Artificial intelligence-based work data decision hub risk identification method and system

Abstract

This application relates to the technical field of enterprise intelligent digitalization, and discloses a risk identification method and system for the decision-making center based on artificial intelligence, including: collecting multi-source heterogeneous work data and performing full-link correlation alignment to obtain a standardized decision-making full-link dataset; extracting multi-dimensional features to generate a multi-modal feature set; outputting time-series anomaly scores and correlation anomaly scores through a dual-path risk identification AI model; fusing and calculating a comprehensive risk score, matching risk levels, and generating an interpretable risk description document; pushing the identification results and triggering control operations to complete incremental model optimization. This invention realizes dynamic, comprehensive, and interpretable risk identification and proactive intervention throughout the entire decision-making process, improving the risk management capabilities and scenario adaptability of the enterprise decision-making center.

Description

Technical Field

[0001] This application relates to the technical field of enterprise intelligent digitalization, and in particular to a risk identification method and system for a work data decision-making center based on artificial intelligence. Background Technology

[0002] As enterprises deepen their digital transformation, core activities such as business operations, management, execution, and resource allocation are all completed through digital systems, forming work data covering the entire process from decision initiation, approval, implementation, and feedback. The work data decision-making hub, as the core carrier of enterprise digital management, plays a crucial role in aggregating all work data, supporting decision-making at all levels, and controlling the decision execution process. Its risk identification capabilities directly determine the compliance, stability, and resilience of enterprise operations.

[0003] The existing risk identification schemes in decision-making centers have many technical flaws: First, most of them adopt a threshold judgment mode based on fixed rules, which relies on manually preset risk rules and verification thresholds. They can only identify known and standardized risk scenarios. For dynamic risks brought about by changes in enterprise business models and organizational structure, as well as implicit risks without clear rules, there are serious problems of missed judgment and misjudgment. They cannot adapt to the complex and ever-changing decision-making scenarios of enterprises.

[0004] Secondly, most existing AI-based risk identification solutions analyze single-dimensional work data, such as building identification models only for financial data, approval data, or business execution data. They fail to connect multi-source heterogeneous work data across the entire decision-making process, making it impossible to capture the risk transmission relationship from decision initiation to execution feedback. This results in limited coverage of risk identification and an inability to identify complex risks that cross links and entities.

[0005] Third, existing AI risk identification models mostly adopt a black-box deep learning architecture, which can only output risk identification results and cannot clearly explain the decision-making process, related basis and triggering logic of risk generation. However, the risk identification results of the decision-making center need to meet the traceability requirements of internal audit and industry regulatory compliance. The insufficient interpretability of black-box models makes it impossible to directly apply the identification results to the compliance control and risk interception of the decision-making process.

[0006] Fourth, most existing solutions adopt a post-event risk tracing identification model, which can only conduct risk screening after the decision is implemented. They cannot achieve real-time risk identification and pre-intervention during the decision initiation and approval process, resulting in risk events being dealt with only after they occur. This makes it impossible to reduce the actual losses caused by risk events, and the timeliness of risk management is seriously insufficient.

[0007] Fifth, the iterative optimization of existing models relies on manually labeled risk samples and manually adjusted model parameters. For low-frequency, high-hazard unconventional risk events, there is a problem of poor model generalization ability due to insufficient sample size. The model cannot achieve autonomous optimization based on real-time operating data from the decision center, and the recognition accuracy continues to decline after long-term use. Summary of the Invention

[0008] To address the aforementioned technical issues, this application provides a risk identification method and system for work data decision-making centers based on artificial intelligence.

[0009] Firstly, this application provides a risk identification method for work data decision-making centers based on artificial intelligence, including the following steps: S1. Collect multi-source heterogeneous working data corresponding to the working data decision center, perform data cleaning, standardization mapping and full-link association alignment on the collected multi-source heterogeneous working data, and obtain a standardized decision full-link dataset. S2. Perform multi-dimensional feature extraction on the standardized decision-making full-link dataset to generate a multimodal feature set containing time-series features, subject association features, behavioral features, and compliance verification features; S3. Input the multimodal feature set into the pre-trained dual-path risk identification AI model, and output the decision link temporal anomaly score and the subject association anomaly score respectively through the dual-path risk identification AI model; wherein, the dual-path risk identification AI model includes a temporal causal reasoning sub-model and an association graph neural network sub-model; S4. Based on the fusion calculation of the time-series anomaly score and the subject-related anomaly score of the decision link, a comprehensive risk score of the target decision link is obtained. According to the comprehensive risk score, the corresponding risk level is matched, and an interpretable risk description document corresponding to the risk level is generated. S5. The comprehensive risk score, risk level, and interpretable risk description document are pushed to the work data decision center in real time, triggering the decision process control operation corresponding to the risk level. At the same time, the full-link data of this risk identification and the results of manual review are fed back to the dual-path risk identification AI model to complete the incremental optimization of the model.

