nodes, impact scope, and propagation rate of risk propagation, providing standardized, high-quality feature data for cluster analysis. The Euclidean distance-based clustering algorithm can accurately distinguish risk paths with different propagation characteristics. Combining historical cases to determine clustering parameters ensures that the obtained typical pattern categories closely match the actual risk evolution characteristics of educational scenarios, reducing errors caused by subjective classification. The constructed risk propagation pattern recognition model achieves rapid and accurate matching between the current risk activation state and typical propagation patterns, solving the problem that traditional methods cannot effectively identify risk propagation patterns and predict risk evolution trends.

[0104] In one embodiment, the educational ethics risk prediction parameters are obtained through the following steps, which may include:

[0105] Step S501: Obtain time series data corresponding to the propagation path type; the time series data contains risk propagation characteristics of multiple consecutive time stages.

[0106] Step S502: Input the time series data into the pre-trained long short-term memory network for time series modeling processing, and output the risk value of the education ethics stage corresponding to each time stage.

[0107] Furthermore, the input time-series data corresponds to the defined propagation path type. This data, structured around time, includes risk propagation characteristics across multiple consecutive time phases, specifically covering core quantitative information such as the risk outbreak time points, the set of educational stakeholders with the greatest impact, and the time-series data on risk propagation speed. Before input, this time-series data needs to be normalized to eliminate dimensional differences between different propagation characteristics, mapping all data to a unified value range to ensure modeling accuracy. The Long Short-Term Memory (LSTM) network used has been pre-trained based on historical time-series data of multi-agent ethical risks in the education field. During training, it has learned the time dependencies and evolutionary patterns of risk propagation characteristics, enabling it to accurately capture long-term correlation information in the time-series data. The preprocessed time series data is input into the pre-trained long short-term memory network. The network receives data through the input layer and captures the temporal correlation of risk propagation characteristics at each time stage through the hidden layer (such as the impact of the propagation speed of the previous stage on the risk level of the current stage). Then, the output layer outputs the educational ethics stage risk value corresponding to each time stage. This risk value is a quantitative value, and its value range is consistent with the stage risk evolution index ([0,10]). It is used to accurately characterize the severity of educational ethics risks within the corresponding time stage and provide standardized input data for the subsequent construction of a time-series risk value sequence.

[0108] Step S503: Connect the risk values ​​of each stage in chronological order to construct a time-series risk value sequence.

[0109] Step S504: Perform sliding window analysis on the time-series risk value sequence using a preset time window length, and calculate the fluctuation range of the risk value within each window.

[0110] Step S505: If the fluctuation amplitude corresponding to a consecutive preset positive integer number of time periods all exceeds the preset risk fluctuation threshold, then extract the statistical features of the current time-series risk value sequence.

[0111] Step S506: The statistical features are weighted and fused with the real-time interaction data of multiple agents in the field of education to generate updated educational ethics risk prediction parameters.

[0112] The real-time interaction process data includes interaction logs, feature extraction results, and risk node activation status data.

[0113] Specifically, time-series data corresponding to the propagation path type of the current educational ethics risk evolution is obtained. This time-series data, with time as the axis, includes risk propagation characteristics across multiple consecutive time stages, specifically covering core information such as the risk outbreak time points, the set of educational entities with the greatest impact, and the time-series data of risk propagation speed, consistent with the extracted propagation pattern characteristics. This time-series data is used as input and imported into a pre-trained Long Short-Term Memory (LSTM) network for temporal modeling. This LSTM network, pre-trained based on historical time-series data of multi-agent ethical risks in the education field, can accurately capture the temporal dependencies of risk propagation characteristics. Through modeling, it outputs the educational ethics stage risk values ​​corresponding to each time stage, quantifying the severity of the risk at each stage. The educational ethics stage risk values ​​corresponding to each time stage are concatenated sequentially to construct a temporally sequenced risk value sequence, which fully presents the evolution trajectory of the risk values ​​over time. Using a preset time window length (set according to the risk evolution cycle of the educational scenario), sliding window analysis is performed on the constructed temporally sequenced risk value sequence, calculating the fluctuation amplitude of the risk value window by window. The fluctuation amplitude is quantified by the extreme value difference or standard deviation of the risk values ​​within the window. A preset positive integer (calibrated according to the accuracy requirements of risk control) and a preset risk fluctuation threshold (determined based on historical risk out-of-control case data) are set. The fluctuation amplitude of each window is judged. If the fluctuation amplitude of each consecutive preset positive integer time period exceeds the preset risk fluctuation threshold, it indicates that the current risk evolution trend is unstable and the prediction parameters need to be updated. At this time, the statistical characteristics of the current time-series risk value sequence are extracted, specifically including key quantitative indicators such as the mean, variance, trend slope, and peak density of the sequence. At the same time, real-time interaction process data of multi-agents in the field of education is collected. This data includes interaction logs (raw interaction information), feature extraction results (interaction features extracted at each step, risk features, etc.), and risk node activation status data (activation / inactivation status of risk nodes at each time). The extracted statistical features and the real-time interaction process data are weighted and fused according to preset weights (the weights are calibrated according to the degree of influence of the data on risk prediction). Finally, updated educational ethics risk prediction parameters are generated for subsequent dynamic optimization of risk prediction.

