CN-122022498-A - Multidimensional behavior modeling and risk quantification method and system for enterprise AI (advanced technology attachment) agent

Abstract

The application provides a multidimensional behavior modeling and risk quantifying method and system for an enterprise AI intelligent agent, which relate to the technical field of artificial intelligence, and are characterized in that an intelligent agent behavior data stream which is generated by the operation of the enterprise AI intelligent agent and contains a behavior event record and a time sequence association relation thereof is firstly collected and then is input into a multidimensional behavior decoupling network to extract intention-oriented and environment-responsive behavior characteristics, and then constructing a dynamic interaction mapping relation between the two, determining driving weight and feedback correction parameters, generating a behavior evolution path feature sequence according to iterative fusion of the parameters, and finally carrying out behavior pattern clustering on the behavior evolution path feature sequence, and matching with a preset risk behavior feature library to generate a behavior risk quantification evaluation result. The method can scientifically and accurately evaluate the behavior risk of the AI intelligent agent of the enterprise and ensure the safe and stable operation of the AI intelligent agent.

Inventors

• PU LIANG

Assignees

• 云纷(上海)信息科技有限公司

Dates

Publication Date

20260512

Application Date

20260413

Claims (10)

1. 1. The method for multidimensional behavior modeling and risk quantification for the enterprise AI intelligent agent is characterized by comprising the following steps: collecting an agent behavior data stream generated by an enterprise AI agent in the operation process, wherein the agent behavior data stream comprises behavior event records generated when the agent interacts with an external environment and time sequence association relations among the behavior event records; Inputting the agent behavior data stream into a pre-built multidimensional behavior decoupling network for behavior feature separation processing, extracting the agent intention guiding type behavior feature from the behavior event record through an intention decoder in the multidimensional behavior decoupling network, and extracting the agent environment response type behavior feature from the behavior event record through an environment response decoder in the multidimensional behavior decoupling network; Constructing a dynamic interaction mapping relation between the intention-oriented behavior feature and the environment-responsive behavior feature, wherein the dynamic interaction mapping relation comprises a driving weight parameter of the intention-oriented behavior feature to the environment-responsive behavior feature and a feedback correction parameter of the environment-responsive behavior feature to the intention-oriented behavior feature; Performing iterative fusion processing on the intention-oriented behavior feature and the environment-responsive behavior feature according to the driving weight parameter and the feedback correction parameter to generate a behavior evolution path feature sequence reflecting the behavior evolution trend of the intelligent agent; And performing behavior pattern clustering processing on the behavior evolution path feature sequence, determining a behavior pattern type label of the agent according to a clustering result, and performing matching processing on the behavior pattern type label and a preset risk behavior feature library to generate an agent behavior risk quantification evaluation result.
2. 2. The method for multidimensional behavior modeling and risk quantification for an enterprise AI agent according to claim 1, wherein the collecting agent behavior data streams generated by the enterprise AI agent during operation comprises: Monitoring interactive event nodes generated in the process of carrying out data exchange with an external data source when an enterprise AI intelligent agent executes an enterprise business process task, and recording each interactive event node as a basic behavior unit, wherein the basic behavior unit comprises identification information of an interaction initiator, address identification information of an interaction receiver, type identification code of interaction operation, time stamp information of the interaction operation and data load feature description carried by the interaction operation; Extracting data load feature description carried in the basic behavior unit, and carrying out semantic analysis processing on the data load feature description to obtain a business intention type label corresponding to the basic behavior unit and a confidence score of the business intention type label; According to the time stamp information of the interactive operation in the basic behavior unit, constructing a sequence relation of the basic behavior unit in a time dimension, connecting the basic behavior units with direct front-to-back relation through directed edges, and forming a time sequence association relation between behavior event records, wherein the time sequence association relation comprises a time interval parameter between a precursor behavior unit identifier and a back-to-back behavior unit identifier and a type transition probability between the interactive operation type of the precursor behavior unit identifier and the interactive operation type of the back-to-back behavior unit identifier; Organizing the basic behavior units and the time sequence association relation into an agent behavior data stream according to the ascending order of time stamp information, wherein each basic behavior unit in the agent behavior data stream carries a forward reference pointer pointing to a precursor basic behavior unit and a backward reference pointer pointing to a subsequent basic behavior unit; The basic behavior units with the same interaction initiator identity information in the intelligent agent behavior data stream are subjected to aggregation processing to form an individual behavior track sequence taking a single enterprise AI intelligent agent as a dimension, wherein the individual behavior track sequence comprises all basic behavior units generated by the enterprise AI intelligent agent at different time points and a complete