[0010] Optionally, in step S1, the multi-source heterogeneous working data includes decision initiation data, approval flow data, business execution data, financial voucher data, organizational authority data, and compliance rule data; the full-link association alignment specifically involves using a unique decision identifier as the primary key to associate and map the full-process data corresponding to the same decision item, filling in the missing data nodes in the decision link, and generating a full-link closed-loop dataset based on decision items.

[0011] Optionally, in step S2, the multi-dimensional feature extraction specifically includes: extracting time-series features based on the time nodes and flow duration of the entire decision-making process; extracting subject-related features based on the permissions, relationships, and historical behaviors of decision-making participants; extracting behavioral features based on the operational behaviors, content modification trajectories, and execution actions of decision-making matters; and extracting compliance verification features based on the verification items of the compliance rule base. The extracted features are then normalized and filtered, and redundant features are removed to generate a multimodal feature set.

[0012] Optionally, in step S3, the temporal causal reasoning sub-model adopts a temporal causal Transformer architecture based on an attention mechanism. It takes temporal features, behavioral features, and compliance verification features from the multimodal feature set as input, mines the causal dependencies of each node in the decision-making link through the causal attention layer, identifies non-compliant temporal abnormal nodes, and outputs the temporal abnormality score of the decision-making link. The attention weight calculation of the causal attention layer excludes the influence of future time nodes on past time nodes and only retains the causal relationship between the preceding node and the following node in the decision-making link.

[0013] Optionally, in step S3, the relational graph neural network sub-model adopts a heterogeneous graph neural network architecture. First, a heterogeneous graph of decision-making associations is constructed with decision-making participants as nodes and decision-making associations, permission associations, and business associations between subjects as edges. The subject association features and behavioral features in the multimodal feature set are mapped to the node features of the corresponding nodes. The neighborhood features of the nodes are aggregated through graph convolutional layers to identify abnormal association relationships and abnormal behavior patterns of the nodes and output the abnormal subject association score.

[0014] Optionally, in step S4, the fusion calculation specifically involves weighting and summing the decision link time-series anomaly score and the subject-related anomaly score using a preset weight allocation coefficient to obtain a comprehensive risk score; the interpretable risk description document includes the decision node that triggers the risk, the related subject, the basis for matching anomaly features, the causal relationship logic, and the corresponding compliance rules and clauses.

[0015] Optionally, in step S5, the decision-making process control operations include normal release, risk warning, approval escalation, process interception and compliance audit triggering, with different risk levels corresponding to different intensity control operations; among them, when the comprehensive risk score reaches the preset interception threshold, the current decision-making process is directly terminated and simultaneously pushed to the enterprise's compliance audit department.

[0016] Optionally, in step S5, the incremental optimization of the model specifically involves using the full-link data of this risk identification, the risk labels after manual review, and the corrected risk levels as incremental training samples, and using the mini-batch gradient descent algorithm to fine-tune the parameters of the dual-path risk identification AI model, while updating the model's feature weights and risk thresholds to complete the incremental iterative optimization of the model.

[0017] Secondly, this application provides a risk identification system for a work data decision-making center based on artificial intelligence, which adopts the following technical solution: The AI-based risk identification system for work data decision-making includes a data governance module, a feature engineering module, a dual-path AI risk identification module, a risk classification and interpretation module, and a process control and model optimization module. The data governance module is used to collect multi-source heterogeneous working data corresponding to the working data decision center, and to perform data cleaning, standardization mapping and full-link association alignment on the collected multi-source heterogeneous working data to obtain a standardized decision full-link dataset. The feature engineering module is used to extract multi-dimensional features from the standardized decision-making full-link dataset to generate a multimodal feature set that includes time-series features, subject association features, behavioral features, and compliance verification features. The dual-path AI risk identification module has a pre-trained dual-path risk identification AI model built in, which is used to input the multimodal feature set into the dual-path risk identification AI model and output the decision link temporal anomaly score and the subject association anomaly score respectively; wherein, the dual-path risk identification AI model includes a temporal causal reasoning sub-model and an association graph neural network sub-model; The risk classification and interpretation module is used to perform a fusion calculation based on the time-series anomaly score and the subject-related anomaly score of the decision link to obtain the comprehensive risk score of the target decision link, and generate an interpretable risk description document corresponding to the risk level based on the comprehensive risk score. The process control and model optimization module is used to push the comprehensive risk score, risk level and interpretable risk description document to the work data decision center in real time, trigger the decision process control operation corresponding to the risk level, and at the same time feed back the full-link data of this risk identification and the manual review results to the dual-path risk identification AI model to complete the incremental optimization of the model.