[0114] This embodiment uses time-series data corresponding to propagation path types as the core input. Leveraging the time-series modeling capabilities of a pre-trained Long Short-Term Memory (LSTM) network, it accurately captures the temporal evolution patterns of educational ethics risk values, solving the problem of traditional methods' difficulty in quantifying risk temporal changes. The generated time-series risk value sequence can intuitively reflect the dynamic evolution of risks. Real-time monitoring of risk value fluctuations through sliding window analysis allows for timely identification of unstable risk evolution states, avoiding delayed warnings due to sudden risk changes and ensuring timely updates to prediction parameters. Statistical feature extraction accurately extracts the core patterns of risk fluctuations, providing a scientific basis for parameter updates. Weighted fusion of statistical features with real-time multi-agent interaction data achieves a deep integration of risk temporal patterns and real-time interaction data. Compared to single data input, this effectively improves the relevance and accuracy of updated prediction parameters and reduces prediction errors.

[0115] In one embodiment, such as Figure 2 As shown, this application also provides an ethical risk prediction system for multi-agent cooperative systems, which may include:

[0116] The log time-series analysis module 601 is used to acquire real-time data logs of the interaction behavior of multiple agents in the field of education, and to process the real-time data logs using time-series analysis methods to obtain the dynamic change sequence of the interaction process.

[0117] The stage risk early warning module 602 is used to divide the dynamic change sequence into time stage boundaries, extract the differentiated structure of time stages from the time stage boundaries, calculate the stage risk evolution index of each time stage, and generate an educational ethics risk early warning vector.

[0118] The risk path identification module 603 is used to model the dependency relationship between different risk indicators based on the educational ethics risk warning vector through a Bayesian network to obtain the joint probability distribution of the evolution of educational ethics risks; it is also used to generate multiple risk evolution paths from the joint probability distribution using the Monte Carlo method, and input the risk evolution paths into a machine learning classification model to determine the propagation path type of the current educational ethics risk evolution.

[0119] The time-series parameter generation module 604 is used to input the propagation path type into the pre-trained long short-term memory network to construct a time-series risk value sequence, extract the statistical features of the time-series risk value sequence and weight and fuse it with the real-time interaction process data of multiple agents in the field of education to generate educational ethics risk prediction parameters.

[0120] The aforementioned ethical risk prediction system for multi-agent collaborative systems comprises a log time-series analysis module, a stage risk warning module, a risk path identification module, and a time-series parameter generation module. These modules work collaboratively to form the core module of the system: The log time-series analysis module acquires real-time data logs of multi-agent interactions in the education field (covering key information such as interaction frequency, data transmission content, and operational command feedback). It uses time-series analysis methods to clean, structure, and extract time-dimensional interaction correlation features from the real-time data logs, outputting a dynamic change sequence of the interaction process. The stage risk warning module receives the dynamic change sequence output by the log time-series analysis module, divides it into time stage boundaries (starting with the interaction frequency exceeding a preset threshold), extracts interaction patterns, educational subject participation, and ethically sensitive word features from each time stage sequence, integrates them to form a differentiated structure, calculates the stage risk evolution index for each time stage using a preset formula, and finally generates an educational ethical risk warning vector. The risk path identification module receives the output from the stage risk warning module... The educational ethics risk early warning vector is used to model the conditional dependencies between different risk indicators through a Bayesian network, determine the conditional probability table of network nodes, and obtain the joint probability distribution of the evolution of educational ethics risks. Simultaneously, this module uses the Monte Carlo method to perform multiple rounds of random simulation on the joint probability distribution, generating multiple risk evolution paths. The propagation pattern features of each path are extracted, and typical pattern categories are obtained through cluster analysis. A machine learning classification model is constructed as a risk propagation pattern recognition model. The risk evolution path is input into this model to determine the current propagation path type of educational ethics risk evolution. The time-series parameter generation module receives the propagation path type output by the risk path recognition module, inputs the corresponding time-series data into a pre-trained long short-term memory network for time-series modeling, outputs the educational ethics stage risk value corresponding to each time stage, constructs a time-series risk value sequence, extracts the statistical features of this sequence, and then performs weighted fusion with the real-time interaction process data of multiple agents in the education field (including interaction logs, feature extraction results, and risk node activation state data) to finally generate educational ethics risk prediction parameters.

