A disciplinary inspection multi-source heterogeneous time sequence data behavior attribution evaluation method and device, electronic equipment and storage medium

Abstract

The application discloses a discipline inspection multi-source heterogeneous time sequence data behavior attribution evaluation method and device, electronic equipment and storage medium, belongs to artificial intelligence technical field, the method includes: obtaining discipline inspection multi-source heterogeneous time sequence data and track overall return; the time sequence data is carried out element extraction and vectorization mapping, generates a plurality of structured semantic elements and corresponding semantic feature vectors; according to each semantic feature vector carries out latent reward mapping and decoding processing, obtains the contribution of each structured semantic element corresponding agent immediate reward;Combined with each agent immediate reward and track overall return contribution distribution calculation, determine the contribution degree of each structured semantic element;Each structured semantic element, each multi-dimensional latent reward vector and each contribution degree are input into a large language model to generate a discipline inspection behavior attribution evaluation report. Through the implementation of the application, the problem of inaccurate behavior attribution evaluation of discipline inspection multi-source heterogeneous time sequence data in the prior art can be solved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method, apparatus, electronic device, and storage medium for behavioral attribution assessment of multi-source heterogeneous time-series data in discipline inspection and supervision. Background Technology

[0002] Case leads and review materials in disciplinary inspection and supervision work typically cover lengthy approval processes and financial irregularities. These records constitute extremely complex, multi-source, heterogeneous time-series data for disciplinary inspection and supervision. Accurate behavioral attribution assessment of this type of time-series data is crucial. This not only helps staff clarify the evolution of disciplinary incidents from massive and fragmented case files but also accurately analyzes which specific actions led to the final violation over a long timeline. This is the core foundation for the transformation of disciplinary inspection and supervision work from time-consuming and laborious manual review to efficient and intelligent supervision.

[0003] However, existing intelligent assisted review methods often struggle to accurately trace and assess intermediate sub-behaviors when processing long-term data, leading to inaccurate behavioral attribution assessments of multi-source heterogeneous time-series data in discipline inspection and supervision. The main reason for this problem is that current assessment processes typically only provide an overall final evaluation conclusion after collecting all sequential data, lacking a computational mechanism to back-allocate the global evaluation result to each historical step. Furthermore, existing methods struggle to extract structured elements from chaotic multi-source text into multi-dimensional numerical calculations. Because a quantitative value mapping relationship cannot be established between the overall result and local actions, existing technologies cannot accurately calculate the actual contribution of each independent action on the timeline to the final result. Consequently, the generated assessment reports often only contain macro-level qualitative conclusions, lacking quantitative attribution evidence down to specific action stages. Summary of the Invention

[0004] This invention provides a method, apparatus, electronic device, and storage medium for behavioral attribution assessment of multi-source heterogeneous time-series data in discipline inspection and supervision, which can solve the problem of inaccurate behavioral attribution assessment of multi-source heterogeneous time-series data in discipline inspection and supervision in the prior art.

[0005] An embodiment of the present invention provides a method for behavioral attribution assessment of multi-source heterogeneous time-series data in discipline inspection and supervision, comprising: Acquire multi-source heterogeneous time-series data of discipline inspection and supervision and the overall trajectory report; wherein, the overall trajectory report is used to characterize the comprehensive evaluation result of the supervision trajectory composed of multi-source heterogeneous time-series data of discipline inspection and supervision; Element extraction is performed on multi-source heterogeneous time-series data of discipline inspection and supervision to generate multiple structured semantic elements; vectorization mapping is performed on each structured semantic element to generate semantic feature vectors. The potential reward mapping is calculated based on each semantic feature vector to generate each multidimensional potential reward vector; the multidimensional potential reward vector is then decoded to generate each agent's instant reward. The contribution of each structured semantic element is determined by allocating and calculating the immediate rewards of each agent and the overall return of the trajectory. Each structured semantic element, each multidimensional potential reward vector, and the contribution degree corresponding to each structured semantic element are input into a preset large language model, so that the large language model generates an attribution assessment report of disciplinary inspection and supervision behavior based on each structured semantic element, each multidimensional potential reward vector, and the contribution degree corresponding to each structured semantic element.