[0018] In summary, this application includes at least the following beneficial technical effects: 1. This application achieves closed-loop management of data throughout the entire process of decision-making, from initiation to feedback, by performing full-link correlation and alignment of multi-source heterogeneous work data in the work data decision-making center. It constructs a dual-path risk identification AI model, which includes a temporal causal reasoning sub-model and a relational graph neural network sub-model. It conducts risk identification from two dimensions: the temporal causal logic of the decision-making link and the subject correlation relationship. It can identify temporal abnormal behaviors that do not conform to causal logic in the decision-making process, and can also explore implicit correlations and abnormal behavior patterns between decision-making subjects. It gets rid of the dependence of traditional fixed rule engines on known risk scenarios, can effectively adapt to the dynamic business and organizational scenarios of enterprises, and significantly reduce the false negative rate of implicit and unconventional risks. By generating interpretable risk description documents corresponding to risk levels, it clarifies the nodes, basis and causal logic of risk triggering, fully meets the traceability requirements of enterprise internal audit and industry supervision, and enables the risk identification results to be directly applied to the compliance control of the decision-making process, solving the problem of insufficient interpretability of traditional black box AI models. 2. This application achieves proactive risk identification and real-time intervention by pushing risk identification results and triggering corresponding control operations in real time during the decision-making process. It can complete control and interception before risk events occur, avoiding operational losses caused by post-event tracing, and significantly improving the timeliness and security of enterprise decision-making and control. Through the incremental optimization mechanism, the model can be autonomously iterated, continuously optimizing model parameters and identification accuracy based on actual operating data and manual review results. This solves the problem of poor model generalization ability caused by insufficient low-frequency risk samples, and ensures the identification stability and adaptability of the model in long-term use. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the risk identification method for the work data decision center based on artificial intelligence in this application. Detailed Implementation

[0020] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0021] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0022] This application discloses a risk identification method for a work data decision-making center based on artificial intelligence, referring to... Figure 1 It includes the following steps: S1. Collect multi-source heterogeneous working data corresponding to the working data decision center, perform data cleaning, standardization mapping and full-link association alignment on the collected multi-source heterogeneous working data, and obtain a standardized decision full-link dataset. S2. Perform multi-dimensional feature extraction on the standardized decision-making full-link dataset to generate a multimodal feature set containing time-series features, subject association features, behavioral features, and compliance verification features; S3. Input the multimodal feature set into the pre-trained dual-path risk identification AI model, and output the decision link temporal anomaly score and the subject association anomaly score respectively through the dual-path risk identification AI model; wherein, the dual-path risk identification AI model includes a temporal causal reasoning sub-model and an association graph neural network sub-model; S4. Based on the fusion calculation of the time-series anomaly score and the subject-related anomaly score of the decision link, a comprehensive risk score of the target decision link is obtained. According to the comprehensive risk score, the corresponding risk level is matched, and an interpretable risk description document corresponding to the risk level is generated. S5. The comprehensive risk score, risk level, and interpretable risk description document are pushed to the work data decision center in real time, triggering the decision process control operation corresponding to the risk level. At the same time, the full-link data of this risk identification and the results of manual review are fed back to the dual-path risk identification AI model to complete the incremental optimization of the model.

[0023] The specific implementation process of the AI-based risk identification method for work data decision-making centers described in this embodiment is as follows: Step S1: Collect multi-source heterogeneous working data corresponding to the working data decision center, perform data cleaning, standardization mapping and full-link association alignment on the collected multi-source heterogeneous working data, and obtain a standardized decision full-link dataset.