[0121] This embodiment features clearly defined modules, effectively addressing the problems of static review by single agents and difficulty in adapting to complex collaborative scenarios in existing technologies. The log time-series analysis module provides high-quality, time-series-based foundational data for subsequent processing, ensuring the objectivity of risk analysis. The staged risk warning module enables staged quantitative assessment of risks, providing standardized input for risk modeling. The risk path identification module accurately characterizes the dependencies and propagation patterns of risk indicators, overcoming the limitations of traditional methods in effectively tracking the time-series evolution path of risks. The time-series parameter generation module enables dynamic optimization of risk prediction parameters, improving the adaptability and accuracy of risk prediction. The four modules work synergistically to ensure the timeliness, accuracy, and operability of educational ethics risk prediction, providing educational administrators with dynamic control measures. This effectively avoids ethical risks such as data privacy leaks and algorithmic biases that could infringe upon educational fairness and students' legitimate rights, providing technical support for the standardized and safe development of the field of artificial intelligence education.

[0122] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0123] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the previously described method for predicting ethical risks in multi-agent cooperative systems.

[0124] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0125] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0126] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for predicting ethical risks in multi-agent cooperative systems, characterized in that, The method includes: Real-time data logs of multi-agent interaction behavior in the field of education are obtained, and time series analysis is used to process the real-time data logs to obtain the dynamic change sequence of the interaction process. The dynamic change sequence is divided into time stage boundaries, the differentiated structure of the time stages is extracted from the time stage boundaries, and the stage risk evolution index of each time stage is calculated to generate an educational ethics risk early warning vector. Based on the aforementioned educational ethics risk warning vector, the dependency relationship between different risk indicators is modeled using a Bayesian network to obtain the joint probability distribution of the evolution of educational ethics risks. The Monte Carlo method is used to generate multiple risk evolution paths for the joint probability distribution, and the risk evolution paths are input into a machine learning classification model to determine the propagation path type of the current educational ethics risk evolution. The propagation path type is input into a pre-trained long short-term memory network to construct a temporal risk value sequence. The statistical features of the temporal risk value sequence are extracted and weighted and fused with real-time interaction data of multiple agents in the field of education to generate educational ethics risk prediction parameters.

2. The method according to claim 1, characterized in that, The process of dividing the dynamic change sequence into time stage boundaries, extracting the differentiated structure of the time stages from the time stage boundaries, calculating the stage risk evolution index for each time stage, and generating an educational ethics risk early warning vector includes: Acquire interaction frequency data at each time point in the dynamic change sequence of multi-agent interactions in the field of education; The interaction frequency data is compared with a preset threshold one by one. If the interaction frequency data at the current time point exceeds the preset threshold, the time point is defined as the starting boundary of the new stage. Using the determined starting boundaries of each stage, the dynamic change sequence is segmented to obtain multiple continuous time stage sequences; Interaction pattern features, educational subject participation features, and ethically sensitive word features are extracted for each time stage sequence and integrated to form a differentiated structure corresponding to each time stage; From the differentiated structures at each time stage, we extract the characteristics of risk transmission intensity, risk accumulation trend, frequency of ethical conflicts between subjects, and delay in intervention response. Based on the risk propagation intensity characteristics and the risk accumulation trend characteristics, the stage risk evolution index for each time period is calculated using a preset algorithm. The stage risk evolution index of each time stage is weighted and fused with the corresponding inter-subject ethical conflict frequency characteristics to generate an educational ethics risk early warning vector for each time stage.

3. The method according to claim 2, characterized in that, The stage risk evolution index is calculated using the following formula: in, This represents the stage risk evolution index, with a value range of [value range missing]. , This represents the quantitative value of the risk transmission intensity characteristic, with a value range of [0, 10]. This represents a quantitative value indicating the cumulative trend of risk, with a value range of [value range missing]. , The weighting coefficient representing the intensity of risk transmission has a range of values. , The coefficient representing the synergistic adjustment of risk propagation and accumulation, with a value range of [value missing]. .