[0006] Furthermore, element extraction is performed on the multi-source heterogeneous time-series data of discipline inspection and supervision to generate multiple structured semantic elements, including: Multiple entities and events were extracted from the multi-source heterogeneous time-series data of discipline inspection and supervision, along with the roles, timestamps, and attributes corresponding to each entity and event. Each entity and each event is identified as a node in the semantic graph; Based on the roles, timestamps, and attributes of each entity and event, determine the associated edges connected to each node; For each node, the current node and the associated edges connected to the current node are combined to form the structured semantic elements corresponding to the current node.

[0007] Furthermore, each structured semantic element is vectorized and mapped to generate semantic feature vectors, including: The nodes contained in each structured semantic element are initially embedded to generate the initial vector representation corresponding to each node; For each structured semantic element, the nodes contained in the current structured semantic element are determined as the current node; the nodes other than the current node corresponding to the associated edges connected to the current node are taken as the associated nodes of the current node; based on the associated edges contained in the current structured semantic element, feature aggregation calculation is performed on the initial vector representation corresponding to the current node and the initial vector representation corresponding to the associated nodes of the current node to generate the semantic feature vector corresponding to the current structured semantic element.

[0008] Furthermore, based on each semantic feature vector, potential reward mapping is calculated to generate various multi-dimensional potential reward vectors, including: For each semantic feature vector, the reward feature components of the current semantic feature vector under each preset supervision and evaluation dimension are calculated based on the preset reward mapping function. The reward feature components of the current semantic feature vector are concatenated and combined to generate a multidimensional potential reward vector corresponding to the current semantic feature vector.

[0009] Furthermore, the multi-dimensional potential reward vectors are decoded to generate instant rewards for each agent, including: For each multidimensional potential reward vector, based on the preset evaluation weights corresponding to each supervised evaluation dimension, the reward feature components contained in the current multidimensional potential reward vector are weighted and fused to generate the fused reward value corresponding to the current multidimensional potential reward vector. The fusion reward value corresponding to the current multidimensional potential reward vector is normalized to generate the agent instant reward corresponding to the current multidimensional potential reward vector.

[0010] Furthermore, based on the immediate rewards for each agent and the overall trajectory reward, the contribution of each structured semantic element is determined, including: Based on the timestamps corresponding to each structured semantic element, the structured semantic elements are arranged in chronological order to generate a global time series sequence. For each structured semantic element, based on a preset time discount factor, the agent's immediate reward corresponding to the current structured semantic element and the agent's immediate reward corresponding to each structured semantic element located after the current structured semantic element in the global time series are calculated to generate the cumulative return prediction value corresponding to the current structured semantic element. Calculate the difference between the cumulative return prediction value corresponding to the current structured semantic element and the overall trajectory return to determine the contribution of the current structured semantic element.

[0011] Furthermore, the pre-defined large language model is trained in the following ways: Obtain several training samples; each training sample includes historical structured semantic elements, historical multidimensional potential reward vectors, historical contribution, and corresponding historical attribution evaluation labels; the historical attribution evaluation labels are historical attribution evaluation reports; Each training sample is input into the large language model to train it until a preset number of training iterations are reached. Each time the large language model receives a training sample, it outputs a prediction attribution evaluation report corresponding to that training sample. Based on the prediction attribution evaluation report and the corresponding historical attribution evaluation labels, a loss function value is calculated. The large language model is then updated based on the loss function value.