time sequence association relation network between the basic behavior units; The basic behavior units with the same business intention type labels in the individual behavior track sequence are subjected to sectional aggregation according to the timestamp information to generate a sustainable behavior fragment set reflecting the driving of the enterprise AI intelligent agent under the specific business intention, wherein the sustainable behavior fragment set comprises a behavior fragment starting timestamp, a behavior fragment ending timestamp and the arrangement sequence of the basic behavior units in the behavior fragments; Extracting the type identification code of the interactive operation of the first basic behavior unit of each continuous behavior fragment in the continuous behavior fragment set as a fragment entry behavior type identification, and extracting the type identification code of the interactive operation of the last basic behavior unit of each continuous behavior fragment in the continuous behavior fragment set as a fragment exit behavior type identification; Constructing a boundary behavior type conversion path of the persistence behavior segment according to the segment inlet behavior type identifier and the segment outlet behavior type identifier, wherein the boundary behavior type conversion path comprises an intermediate behavior type sequence from an inlet behavior type to an outlet behavior type and a transfer frequency statistic value between adjacent behavior types; Classifying the sustainable behavior fragments with the same boundary behavior type conversion paths in the sustainable behavior fragment set to generate a similar behavior fragment cluster taking the boundary behavior type conversion paths as a clustering center, wherein the similar behavior fragment cluster comprises all the sustainable behavior fragments belonging to the same boundary behavior type conversion paths and the occurrence frequency weight of each sustainable behavior fragment in the similar behavior fragment cluster; Calculating duration deviation parameters of the duration of each duration of the similar behavior fragment cluster and the average duration of all the duration of the similar behavior fragments, marking the duration of the duration deviation parameter exceeding a preset tolerance threshold as an abnormal behavior fragment, and marking a boundary behavior type conversion path corresponding to the abnormal behavior fragment as an abnormal behavior path identifier.
3. 3. The method for multidimensional behavior modeling and risk quantification of an enterprise AI agent according to claim 1, wherein the inputting the agent behavior data stream into a pre-built multidimensional behavior decoupling network performs behavior feature separation processing, extracting, by an intention decoder in the multidimensional behavior decoupling network, an intention-directed behavior feature of the agent from the behavior event record, and extracting, by an environmental response decoder in the multidimensional behavior decoupling network, an environmental response type behavior feature of the agent from the behavior event record, includes: Inputting a basic behavior unit in the intelligent agent behavior data stream into a shared feature coding module of the multidimensional behavior decoupling network, wherein the shared feature coding module performs embedding vectorization processing on identity identification information of an interaction initiator of the basic behavior unit, address identification information of an interaction receiver, a type identification code of interaction operation and data load feature description to generate an original behavior embedded vector of the basic behavior unit; The original behavior embedded vector is input into an intention decoder and an environment response decoder in parallel, the intention decoder carries out semantic intention analysis processing on the original behavior embedded vector through a self-attention mechanism, behavior driving factor characteristics which are directly related to business target achievement are separated from the original behavior embedded vector and serve as intention-oriented behavior characteristics, and the intention-oriented behavior characteristics comprise business target directivity parameters and business target persistence parameters; The environment response decoder carries out context dependency modeling processing on the original behavior embedded vector through a conditional random field model, and separates behavior adaptation factor characteristics associated with external environment state change from the original behavior embedded vector to serve as environment response type behavior characteristics, wherein the environment response type behavior characteristics comprise environment state sensitivity parameters and behavior feedback timeliness parameters; Inputting the intention-oriented behavior feature output by the intention decoder and the environment-responsive behavior feature output by the environment-responsive decoder into a feature orthogonalization constraint layer of the multidimensional behavior decoupling network, calculating the mutual information quantity between the intention-oriented behavior feature and the environment-responsive behavior feature through the feature orthogonalization constraint layer, and carrying out joint optimization adjustment on network weight parameters of the intention decoder and the environment-responsive decoder based on a minimization target of the mutual information quantity; Carrying out residual connection processing on the intention guiding type behavior feature processed by the feature orthogonalization constraint layer and the original behavior embedded vector to generate an enhanced intention guiding type behavior feature with original behavior information compensation, and carrying out residual connection processing on the environment response type behavior feature processed by the feature orthogonalization constraint layer and the original behavior embedded vector to generate an enhanced environment response type behavior feature with original behavior information compensation; Inputting the enhanced intention-oriented behavior feature into a first time sequence convolution module of the multidimensional behavior decoupling network, performing sliding window convolution