[0024] The multi-source heterogeneous working data specifically includes decision initiation data, approval flow data, business execution data, financial voucher data, organizational authority data, and compliance rule data. Decision initiation data includes the initiating entity, initiation time, content, expected goals, and related project and resource application information for the decision. Approval flow data includes approval nodes, approval entities, approval time, approval opinions, approval modification records, and flow duration information. Business execution data includes the executing entity, execution actions, execution progress, execution results, and effect feedback information for the decision. Financial voucher data includes fund applications, payment vouchers, invoice information, budget execution data, and financial accounting information related to the decision. Organizational authority data includes the organizational structure, job authority, approval authority, business scope, and historical performance information of the decision-making participants. Compliance rule data includes the enterprise's internal management system, industry regulatory standards, and compliance verification rules and prohibitive clauses corresponding to laws and regulations. Data cleaning specifically involves filling missing values, removing duplicate data, filtering outliers, and transforming unstructured data into structured data from the collected raw data. For unstructured text data such as approval opinions and explanations of matters, natural language processing techniques are used to extract entities and perform semantic standardization transformation to obtain structured text feature data.

[0025] Standardized mapping specifically involves constructing a unified data standard dictionary for the enterprise, mapping work data from different sources and in different formats to unified data standard fields, eliminating data format differences between different business systems, and ensuring data consistency and comparability. End-to-end association and alignment specifically involves using a unique decision identifier as the primary key to associate and map all data corresponding to the same decision item throughout the entire process from initiation, approval, execution to feedback, filling in missing data nodes in the decision-making chain, verifying the integrity and logical consistency of the chain data, and generating a complete-chain closed-loop dataset based on decision items, i.e., a standardized decision-making end-to-end dataset.

[0026] Step S2 involves extracting multi-dimensional features from the standardized decision-making full-link dataset to generate a multimodal feature set that includes time-series features, subject association features, behavioral features, and compliance verification features.

[0027] Multi-dimensional feature extraction is specifically divided into four dimensions of feature extraction. The first is temporal feature extraction, which extracts temporal features based on the time nodes, flow duration, time intervals and time sequence order of each node in the entire decision-making process. These features include the total duration of the entire decision-making process, the dwell time of each approval node, the time interval between adjacent nodes, the deviation value of the node temporal sequence from the standard process, and the relative deviation value of the flow duration of similar decision-making items. These features are used to characterize the rationality of the temporal flow of the decision-making process. The second is the extraction of subject association features. Based on the organizational structure, job authority, approval authority, historical behavior records, and inter-subject relationships of the decision-making participants, subject association features are extracted, including the degree of authority matching of decision-making participants, the degree of organizational association between subjects, the frequency of historical cooperation, the risk records of historical joint decision-making matters, and the degree of matching between subject authority and decision-making matters. These features are used to characterize the relationship and authority compliance of decision-making participants. Third is behavioral feature extraction. Based on the operational behavior, content modification trajectory, approval actions, and execution actions in the entire decision-making process, behavioral features such as the frequency of modification of decision content, modification range, modification nodes, frequency of approval rejection and re-initiation, consistency of approval opinions, and deviation between execution actions and decision content are extracted to characterize the abnormality of operational behavior in the entire decision-making process. Fourthly, compliance verification feature extraction is performed. Based on compliance rule data, compliance verification is conducted on the entire process data of decision-making matters, extracting compliance verification features such as matching items between decision-making matters and compliance rules, triggering of prohibitive clauses, compliance verification pass status, and compliance deviation values ​​of similar matters. These features are used to characterize the basic compliance of decision-making matters. After feature extraction, the extracted features are subjected to min-max normalization to map all feature values ​​to the [0,1] interval, eliminating the influence of features with different dimensions on the model. Then, a feature screening method based on mutual information is used to remove redundant features with mutual information below a preset threshold with the risk identification target, and a multimodal feature set is generated after retaining the effective features.