[0012] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0013] One embodiment of the present invention provides a behavioral attribution assessment device for multi-source heterogeneous time-series data of discipline inspection and supervision, including: a data acquisition module, a semantic element processing module, a reward calculation module, and an attribution assessment module; The data acquisition module is used to acquire multi-source heterogeneous time-series data of discipline inspection and supervision, as well as the overall trajectory report; wherein, the overall trajectory report is used to characterize the comprehensive evaluation result of the supervision trajectory composed of multi-source heterogeneous time-series data of discipline inspection and supervision. The semantic element processing module is used to extract elements from multi-source heterogeneous time-series data of discipline inspection and supervision, generate multiple structured semantic elements, and perform vectorization mapping on each structured semantic element to generate each semantic feature vector. The reward calculation module is used to perform potential reward mapping calculation based on each semantic feature vector to generate each multidimensional potential reward vector; decode each multidimensional potential reward vector to generate each agent's instant reward; and perform allocation calculation based on each agent's instant reward and the overall trajectory reward to determine the contribution degree corresponding to each structured semantic element. The attribution assessment module is used to input each structured semantic element, each multidimensional potential reward vector, and the contribution degree corresponding to each structured semantic element into a preset large language model, so that the large language model generates an attribution assessment report of disciplinary inspection and supervision behavior based on each structured semantic element, each multidimensional potential reward vector, and the contribution degree corresponding to each structured semantic element.

[0014] Based on the above method embodiments, the present invention provides corresponding electronic device embodiments.

[0015] An embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the behavioral attribution evaluation method for multi-source heterogeneous time-series data of discipline inspection and supervision as described in any of the above-described method embodiments.

[0016] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments.

[0017] One embodiment of the present invention provides a storage medium storing a computer program thereon, wherein, when the computer program is running, it controls the device where the storage medium is located to execute the behavioral attribution evaluation method for multi-source heterogeneous time-series data of discipline inspection and supervision as described in any of the above-described method embodiments.

[0018] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a method, apparatus, electronic device, and storage medium for behavioral attribution assessment of multi-source heterogeneous time-series data in discipline inspection and supervision. The method acquires multi-source heterogeneous time-series data and overall trajectory returns from discipline inspection and supervision; extracts and vectorizes the time-series data to generate multiple structured semantic elements and corresponding semantic feature vectors; performs potential reward mapping and decoding processing based on each semantic feature vector to obtain the proxy immediate reward corresponding to each structured semantic element; calculates the contribution allocation by combining each proxy immediate reward with the overall trajectory return to determine the contribution degree of each structured semantic element; and inputs each structured semantic element, each multi-dimensional potential reward vector, and each contribution degree into a preset large language model to generate a discipline inspection and supervision behavioral attribution assessment report.

[0019] This invention addresses the technical challenge of extracting structured elements from chaotic text into multi-dimensional numerical calculations by extracting and vectorizing elements from multi-source heterogeneous time-series data. Furthermore, it calculates immediate agent rewards based on the vectorized semantic features and combines this with the overall trajectory reward for allocation calculation to determine the true contribution of each element. This successfully constructs a computational mechanism for value mapping and reverse allocation from global evaluation results to historical local actions. Finally, this solution inputs the quantified contribution and related multi-dimensional features into a large language model to generate a report, directly overcoming the limitations of traditional reports that only offer macro-level qualitative conclusions. It supplements the data with quantitative attribution evidence down to specific action stages, significantly improving the accuracy of long-term behavioral attribution assessments in discipline inspection and supervision. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a behavioral attribution assessment method for multi-source heterogeneous time-series data in discipline inspection and supervision, provided by an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of a behavioral attribution assessment device for multi-source heterogeneous time-series data in discipline inspection and supervision, provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, to address the problem of inaccurate behavioral attribution assessment of multi-source heterogeneous time-series data in discipline inspection and supervision in existing technologies, an embodiment of the present invention provides a method for behavioral attribution assessment of multi-source heterogeneous time-series data in discipline inspection and supervision, comprising at least the following steps: Step S1: Obtain multi-source heterogeneous time-series data and overall trajectory report of discipline inspection and supervision; wherein, the overall trajectory report is used to characterize the comprehensive evaluation result of the supervision trajectory composed of multi-source heterogeneous time-series data of discipline inspection and supervision.