operation on the enhanced intention-oriented behavior feature in a time dimension through the first time sequence convolution module, and extracting a change trend feature of the enhanced intention-oriented behavior feature in a continuous time window as an intention-oriented behavior evolution track; Inputting the enhanced environmental response type behavior feature into a second time sequence convolution module of the multidimensional behavior decoupling network, performing sliding window convolution operation on the enhanced environmental response type behavior feature in a time dimension through the second time sequence convolution module, and extracting a change trend feature of the enhanced environmental response type behavior feature in a continuous time window as an environmental response type behavior evolution track; Inputting the intention-oriented behavior evolution track and the environment-responsive behavior evolution track into a track alignment module of the multidimensional behavior decoupling network, calculating phase offset of the intention-oriented behavior evolution track and the environment-responsive behavior evolution track in a time dimension through the track alignment module, and performing time axis translation calibration processing on the intention-oriented behavior evolution track according to the phase offset to generate the environment-responsive behavior evolution track aligned with the time of the intention-oriented behavior evolution track; Performing feature splicing processing on the time-aligned environment response type behavior evolution track and the intention-oriented type behavior evolution track point by point to generate a multidimensional coupling behavior feature sequence with time synchronization relevance, wherein feature vectors corresponding to each time point in the multidimensional coupling behavior feature sequence simultaneously comprise intention-oriented type behavior feature information and environment response type behavior feature information of the time point.
4. 4. The method for multidimensional behavioral modeling and risk quantification for an enterprise AI agent of claim 1, wherein the constructing a dynamic interactive mapping relationship between the intent-directed behavioral characteristics and the environmental responsive behavioral characteristics comprises: The intention-oriented behavior feature vector of the adjacent time point is used as a forward driving input, the environment-responsive behavior feature vector of the adjacent time point is used as a forward driving output, and a forward driving mapping function of the intention to respond to the environment is constructed; Calculating a linear transformation matrix required by mapping the intention-oriented behavior feature vector to the environment-responsive behavior feature vector in each adjacent time point by the forward driving mapping function, extracting off-diagonal elements in the linear transformation matrix, and generating driving weight parameters of the intention-oriented behavior feature to the environment-responsive behavior feature, wherein the driving weight parameters are used for representing driving action intensity of each dimension component in the intention-oriented behavior feature to each dimension component in the environment-responsive behavior feature; The environment response type behavior feature vector of the adjacent time point is used as a reverse correction input, the intention guide type behavior feature vector of the adjacent time point is used as a reverse correction output, and a reverse correction mapping function of the environment response type behavior feature vector is constructed; Calculating a linear transformation matrix required by mapping the environment response type behavior feature vector to the intention-oriented behavior feature vector in each adjacent time point through the reverse correction mapping function, extracting off-diagonal elements in the linear transformation matrix, and generating feedback correction parameters of the environment response type behavior feature to the intention-oriented behavior feature, wherein the feedback correction parameters are used for representing correction action intensity of each dimension component in the environment response type behavior feature to each dimension component in the intention-oriented behavior feature; organizing the driving weight parameters into driving weight parameter time sequence sequences according to the sequence of time dimensions, performing differential operation on the driving weight parameter time sequence sequences, and extracting the variation amplitude values of the driving weight parameters between adjacent time points as driving weight dynamic fluctuation characteristics; organizing the feedback correction parameters into a feedback correction parameter time sequence according to the sequence of the time dimension, performing differential operation processing on the feedback correction parameter time sequence, and extracting the variation amplitude value of the feedback correction parameters between adjacent time points as a feedback correction dynamic fluctuation characteristic; Inputting the driving weight dynamic fluctuation feature and the feedback correction dynamic fluctuation feature into an interactive coupling analysis module, calculating a covariance value of the driving weight dynamic fluctuation feature and the feedback correction dynamic fluctuation feature at the same time point through the interactive coupling analysis module, and taking the covariance value as an interactive coupling strength index between the intention-oriented behavior feature and the environment response type behavior feature; Screening the driving weight parameters and the feedback correction parameters according to the interactive coupling strength indexes, extracting driving weight parameters at corresponding time points of the interactive coupling strength indexes exceeding a preset coupling threshold as effective driving weight parameters, and extracting feedback correction parameters at corresponding time points of the interactive coupling strength indexes exceeding the preset coupling threshold as effective feedback correction parameters; and organizing the effective driving weight parameters into effective driving weight sequences according to the sequence of the corresponding time points, organizing the effective feedback correction parameters into effective feedback correction sequences according to the sequence of the corresponding time points, and constructing a corresponding relation mapping table between the effective driving weight sequences and the effective feedback correction sequences, wherein each entry in the corresponding relation mapping table contains the effective driving weight parameters and the effective feedback correction parameters of the same time point.