[0028] Step S3 involves inputting the multimodal feature set into the pre-trained dual-path risk identification AI model. The model then outputs the decision link temporal anomaly score and the subject association anomaly score, respectively. The dual-path risk identification AI model includes a temporal causal reasoning sub-model and an association graph neural network sub-model. Before this step, the dual-path risk identification AI model needs to be pre-trained. The pre-training steps are as follows: First, a training dataset is constructed by collecting full-chain operational data of historical decision-making events of enterprises and labeling them with corresponding risk tags and risk types. Risk tags include four levels: normal, low risk, medium risk, and high risk. Risk types include compliance risk, fraud risk, operational risk, and operational risk. Data augmentation is performed on the labeled dataset, including time-series perturbation, feature masking, and negative sample generation, to expand the training sample size to solve the problem of insufficient low-frequency risk samples. The dataset is divided into training set, validation set, and test set in a ratio of 8:1:1. Second, model initialization: parameters were initialized for the temporal causal reasoning sub-model and the relational graph neural network sub-model respectively. The learning rate of the model was set to 0.0001, the batch size to 32, the number of iterations to 100, and the Adam optimizer was used, with the cross-entropy loss function as the model's loss function. Third, joint training involves inputting the multimodal feature set of the training set into two sub-models, which output temporal anomaly scores and correlation anomaly scores respectively. After fusion, a comprehensive risk score is obtained. The model loss is calculated based on the comprehensive risk score and the labeled risk tags. The parameters of the two sub-models are updated through backpropagation. The hyperparameters of the model are adjusted through the validation set. Training is stopped early when the validation set loss no longer decreases for 10 consecutive rounds. The recognition accuracy of the model is verified through the test set, thus completing the pre-training of the model.

[0029] The temporal causal reasoning sub-model adopts a temporal causal Transformer architecture based on an attention mechanism. The architecture specifically includes an input layer, a position encoding layer, a multi-layer causal Transformer encoder, and an output layer. The causal Transformer encoder includes a causal attention layer, a feedforward neural network layer, a layer normalization layer, and a residual connection structure.

[0030] This sub-model takes temporal features, behavioral features, and compliance verification features from the multimodal feature set as input. First, it transforms the dimensions of the input features through the input layer to match the input dimensions of the model. Then, it adds temporal positional encoding to the input features through the positional encoding layer to preserve the temporal sequence information of the decision link. The features with added positional encoding are then input into a multi-layer causal Transformer encoder. The causal attention layer mines the causal dependencies of each node in the decision link and identifies non-compliant temporal anomaly nodes.

[0031] The formula for calculating the attention weights of the causal attention layer is: ; The meanings of each parameter in the formula are as follows: The query matrix is ​​obtained by linear transformation of the input features and is used to represent the feature information of the current decision node. The key matrix is ​​obtained by linear transformation of the input features and is used to represent the feature information of all preceding nodes in the decision-making chain. The value matrix is ​​obtained by linear transformation of the input features and is used to represent the feature weight information of all preceding nodes in the decision-making link. The dimension of the key matrix is ​​used to scale the matrix dot product result to prevent the dot product value from being too large, which would cause the gradient of the softmax function to vanish. This is a causal mask matrix, an upper triangular matrix, where the values ​​of the elements corresponding to future time points are... The element value corresponding to the past and present time nodes is 0, used to shield the influence of future time nodes on past time nodes, retaining only the causal relationship between the preceding node and the following node in the decision-making chain, ensuring that the model only performs causal reasoning based on the node information that has already occurred in the decision-making process, which conforms to the actual temporal logic of the decision-making process; softmax is a normalization exponential function used to normalize the attention weights to the [0,1] interval, so that the sum of the attention weights of all nodes is 1. After feature extraction by a multi-layer causal Transformer encoder, the model performs feature mapping and normalization processing through the output layer, outputting the decision-making chain temporal anomaly score in the 0-1 interval. The higher the score, the higher the degree of temporal causal logic anomaly in the decision-making chain, and the higher the corresponding risk.

[0032] The relational graph neural network sub-model adopts a heterogeneous graph neural network architecture, which specifically includes a graph construction layer, a node feature mapping layer, a multi-layer graph convolutional layer, a readout layer, and an output layer. First, the graph construction layer constructs a heterogeneous graph of decision-making relationships, using decision-making participants as nodes and decision-making, permission, and business relationships between participants as edges. Nodes include different types of decision-making participants such as decision initiators, approvers, executors, and supervisors, while edges include different types of relationships with corresponding weight coefficients. Next, the node feature mapping layer maps the participant relationship features and behavioral features from the multi-modal feature set to the node features of the corresponding nodes, completing the feature initialization of the heterogeneous graph. Then, the initialized heterogeneous graph of decision-making relationships is input into the multi-layer graph convolutional layer. The graph convolutional layer aggregates the neighborhood features of each node, mines implicit relationships between nodes, and identifies abnormal relationships and abnormal behavior patterns. The node feature update calculation formula for the graph convolutional layer is: ; The meanings of each parameter in the formula are as follows: For the first Nodes output by the layer graph convolutional layer eigenvectors; For the first Nodes at the input of a layer graph convolutional layer eigenvectors, For nodes Neighboring nodes eigenvectors; The ReLU activation function is used as a non-linear activation function to introduce non-linear fitting capability into the model. This is the set of all edge types in the heterogeneous graph of decision association, including decision association, permission association, and business association; For nodes In edge type The set of neighboring nodes; The normalization coefficient takes values ​​from the nodes. In edge type The number of neighboring nodes is used to normalize the neighborhood feature aggregation results and avoid feature bias caused by differences in the number of neighboring nodes. For the first Edge type of layer graph convolutional layer A trainable weight matrix is ​​used to perform linear transformations on the neighborhood features of different types of edges; For the first The layered graph convolutional layer corresponds to a trainable weight matrix of the node's own features, which is used to preserve the node's own feature information.