[0024] Specifically, this involves acquiring multi-source heterogeneous time-series data on discipline inspection and supervision, as well as an overall trajectory report. The overall trajectory report is used to characterize the comprehensive evaluation result of the supervision trajectory composed of multi-source heterogeneous time-series data on discipline inspection and supervision.

[0025] In the implementation of discipline inspection and supervision work, the first step is to collect raw data from multiple sources, in different formats, and with time-series attributes related to the supervised entities. This multi-source, heterogeneous time-series data encompasses materials from letters and visits reporting violations, case-handling inquiry records, important meeting minutes, financial audit vouchers, online public opinion information, and routine office documents. These various types of heterogeneous data are interconnected on a timeline, collectively forming a complete business trajectory from accepting leads, initiating investigations, to closing cases and imposing disciplinary actions or a closed-loop supervision system. All the acquired discrete information is integrated according to the chronological order of occurrence to form a supervisory trajectory that reflects the evolution path of disciplinary risks.

[0026] The overall trajectory return is a monitoring signal obtained by quantitatively evaluating the final execution effectiveness of a specific monitored trajectory; it is a domain knowledge-driven evaluation result. To ensure the objectivity of the evaluation results, a multi-dimensional evaluation index is used to weight and sum the overall trajectory return. The specific calculation formula is as follows: In the above formula, Indicates the overall return of the trajectory; Indicates the monitoring trajectory; This indicates the pre-defined weighting of the outcome. This indicates the preset timeliness weight; Indicates the preset compliance weight; This indicates the pre-defined weight of the impact of the rectification measures; The score represents the outcome effectiveness indicator, used to reflect the objective and quantitative degree of the case's qualitative outcome; The score represents the timeliness indicator, used to measure processing efficiency; This represents the compliance indicator score, used to calculate whether the process complies with legally mandated procedural norms; The score for the social rectification impact indicator reflects the effectiveness of post-case warning education and rectification implementation. The weights for each of these evaluation criteria are typically pre-set by staff based on the specific type of supervisory task.

[0027] For the specific calculation of the timeliness index score, it is necessary to compare the difference between the actual completion time and the standard time limit, and use the following formula for normalization: In the formula for calculating timeliness, The base of the natural logarithm; This indicates the preset timeout penalty parameter; This represents the time deviation value. The time deviation value is calculated as the difference between the actual total processing time of the monitored trajectory and the statutory standard completion time limit. When the actual total processing time is greater than the statutory standard completion time limit, the time deviation value is equal to the positive difference obtained by subtracting the two time values; when the actual total processing time is less than or equal to the statutory standard completion time limit, the time deviation value is set to zero.

[0028] By acquiring multi-source heterogeneous time-series data and overall trajectory reports from discipline inspection and supervision, we not only achieved efficient integration of massive amounts of fragmented supervision information, but also provided a globally authoritative benchmark for the subsequent refined allocation of behavioral contributions.

[0029] Step S2: Extract elements from the multi-source heterogeneous time-series data of discipline inspection and supervision to generate multiple structured semantic elements; perform vectorization mapping on each structured semantic element to generate each semantic feature vector.

[0030] In a preferred embodiment, element extraction is performed on multi-source heterogeneous time-series data of discipline inspection and supervision to generate multiple structured semantic elements, including: Multiple entities and events were extracted from the multi-source heterogeneous time-series data of discipline inspection and supervision, along with the roles, timestamps, and attributes corresponding to each entity and event. Each entity and each event is identified as a node in the semantic graph; Based on the roles, timestamps, and attributes of each entity and event, determine the associated edges connected to each node; For each node, the current node and the associated edges connected to the current node are combined to form the structured semantic elements corresponding to the current node.