5. 5. The method for multidimensional behavior modeling and risk quantification for an enterprise AI agent according to claim 4, wherein the performing iterative fusion processing on the intent-oriented behavior feature and the environmental response type behavior feature according to the driving weight parameter and the feedback correction parameter to generate a behavior evolution path feature sequence reflecting a behavior evolution trend of the agent comprises: Organizing the intention-oriented behavior features into an intention feature time sequence matrix according to the sequence of the time dimension, organizing the environment-responsive behavior features into an environment feature time sequence matrix according to the sequence of the time dimension, organizing the effective driving weight sequence into a driving weight time sequence matrix according to the sequence of the time dimension, and organizing the effective feedback correction sequence into a feedback correction time sequence matrix according to the sequence of the time dimension; Performing feature stitching operation on the driving weight time sequence matrix and the intention feature time sequence matrix to generate a combination feature matrix of intention features and driving weights, and performing feature stitching operation on the feedback correction time sequence matrix and the environment feature time sequence matrix to generate a combination feature matrix of environment features and feedback correction; Performing feature stitching operation on the combination feature matrix of the intention feature and the driving weight and the combination feature matrix of the environment feature and the feedback correction to generate a first wheel combination feature matrix, wherein the feature vector of each time point in the first wheel combination feature matrix simultaneously comprises an intention feature component weighted by the driving weight and an environment feature component weighted by the feedback correction; Inputting the first wheel combination feature matrix into a feature evolution prediction module, performing time sequence prediction processing on the first wheel combination feature matrix through the feature evolution prediction module to generate a prediction feature vector of the first wheel combination feature matrix at the next time point, and performing splicing processing on the prediction feature vector and the first wheel combination feature matrix to generate an expanded first wheel combination feature matrix; splicing the expanded first wheel combination characteristic matrix, the driving weight time sequence matrix and the feedback correction time sequence matrix to generate a second wheel combination characteristic matrix; Repeatedly executing an iterative process of inputting the previous wheel combination feature matrix into a feature evolution prediction module to generate an expansion feature, and then splicing the expansion feature, the driving weight time sequence matrix and the feedback correction time sequence matrix until the preset iteration round number is reached, so as to obtain a final round combination feature matrix; extracting the space position coordinates of the feature vector of each time point in the final round combination feature matrix in a feature space, connecting the space position coordinates of each time point into space position coordinate track lines according to the sequence of time dimensions, and taking the space position coordinate track lines as behavior evolution path feature sequences reflecting the behavior evolution trend of the intelligent agent; Calculating Euclidean distance values between space position coordinates of adjacent time points in the behavior evolution path feature sequence, organizing the Euclidean distance values into a behavior evolution step sequence according to the sequence of time dimensions, calculating the ratio between adjacent step values in the behavior evolution step sequence, and adding the ratio into the behavior evolution path feature sequence as a behavior evolution speed change rate sequence.
6. 6. The method for multidimensional behavior modeling and risk quantification of enterprise AI-oriented agents according to claim 1, wherein the performing behavior pattern clustering processing on the behavior evolution path feature sequence, determining a behavior pattern class label of an agent according to a clustering result, performing matching processing on the behavior pattern class label and a preset risk behavior feature library, and generating an agent behavior risk quantification evaluation result comprises: inputting the behavior evolution path feature sequence into a behavior pattern clustering network, and performing dimension reduction processing on the behavior evolution path feature sequence by the behavior pattern clustering network through a self-encoder structure to generate a hidden layer representation vector sequence of the behavior evolution path feature sequence in a low-dimensional hidden space; inputting the hidden layer representation vector sequence into a clustering distribution layer of the behavior pattern clustering network, calculating membership probability distribution of each hidden layer representation vector to each preset clustering center by the clustering distribution layer through a soft distribution mechanism, and taking a clustering center number corresponding to the maximum probability value in the membership probability distribution as a behavior pattern category label corresponding to the hidden layer representation vector; Performing segmentation labeling processing on the behavior evolution path feature sequence according to the behavior pattern class labels, and aggregating behavior evolution path feature sequence fragments corresponding to continuous time points with the same behavior