[0033] After feature aggregation and updating through multiple graph convolutional layers, the readout layer performs global pooling and aggregation on the features of all nodes to obtain the global features of the entire decision association heterogeneous graph. Then, the output layer performs feature mapping and normalization to output the subject association anomaly score in the range of 0-1. The higher the score, the higher the degree of abnormality in the association relationship and behavior pattern of the decision participants, and the higher the corresponding risk.

[0034] Step S4: Based on the fusion calculation of the decision-making link time-series anomaly score and the subject-related anomaly score, a comprehensive risk score for the target decision-making link is obtained. The corresponding risk level is matched according to the comprehensive risk score, and an interpretable risk description document corresponding to the risk level is generated. Specifically, the fusion calculation involves weighting and summing the decision-making link time-series anomaly score and the subject-related anomaly score using preset weighting coefficients to obtain the comprehensive risk score. The calculation formula is as follows: ; The meanings of each parameter in the formula are as follows: The comprehensive risk score for the target decision-making link is defined, with a value range of [0,1]. The decision link timing anomaly score is taken in the range of [0,1]. The score represents the main correlation anomaly score, with a value range of [0,1]. The weighting coefficients for the time series anomaly scores. The weighting coefficients for the correlation anomaly scores. The weighting coefficients can be adjusted based on the company's industry attributes and risk management preferences; the default value is [value to be filled in]. =0.5, =0.5.

[0035] Then, based on the comprehensive risk score, the corresponding risk level is matched, specifically when... When corresponding to a low-risk level, When the corresponding medium risk level is reached, The risk level corresponds to a high risk level. Subsequently, an interpretable risk statement document corresponding to the risk level is generated. The document includes the decision-making node that triggers the risk, the related entities, the basis for matching abnormal characteristics, the causal relationship logic, and the corresponding compliance rules and clauses. For decision-making matters at the high risk level, the document also includes risk handling suggestions and compliance audit tips to ensure the traceability and verifiability of risk identification results and meet the requirements of internal audit and regulatory compliance of the enterprise.

[0036] Step S5: The comprehensive risk score, risk level, and interpretable risk description document are pushed to the work data decision center in real time, triggering the decision process control operation corresponding to the risk level. At the same time, the full-link data of this risk identification and the results of manual review are fed back to the dual-path risk identification AI model to complete the incremental optimization of the model.

[0037] The decision-making process control operations are linked one-to-one with risk levels. Low-risk levels trigger normal release operations, while simultaneously pushing risk warning information to the decision-making initiator and the approval entity, without affecting the normal flow of the decision-making process. Medium-risk levels trigger approval escalation operations, adding an approval node from the next higher-level supervisor or compliance department to the original approval process, while simultaneously pushing an explanatory risk statement to the new approval node. The process can only continue after the escalation approval is completed. High-risk levels trigger process interception and compliance audit trigger operations, directly terminating the current decision-making process, while simultaneously pushing the explanatory risk statement and all decision-making data to the enterprise's compliance audit department, initiating a special compliance audit process. The incremental optimization of the model is as follows: after the results of this risk identification are manually reviewed, the full-link data of this risk identification, the manually reviewed risk labels, and the corrected risk levels are used as incremental training samples. The parameters of the dual-path risk identification AI model are fine-tuned using the mini-batch gradient descent algorithm. At the same time, the feature weights and risk thresholds of the model are updated to complete the incremental iterative optimization of the model. The learning rate of incremental training is set to 1 / 10 of the pre-training learning rate to avoid catastrophic forgetting of the model and to ensure that the model retains the ability to identify the original risk scenarios while adapting to new risk scenarios.