[0031] In a preferred embodiment, each structured semantic element is vectorized and mapped to generate semantic feature vectors, including: The nodes contained in each structured semantic element are initially embedded to generate the initial vector representation corresponding to each node; For each structured semantic element, the nodes contained in the current structured semantic element are determined as the current node; the nodes other than the current node corresponding to the associated edges connected to the current node are taken as the associated nodes of the current node; based on the associated edges contained in the current structured semantic element, feature aggregation calculation is performed on the initial vector representation corresponding to the current node and the initial vector representation corresponding to the associated nodes of the current node to generate the semantic feature vector corresponding to the current structured semantic element.

[0032] Specifically, for the acquired multi-source heterogeneous time-series data on discipline inspection and supervision, in order to break down the analytical barriers caused by unstructured text, it is necessary to carry out underlying information mining and transformation operations. Specifically, multiple entities and events are extracted from the multi-source heterogeneous time-series data on discipline inspection and supervision, and the roles, timestamps, and attribute information corresponding to each entity and event are simultaneously parsed. Entities include case handlers, involved parties, and related organizations; events include approval and processing nodes, financial loan disbursements, and violations of discipline and regulations; roles define the specific responsibilities of a particular entity in a specific event; attributes include the amount involved, the department where the incident occurred, and key case-related terms. To rigorously represent the above extraction results, the following element extraction mapping relationship is constructed: In the formula, This indicates the mapping relationship for element extraction; This represents the input of heterogeneous time-series data from multiple sources related to discipline inspection and supervision. Indicates the extracted first An entity or event; Indicates the first The role corresponding to an entity or event; Indicates the first The timestamp corresponding to an entity or event; Indicates the first The attributes corresponding to an entity or event; This represents the total number of entities and events extracted.

[0033] After completing the basic information mining, all the entities and events obtained are uniformly transformed and identified as nodes in the semantic graph. Then, based on the roles, timestamps, and attribute information corresponding to each entity and event, the connecting edges between the nodes are clearly defined. These connecting edges not only reflect the logical dependencies between fund flows and business approvals but also map causal relationships occurring sequentially. For each node within the semantic graph, the current node and the connecting edges connecting it are fused and assembled to derive and generate the structured semantic elements corresponding to the current node. By using a construction method that overlays local connections around a single node, it is ensured that each structured semantic element can internalize the micro-level business context.

[0034] After obtaining a series of structured semantic elements, cross-modal numerical space transformation is required, which involves vectorizing all structured semantic elements. First, initial embedding processing is performed on the nodes contained within each structured semantic element, mapping the node features of the textual description into a continuous low-dimensional numerical matrix, thereby generating the initial vector representation corresponding to each node. The mathematical logic of the initial embedding processing is as follows: In the formula, Indicates the first The initial vector representation corresponding to each entity or event; This represents a predefined text feature embedding model.

[0035] Next, for each structured semantic element, the nodes contained within the currently processed structured semantic element are identified as the current node. Following the previously established semantic graph topology, the nodes connected to the current node by its associated edges (excluding the current node itself) are designated as its associated nodes. Then, based on the associated edges contained in the current structured semantic element, feature aggregation calculations are performed on the initial vector representation corresponding to the current node and the initial vector representations corresponding to its associated nodes, ultimately generating the semantic feature vector corresponding to the current structured semantic element. Feature aggregation calculations, by aggregating neighborhood information, enable the final generated feature parameters to deeply integrate global temporal span and local connectivity attributes. The specific calculation formula is as follows: In the formula, This represents the generated semantic feature vector; This represents a predefined graph aggregation operator; This represents the initial vector representation of the current node; This represents the initial vector representation of the associated node; This represents the weight value of the edge connecting the current node and its associated nodes.

[0036] Through the above-mentioned element extraction and vectorization mapping process, the data format barriers in multi-source case materials were successfully eliminated, and the transformation from discrete text to high-dimensional compact continuous features was completed, laying a feature foundation with deep semantics and logical connections for subsequent quantification of supervisory value and attribution allocation.

[0037] Step S3: Calculate the potential reward mapping based on each semantic feature vector to generate each multidimensional potential reward vector; decode each multidimensional potential reward vector to generate each agent's instant reward.