pattern class labels into single behavior pattern fragments to generate a behavior pattern fragment set, wherein the behavior pattern fragment set comprises a starting time point position, an ending time point position and a behavior pattern class label of each behavior pattern fragment; extracting the space geometrical shape characteristics of the behavior evolution path characteristic sequence segments of each behavior pattern segment in the behavior pattern segment set, wherein the space geometrical shape characteristics comprise the number of curvature change extreme points of the behavior evolution path characteristic sequence segments in a characteristic space and the flexible rate change accumulation quantity of the behavior evolution path characteristic sequence segments in the characteristic space; The space geometrical morphology features and the behavior pattern class labels are stored in an associated mode to generate a behavior pattern labeling fragment set with space geometrical morphology feature labels, and behavior pattern fragments with the same behavior pattern class labels in the behavior pattern labeling fragment set are arranged according to a time sequence to form a behavior pattern time sequence evolution chain under a single behavior pattern class; Inputting the behavior pattern time-sequence evolution chain into a risk behavior feature library for matching treatment, wherein the risk behavior feature library comprises a space geometric feature template of a preset risk behavior pattern and a time-sequence evolution template of the preset risk behavior pattern, and calculating the morphological similarity between the space geometric feature of the behavior pattern time-sequence evolution chain and the space geometric feature template of the preset risk behavior pattern; calculating a sequence editing distance between a transfer sequence of a behavior pattern class label in a behavior pattern time sequence evolution chain and a transfer sequence of a behavior pattern class label in a time sequence evolution template of a preset risk behavior pattern, and normalizing the sequence editing distance to obtain a time sequence evolution matching degree; Calculating a comprehensive risk matching score according to the morphological similarity and the time sequence evolution matching degree, marking a corresponding behavior mode time sequence evolution chain with the comprehensive risk matching score exceeding a preset risk threshold as a risk behavior evolution chain, and outputting a behavior mode class label sequence corresponding to the risk behavior evolution chain as an agent behavior risk quantitative evaluation result.
7. 7. The method for multidimensional behavior modeling and risk quantification of an enterprise AI agent according to claim 1, wherein the inputting the agent behavior data stream into a pre-built multidimensional behavior decoupling network performs behavior feature separation processing, the intent-directed behavior feature of the agent is extracted from the behavior event record by an intent decoder in the multidimensional behavior decoupling network, and the environmental-responsive behavior feature of the agent is extracted from the behavior event record by an environmental-response decoder in the multidimensional behavior decoupling network, further comprising: Acquiring data load characteristic description carried by interactive operation of a basic behavior unit in the agent behavior data stream, performing content type identification processing on the data load characteristic description, and dividing the data load characteristic description into a structured data load part and an unstructured data load part; Inputting the structured data load part into a first feature extraction branch of the intention decoder, wherein the first feature extraction branch carries out numerical feature mapping processing on the structured data load part through a multi-layer perceptron network to generate a structured intention feature component; Inputting the unstructured data load part into a second feature extraction branch of the intention decoder, wherein the second feature extraction branch performs semantic feature extraction processing on the unstructured data load part through a convolutional neural network to generate unstructured intention feature components; performing feature dimension alignment processing on the structured intention feature component and the unstructured intention feature component, performing feature stitching processing on the aligned structured intention feature component and unstructured intention feature component, and generating a fusion type intention guiding type behavior feature; inputting the structured data load part into a third feature extraction branch of the environment response decoder, wherein the third feature extraction branch carries out time sequence modeling processing on a numerical variation mode of the structured data load part through a cyclic neural network to generate a structured environment response feature component; Inputting the unstructured data load part into a fourth feature extraction branch of the environment response decoder, and performing context correlation analysis processing on semantic content of the unstructured data load part through an attention mechanism by the fourth feature extraction branch to generate unstructured environment response feature components; And carrying out characteristic dimension alignment processing on the structured environment response characteristic component and the unstructured environment response characteristic component, and carrying out characteristic splicing processing on the aligned structured environment response characteristic component and the unstructured environment response characteristic component to generate a fusion type environment response type behavior characteristic.