4. The method according to claim 1, characterized in that, The method of modeling the dependencies between different risk indicators using a Bayesian network based on the educational ethics risk early warning vector to obtain the joint probability distribution of the evolution of educational ethics risks includes: Based on the risk indicators in the aforementioned educational ethics risk warning vector, a Bayesian network structure is constructed. The Bayesian network structure represents the conditional dependencies between different risk indicators; The Bayesian network structure is trained using a Bayesian estimation method based on historical sample data of educational ethics risks to determine the conditional probability table corresponding to each node in the network. The conditional probability table quantifies the probabilistic relationships of conditional dependencies between nodes; The educational ethics risk warning vector is input into a Bayesian network with a defined conditional probability table. By calculating the posterior probability distribution corresponding to the warning vector, the joint probability distribution of the evolution of educational ethics risk is obtained. The joint probability distribution reflects the probabilistic laws governing the coordinated evolution of various risk indicators.

5. The method according to claim 1, characterized in that, The Monte Carlo method is used to generate multiple risk evolution paths from the joint probability distribution. These risk evolution paths are then input into a machine learning classification model to determine the propagation path type of the current educational ethics risk evolution, including: Based on the joint probability distribution, the Monte Carlo method is used to perform multiple rounds of random simulation to generate multiple time-series risk evolution paths; each risk evolution path records the activation state sequence of each educational ethics risk node at each moment. For each of the aforementioned risk evolution paths, risk propagation pattern features are extracted; the risk propagation pattern features include the risk outbreak time point, the set of educational subject nodes with the greatest impact, and the time sequence of risk propagation speed. A clustering algorithm based on Euclidean distance was used to perform cluster analysis on the propagation pattern characteristics of all paths, resulting in several typical pattern categories of educational ethics risk propagation. A machine learning classification model is constructed based on the typical pattern categories as a risk propagation pattern recognition model; the input of the risk propagation pattern recognition model is the activation state vector of the current educational ethics risk node, and the output is the typical pattern category of risk propagation. After converting the real-time collected sequence of educational ethics risk events into risk node activation state vectors, the data is input into the risk propagation pattern recognition model for pattern matching to determine the propagation path type to which the current educational ethics risk evolution belongs.

6. The method according to claim 1, characterized in that, The educational ethics risk prediction parameters are obtained through the following steps: Obtain the time series data corresponding to the propagation path type; the time series data contains risk propagation characteristics of multiple consecutive time stages; The time series data is input into a pre-trained long short-term memory network for time series modeling, and the corresponding educational ethics stage risk value is output for each time stage. The risk values ​​of each stage are linked together in chronological order to construct a time-series risk value sequence. A sliding window analysis is performed on the time-series risk value sequence using a preset time window length to calculate the fluctuation range of the risk value within each window. If the fluctuation amplitude corresponding to a consecutive preset positive integer number of time periods all exceeds the preset risk fluctuation threshold, then the statistical characteristics of the current time-series risk value sequence are extracted. The statistical features are weighted and fused with real-time interaction data of multiple agents in the field of education to generate updated educational ethics risk prediction parameters. The real-time interaction process data includes interaction logs, feature extraction results, and risk node activation status data.

7. An ethical risk prediction system for multi-agent collaborative systems, characterized in that, The system includes: The log time-series analysis module is used to acquire real-time data logs of the interaction behavior of multiple agents in the field of education, and to process the real-time data logs using time-series analysis methods to obtain the dynamic change sequence of the interaction process. The stage risk warning module is used to divide the dynamic change sequence into time stage boundaries, extract the differentiated structure of the time stages from the time stage boundaries, calculate the stage risk evolution index of each time stage, and generate an educational ethics risk warning vector. The risk path identification module is used to model the dependency relationship between different risk indicators based on the educational ethics risk warning vector through a Bayesian network to obtain the joint probability distribution of the evolution of educational ethics risks; it is also used to generate multiple risk evolution paths from the joint probability distribution using the Monte Carlo method, and input the risk evolution paths into a machine learning classification model to determine the propagation path type of the current educational ethics risk evolution. The time-series parameter generation module is used to input the propagation path type into a pre-trained long short-term memory network to construct a time-series risk value sequence, extract the statistical features of the time-series risk value sequence and weight and fuse it with the real-time interaction process data of multiple agents in the field of education to generate educational ethics risk prediction parameters.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images