[0038] This embodiment also provides an AI-based risk identification system for a work data decision-making center, including a data governance module, a feature engineering module, a dual-path AI risk identification module, a risk classification and interpretation module, and a process control and model optimization module. The functions of each module correspond one-to-one with the aforementioned method steps, realizing the aforementioned risk identification method. The data governance module connects to various business systems within the work data decision-making center, collects corresponding multi-source heterogeneous work data, performs data cleaning, standardization mapping, and full-link association alignment on the collected multi-source heterogeneous work data to obtain a standardized decision-making full-link dataset, and pushes the standardized decision-making full-link dataset to the feature engineering module. The feature engineering module receives the standardized decision-making full-link dataset pushed by the data governance module, performs multi-dimensional feature extraction on the dataset, generates a multimodal feature set including time-series features, subject-related features, behavioral features, and compliance verification features, and pushes the multimodal feature set to the dual-path AI risk identification module. The dual-path AI risk identification module incorporates a pre-trained dual-path risk identification AI model, including a temporal causal reasoning sub-model and a relational graph neural network sub-model. It receives multimodal feature sets pushed by the feature engineering module, inputs them into the dual-path risk identification AI model, and outputs temporal anomaly scores for the decision-making link and subject-related anomaly scores, respectively. These two scores are then pushed to the risk grading and interpretation module. The risk grading and interpretation module receives the two anomaly scores from the dual-path AI risk identification module, performs a fusion calculation to obtain a comprehensive risk score, matches it to the corresponding risk level, generates an interpretable risk description document corresponding to the risk level, and simultaneously pushes the comprehensive risk score, risk level, and interpretable risk description document to the process control and model optimization module and the work data decision center. The process control and model optimization module receives relevant data from the risk grading and interpretation module, triggers decision-making process control operations corresponding to the risk level, and simultaneously collects the entire risk identification data and manual review results, feeding them back to the dual-path risk identification AI model to complete incremental model optimization.

[0039] This embodiment achieves closed-loop management of decision-making data throughout the entire process through end-to-end data association and alignment, comprehensively covering various risk scenarios in the enterprise's decision-making center. The collaborative identification of dual-path AI models ensures both the comprehensiveness and accuracy of risk identification, while also adapting to the dynamically changing business scenarios of the enterprise. At the same time, interpretable risk description documents meet the requirements of compliance control, real-time process control enables proactive intervention in risks, and the incremental optimization mechanism ensures the long-term stable operation of the model, comprehensively improving the risk control capabilities and digital management level of the enterprise's work data decision-making center.

[0040] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. An artificial intelligence based work data decision hub risk identification method, characterized by, Includes the following steps: S1. Collect multi-source heterogeneous working data corresponding to the working data decision center, perform data cleaning, standardization mapping and full-link association alignment on the collected multi-source heterogeneous working data, and obtain a standardized decision full-link dataset. S2. Perform multi-dimensional feature extraction on the standardized decision-making full-link dataset to generate a multimodal feature set containing time-series features, subject association features, behavioral features, and compliance verification features; S3. Input the multimodal feature set into the pre-trained dual-path risk identification AI model, and output the decision link temporal anomaly score and the subject association anomaly score respectively through the dual-path risk identification AI model; wherein, the dual-path risk identification AI model includes a temporal causal reasoning sub-model and an association graph neural network sub-model; S4. Based on the fusion calculation of the time-series anomaly score and the subject-related anomaly score of the decision link, a comprehensive risk score of the target decision link is obtained. According to the comprehensive risk score, the corresponding risk level is matched, and an interpretable risk description document corresponding to the risk level is generated. S5. The comprehensive risk score, risk level, and interpretable risk description document are pushed to the work data decision center in real time, triggering the decision process control operation corresponding to the risk level. At the same time, the full-link data of this risk identification and the results of manual review are fed back to the dual-path risk identification AI model to complete the incremental optimization of the model.

2. The artificial intelligence-based work data decision hub risk identification method according to claim 1, characterized by, In step S1, the multi-source heterogeneous working data includes decision initiation data, approval flow data, business execution data, financial voucher data, organizational authority data, and compliance rule data. The full-link association alignment specifically involves using a unique decision identifier as the primary key to associate and map the full-process data corresponding to the same decision item, filling in the missing data nodes in the decision link, and generating a full-link closed-loop dataset based on decision items.