8. 8. The method for multidimensional behavioral modeling and risk quantification for an enterprise AI agent of claim 1, wherein the constructing a dynamic interactive mapping relationship between the intent-directed behavioral characteristics and the environmental responsive behavioral characteristics further comprises: Acquiring the type identification codes of the interactive operation of the basic behavior units in the behavior data stream of the intelligent agent, organizing the type identification codes of the interactive operation into a behavior type sequence according to the sequence of time dimension, performing frequent pattern mining processing on the behavior type sequence, and extracting continuous behavior type subsequences frequently appearing in the behavior type sequence as behavior type transfer patterns; Constructing a behavior type transfer graph structure according to the behavior type transfer mode, wherein the behavior type transfer graph structure comprises graph nodes with interactive type identification codes and graph structure representations with the behavior type transfer mode as directed edges, and each directed edge carries the occurrence frequency weight of the behavior type transfer mode; Mapping the intention-oriented behavior feature to a graph node of the behavior type transition graph structure, generating an intention labeling behavior type transition graph with an intention-oriented behavior feature label, mapping the environment-responsive behavior feature to a graph node of the behavior type transition graph structure, and generating an environment labeling behavior type transition graph with an environment-responsive behavior feature label; Calculating the difference degree of the intention guiding behavior characteristics between adjacent graph nodes in the intention labeling behavior type transfer graph, taking the difference degree as a change amplitude parameter of the intention guiding behavior characteristics in the behavior type transfer process, and adding the change amplitude parameter to a corresponding directed edge to serve as an intention change edge weight; Calculating the difference degree of the environmental response type behavior characteristics between adjacent graph nodes in the environmental labeling behavior type transfer graph, taking the difference degree as a change amplitude parameter of the environmental response type behavior characteristics in the behavior type transfer process, and adding the change amplitude parameter to a corresponding directed edge to serve as an environmental change edge weight; Calculating the ratio of the intent change edge weight to the environment change edge weight corresponding to the oriented edges, generating an intent-to-environment driving strength ratio corresponding to each oriented edge, and taking the intent-to-environment driving strength ratio as a driving weight parameter of an intent-oriented behavior feature to an environment response behavior feature; And calculating the ratio of the corresponding directed edge of the environment change edge weight to the intent change edge weight, generating the environment to intent feedback intensity ratio corresponding to each directed edge, and taking the environment to intent feedback intensity ratio as a feedback correction parameter of the environment response type behavior characteristic to the intent guide type behavior characteristic.
9. 9. The method for multidimensional behavior modeling and risk quantification for an enterprise AI agent according to claim 8, wherein the performing iterative fusion processing on the intent-oriented behavior feature and the environmental response type behavior feature according to the driving weight parameter and the feedback correction parameter to generate a behavior evolution path feature sequence reflecting a behavior evolution trend of the agent further comprises: The intention change edge weight and the environment change edge weight of each directed edge in the behavior type transfer graph structure are obtained, the intention change edge weight and the environment change edge weight of the directed edge are combined into a binary group form, and an interaction characteristic binary group set of each directed edge is formed; Performing weighted average processing on the two groups in the interaction characteristic two-group set according to the occurrence frequency weight of the corresponding directed edge to generate an average interaction characteristic two-group of each directed edge, taking the intention change edge weight component in the average interaction characteristic two-group as the reference driving weight of the directed edge and taking the environment change edge weight component in the average interaction characteristic two-group as the reference feedback weight of the directed edge; Mapping basic behavior units in the intelligent agent behavior data stream onto graph nodes of a behavior type transition graph structure according to the type identification codes of the interactive operation, and taking original behavior embedded vectors corresponding to the basic behavior units as node feature vectors of the graph nodes; taking the reference driving weight as a propagation attenuation coefficient on the directed edge, taking the reference feedback weight as a counter propagation attenuation coefficient on the directed edge, executing node characteristic iterative propagation operation of the graph neural network on a behavior type transfer graph structure, and performing diffusion fusion on node characteristic vectors along the directed edge according to the propagation attenuation coefficient through repeated iterative propagation; Extracting updated node feature vectors of each graph node after the graph neural network iterative propagation operation, and sequencing the updated node feature vectors according to the time stamp information of the corresponding basic behavior units to generate an updated node feature vector time sequence; Inputting the updated node feature vector time sequence into a time sequence feature extraction network, performing long-term and short-term dependency modeling processing on the updated node feature vector time sequence by the time sequence feature extraction network, and extracting a time sequence evolution trend component of the updated node feature vector time sequence as a behavior evolution path feature sequence.
10. 10. The system for multi-dimensional behavior modeling and risk quantification for an enterprise AI agent is characterized by comprising a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores machine-executable instructions, and the machine-executable instructions, when executed by the processor, implement the method for multi-dimensional behavior modeling and risk quantification for an enterprise AI agent according to any one of claims 1-9.