[0038] In a preferred embodiment, potential reward mapping is calculated based on each semantic feature vector to generate multi-dimensional potential reward vectors, including: For each semantic feature vector, the reward feature components of the current semantic feature vector under each preset supervision and evaluation dimension are calculated based on the preset reward mapping function. The reward feature components of the current semantic feature vector are concatenated and combined to generate a multidimensional potential reward vector corresponding to the current semantic feature vector.

[0039] In a preferred embodiment, the multidimensional potential reward vectors are decoded to generate instant rewards for each agent, including: For each multidimensional potential reward vector, based on the preset evaluation weights corresponding to each supervised evaluation dimension, the reward feature components contained in the current multidimensional potential reward vector are weighted and fused to generate the fused reward value corresponding to the current multidimensional potential reward vector. The fusion reward value corresponding to the current multidimensional potential reward vector is normalized to generate the agent instant reward corresponding to the current multidimensional potential reward vector.

[0040] Specifically, after extracting features from structured semantic elements and obtaining corresponding semantic feature vectors, the aforementioned multi-dimensional feature information needs to be transformed into quantified business supervision value. For each obtained semantic feature vector, a pre-defined reward mapping function is used to perform calculations, measuring the reward feature components corresponding to the currently processed semantic feature vector under multiple preset supervision and evaluation dimensions. The preset supervision and evaluation dimensions cover the key dimensions of case handling compliance, timeliness assessment, and integrity risk that are of primary concern in discipline inspection and supervision work.

[0041] The preset reward mapping function is not a fixed static formula, but rather a dynamic encoding logic generated based on the large language model and business verification mechanism. Before formally establishing the preset reward mapping function, the large language model conducts multiple rehearsals based on the historical supervised state of the input, and verifies the rationality of the mapping logic by constructing an error signal. The calculation logic of the error signal is as follows: In the formula, This represents the overall verification error signal used to guide the optimization iteration of large language models; This represents the first penalty weight parameter set for deviations from the numerical range; This indicates the amount of overflow error caused by the test output value deviating from the preset reasonable range. This represents the second penalty weight parameter set for conflicts in business rules; This indicates the logical violation error in the test output, where the value deviates from the common sense and logic of the case. This represents the third penalty weight parameter set to account for fluctuations in the output value; It represents the amount of stable fluctuation error in test results when dealing with small input disturbances.

[0042] Once the overall verification error signal meets the convergence condition, the current mapping logic is locked as the preset reward mapping function for formal invocation. Next, the calculated reward feature components of the current semantic feature vector under all supervised evaluation dimensions are concatenated along a fixed matrix dimension to generate a multi-dimensional potential reward vector specific to the current semantic feature vector.

[0043] After obtaining all the multidimensional potential reward vectors, a dimensionality reduction decoding process is then performed. For each multidimensional potential reward vector, the evaluation weights corresponding to each supervision and evaluation dimension are first obtained. These evaluation weights characterize the importance of different review perspectives within the context of the specific case. Subsequently, based on the obtained evaluation weights, a weighted fusion calculation is performed on all reward feature components covered in the current multidimensional potential reward vector, thereby outputting the fused reward value corresponding to the current multidimensional potential reward vector. To avoid drastic fluctuations in the absolute magnitude of values ​​generated at different case-handling stages, a numerical range constraint operation must be performed on the fused reward value. Through normalization processing, the fused reward value is smoothly mapped to a standard numerical range, ultimately generating the proxy immediate reward corresponding to the current multidimensional potential reward vector.

[0044] The output of the surrogate immediate reward is highly dependent on the parameters within the decoding network. To ensure that the surrogate immediate reward output by the decoding process objectively reflects the true contribution to the global supervision results, the previously obtained overall trajectory reward is introduced as the true supervision signal, and the following reconstruction loss function is constructed to update the network: In the formula, This represents the reconstruction loss function used to update the parameters of the decoding network; Operators that represent the calculation of mathematical expectation; This represents the overall reward value of the acquired trajectory; This represents the total number of time steps of structured semantic elements encompassed within the current complete supervision trajectory; Indicates the first The agent's instant reward value is output after decoding at each time step node.