3. The artificial intelligence-based work data decision hub risk identification method according to claim 1, characterized by, In step S2, the multi-dimensional feature extraction specifically includes: extracting time-series features based on the time nodes and flow duration of the entire decision-making process; extracting subject-related features based on the permissions, relationships, and historical behaviors of decision-making participants; extracting behavioral features based on the operational behaviors, content modification trajectories, and execution actions of decision-making matters; and extracting compliance verification features based on the verification items of the compliance rule base. The extracted features are then normalized and filtered, and redundant features are removed to generate a multimodal feature set.

4. The artificial intelligence-based work data decision hub risk identification method according to claim 1, characterized by, In step S3, the temporal causal reasoning sub-model adopts a temporal causal Transformer architecture based on an attention mechanism. It takes temporal features, behavioral features, and compliance verification features from the multimodal feature set as input, mines the causal dependencies of each node in the decision-making link through the causal attention layer, identifies non-compliant temporal abnormal nodes, and outputs the temporal abnormality score of the decision-making link. The attention weight calculation of the causal attention layer excludes the influence of future time nodes on past time nodes and only retains the causal relationship between the preceding node and the following node in the decision-making link.

5. The artificial intelligence-based work data decision hub risk identification method according to claim 1, characterized by, In step S3, the relational graph neural network sub-model adopts a heterogeneous graph neural network architecture. First, the decision-making participants are used as nodes, and the decision-making associations, permission associations, and business associations between the participants are used as edges to construct a heterogeneous graph of decision associations. The subject association features and behavioral features in the multimodal feature set are mapped to the node features of the corresponding nodes. The neighborhood features of the nodes are aggregated through graph convolutional layers to identify abnormal association relationships and abnormal behavior patterns of the nodes, and output the abnormal subject association score.

6. The risk identification method for a work data decision-making center based on artificial intelligence according to claim 1, characterized in that, In step S4, the fusion calculation specifically involves weighting and summing the decision link time-series anomaly score and the subject-related anomaly score using preset weight allocation coefficients to obtain a comprehensive risk score. The interpretable risk description document includes the decision node that triggers the risk, the related subject, the basis for matching anomaly features, the causal relationship logic, and the corresponding compliance rules and clauses.

7. The risk identification method for a work data decision-making center based on artificial intelligence according to claim 1, characterized in that, In step S5, the decision-making process control operations include normal release, risk warning, approval escalation, process interception and compliance audit triggering. Different risk levels correspond to different levels of control operations. When the comprehensive risk score reaches the preset interception threshold, the current decision-making process is directly terminated and simultaneously pushed to the enterprise's compliance audit department.

8. The risk identification method for a work data decision-making center based on artificial intelligence according to claim 1, characterized in that, In step S5, the incremental optimization of the model specifically involves using the full-link data of this risk identification, the risk labels after manual review, and the corrected risk levels as incremental training samples, and using the mini-batch gradient descent algorithm to fine-tune the parameters of the dual-path risk identification AI model. At the same time, the feature weights and risk thresholds of the model are updated to complete the incremental iterative optimization of the model.

9. A risk identification system for a work data decision-making center based on artificial intelligence, characterized in that: It includes modules for data governance, feature engineering, dual-path AI risk identification, risk classification and interpretation, and process control and model optimization. The data governance module is used to collect multi-source heterogeneous working data corresponding to the working data decision center, and to perform data cleaning, standardization mapping and full-link association alignment on the collected multi-source heterogeneous working data to obtain a standardized decision full-link dataset. The feature engineering module is used to extract multi-dimensional features from the standardized decision-making full-link dataset to generate a multimodal feature set that includes time-series features, subject association features, behavioral features, and compliance verification features. The dual-path AI risk identification module has a pre-trained dual-path risk identification AI model built in, which is used to input the multimodal feature set into the dual-path risk identification AI model and output the decision link temporal anomaly score and the subject association anomaly score respectively; wherein, the dual-path risk identification AI model includes a temporal causal reasoning sub-model and an association graph neural network sub-model; The risk classification and interpretation module is used to perform a fusion calculation based on the time-series anomaly score and the subject-related anomaly score of the decision link to obtain the comprehensive risk score of the target decision link, and generate an interpretable risk description document corresponding to the risk level based on the comprehensive risk score. The process control and model optimization module is used to push the comprehensive risk score, risk level and interpretable risk description document to the work data decision center in real time, trigger the decision process control operation corresponding to the risk level, and at the same time feed back the full-link data of this risk identification and the manual review results to the dual-path risk identification AI model to complete the incremental optimization of the model.

Images