Description

Multidimensional behavior modeling and risk quantification method and system for enterprise AI (advanced technology attachment) agent Technical Field The application relates to the technical field of artificial intelligence, in particular to a multidimensional behavior modeling and risk quantification method and system for an enterprise AI intelligent agent. Background In the process of intelligent development of enterprises, the AI intelligent agents of the enterprises are widely applied to various business scenes, and the complexity and diversity of the behaviors are increasingly prominent. The behavior of the AI agent is driven not only by its own internal logic and algorithms, but also by frequent and complex interactions with the external environment. Accurately understanding and evaluating the behavior of the enterprise AI intelligent agent is important to guaranteeing the operation safety of the enterprise, optimizing the business process and improving the decision quality. However, there are a number of disadvantages to the current research on the behavior of enterprise AI agents. On one hand, the conventional method is often only concerned with the single behavior of the AI agent, and is difficult to comprehensively capture multidimensional features of the behavior, for example, only analyze the efficiency of executing tasks, and neglect the adaptability and other features when the AI agent interacts with the external environment. On the other hand, the risk assessment of the AI intelligent agent behavior lacks a scientific and systematic quantification method, relies on subjective judgment or simple rule matching, cannot accurately reflect the internal relation between the behavior and the risk, and is difficult to meet the requirement of enterprises on accurate management and control of the AI intelligent agent behavior in complex and changeable environments. Disclosure of Invention In view of the above, the present application aims to provide a multidimensional behavior modeling and risk quantification method and system for an enterprise AI intelligent agent. According to a first aspect of the present application, there is provided a multidimensional behavior modeling and risk quantification method for an enterprise AI agent, the method comprising: collecting an agent behavior data stream generated by an enterprise AI agent in the operation process, wherein the agent behavior data stream comprises behavior event records generated when the agent interacts with an external environment and time sequence association relations among the behavior event records; Inputting the agent behavior data stream into a pre-built multidimensional behavior decoupling network for behavior feature separation processing, extracting the agent intention guiding type behavior feature from the behavior event record through an intention decoder in the multidimensional behavior decoupling network, and extracting the agent environment response type behavior feature from the behavior event record through an environment response decoder in the multidimensional behavior decoupling network; Constructing a dynamic interaction mapping relation between the intention-oriented behavior feature and the environment-responsive behavior feature, wherein the dynamic interaction mapping relation comprises a driving weight parameter of the intention-oriented behavior feature to the environment-responsive behavior feature and a feedback correction parameter of the environment-responsive behavior feature to the intention-oriented behavior feature; Performing iterative fusion processing on the intention-oriented behavior feature and the environment-responsive behavior feature according to the driving weight parameter and the feedback correction parameter to generate a behavior evolution path feature sequence reflecting the behavior evolution trend of the intelligent agent; And performing behavior pattern clustering processing on the behavior evolution path feature sequence, determining a behavior pattern type label of the agent according to a clustering result, and performing matching processing on the behavior pattern type label and a preset risk behavior feature library to generate an agent behavior risk quantification evaluation result. According to a second aspect of the present application, there is provided an enterprise AI-agent-oriented multidimensional behavior modeling and risk quantification system, the enterprise AI-agent-oriented multidimensional behavior modeling and risk quantification system including a machine-readable storage medium storing machine-executable instructions and a processor, the processor implementing the aforementioned enterprise AI-agent-oriented multidimensional behavior modeling and risk quantification method when executing the machine-executable instructions. According to any one of the aspects, the application has the technical effects that: Through collecting an agent behavior data stream which is genera