[0045] By executing the aforementioned potential reward mapping calculation and decoding processing mechanism, the obscure and complex text semantic vectors were successfully and accurately transformed into standardized quantitative values ​​that can be directly used to measure business contribution, completely breaking through the technical bottleneck of difficulty in quantifying intermediate subdivision behaviors under the traditional review model.

[0046] Step S4: Calculate the allocation based on the immediate rewards of each agent and the overall trajectory reward to determine the contribution of each structured semantic element.

[0047] In a preferred embodiment, the contribution of each structured semantic element is determined by allocating and calculating based on the immediate rewards of each agent and the overall trajectory reward, including: Based on the timestamps corresponding to each structured semantic element, the structured semantic elements are arranged in chronological order to generate a global time series sequence. For each structured semantic element, based on a preset time discount factor, the agent's immediate reward corresponding to the current structured semantic element and the agent's immediate reward corresponding to each structured semantic element located after the current structured semantic element in the global time series are calculated to generate the cumulative return prediction value corresponding to the current structured semantic element. Calculate the difference between the cumulative return prediction value corresponding to the current structured semantic element and the overall trajectory return to determine the contribution of the current structured semantic element.

[0048] Specifically, after calculating the immediate rewards for each agent, causal attribution calculations need to be performed on these values ​​over time. First, based on the timestamp information corresponding to each structured semantic element, all structured semantic elements are sorted according to their chronological order of occurrence. This sorting step eliminates the clutter in spatial storage of discrete data, thereby generating a global time-series sequence with strict sequential logic. This global time-series sequence forms the logical foundation for subsequent credit attribution calculations, ensuring that each supervisory action is evaluated at the correct point in time.

[0049] After obtaining the global time series, time-series decay calculation is performed for each structured semantic element in the series based on a preset time discount factor. The core logic of the time-series decay calculation is that the value of a behavior at a specific point in time depends not only on the current immediate feedback but also on a series of subsequent chain reactions. Specifically, the immediate agent reward corresponding to the current structured semantic element and the immediate agent rewards corresponding to each structured semantic element following the current structured semantic element in the global time series are weighted and summed. By introducing weights that gradually decay with increasing time span, the cumulative reward prediction value corresponding to the current structured semantic element is generated. The formula for calculating the cumulative reward prediction value is shown below: In the time decay calculation formula Indicates the first The cumulative return prediction value corresponding to each structured semantic element; This indicates the position index of the current structured semantic element in the global temporal sequence; This represents the total number of structured semantic elements contained in the global temporal sequence; This represents a preset time discount factor, the value of which ranges from zero to one; Represents the first in the global time series sequence The agent provides an immediate reward corresponding to the structured semantic elements at each location.

[0050] After obtaining the cumulative return prediction value at each time step, these prediction values ​​are compared and analyzed with the pre-acquired overall trajectory return. The contribution of the current structured semantic element is determined by calculating the difference between the cumulative return prediction value corresponding to the current structured semantic element and the overall trajectory return. The cumulative return prediction value represents the model's estimate of the expected future return in the current state, while the overall trajectory return represents the final objective result of the entire trajectory. The numerical difference between the cumulative return prediction value and the overall trajectory return reflects the degree of deviation or contribution of the current specific behavior to the final supervision effectiveness; a larger numerical difference usually means that the behavior has a more significant impact or contribution on the final result. The specific contribution determination formula is as follows: In the formula, Indicates the first The contribution of each structured semantic element; Indicates the first The cumulative return prediction value corresponding to each structured semantic element; This represents the overall return value of the trajectory.

[0051] By executing the above-mentioned allocation calculation and contribution determination process, the reverse attribution from the global final evaluation to the micro-level historical behavior is realized. This can accurately locate key nodes in the long-term case handling process and significantly improve the quantitative accuracy and objectivity of behavioral attribution assessment.

[0052] Step S5: Input each structured semantic element, each multidimensional potential reward vector, and the contribution degree corresponding to each structured semantic element into a preset large language model, so that the large language model generates an attribution assessment report of disciplinary inspection and supervision behavior based on each structured semantic element, each multidimensional potential reward vector, and the contribution degree corresponding to each structured semantic element.

[0053] In a preferred embodiment, a pre-defined large language model is trained in the following manner: Obtain several training samples; each training sample includes historical structured semantic elements, historical multidimensional potential reward vectors, historical contribution, and corresponding historical attribution evaluation labels; the historical attribution evaluation labels are historical attribution evaluation reports; Each training sample is input into the large language model to train it until a preset number of training iterations are reached. Each time the large language model receives a training sample, it outputs a prediction attribution evaluation report corresponding to that training sample. Based on the prediction attribution evaluation report and the corresponding historical attribution evaluation labels, a loss function value is calculated. The large language model is then updated based on the loss function value.

[0054] Specifically, after quantifying and allocating the contribution of behaviors, the aforementioned numerical calculation results need to be transformed into a behavior attribution assessment report that can be intuitively understood by case handlers. The specific steps involve inputting the obtained structured semantic elements, multi-dimensional potential reward vectors, and the contribution corresponding to each structured semantic element into a pre-defined large language model. This large language model fully leverages natural language reasoning and text generation capabilities, establishing logical connections between numerical potential reward distributions, credit allocation weights, and business semantic elements to deeply analyze the supervisory trajectory. The disciplinary inspection and supervision behavior attribution assessment report generated by the large language model covers the behavior contribution chain, risk evolution trajectory, key supervisory node attributions, and targeted supervisory recommendations. The generated calculation logic can be expressed as the following formula.

[0055] In the formula. This indicates the generated attribution assessment report for disciplinary inspection and supervision actions. This represents the generative mapping operator of the large language model. This represents the set of structured semantic elements input. This represents the set of multidimensional potential reward vectors input. This represents the set of input contribution values.

[0056] To ensure that the output of the large language model conforms to the operational norms of discipline inspection and supervision and possesses a high degree of logical rigor, the large language model needs to undergo specialized adaptation training. First, a large number of attribution assessment training samples are obtained. Each attribution assessment training sample consists of case-handling data accumulated over historical periods, specifically including historical structured semantic elements, historical multidimensional potential reward vectors, historical contribution levels, and historical attribution assessment reports serving as the gold standard supervisory signal. These historical attribution assessment reports are professional documents manually written and reviewed by senior case-handling experts for specific historical behavioral sequences.

[0057] During training, each attribution evaluation training sample is sequentially input into the initial large language model to be optimized. Upon receiving each attribution evaluation training sample, the large language model outputs a predicted attribution evaluation report based on its historical elements and quantitative indicators. By comparing the semantic differences and logical consistency between the predicted attribution evaluation report and the corresponding historical attribution evaluation report, a generation loss function is calculated. This generation loss function quantifies the degree of deviation between the model's output and the expert standard report; the specific calculation formula is as follows.

[0058] In the formula. This represents the training and generation loss of a large language model. This represents the total number of training samples used in the attribution assessment that participated in the training. This indicates the total number of words in the target report. Indicates the first The report number The target vocabulary for each position. Indicates the first The combined input of historical structured semantic elements, historical multidimensional potential reward vectors, and historical contribution values ​​from each sample. This represents the training parameter set of the large language model. Using the backpropagation algorithm and gradient descent operator, the parameters are adjusted based on the numerical value of the generation loss function. Make adjustments in real time.

[0059] The above training process continues until a pre-set number of training iterations is reached, or the value of the generation loss function decreases to a preset convergence threshold. After training, the large language model possesses the ability to cross-modal align underlying mathematical features with high-level disciplinary inspection and supervision legal logic. In practical applications, the large language model can automatically identify abnormal behavior sequence patterns with high contribution and explain the inherent reasoning path for determining the risk of such behavior in natural language, thereby providing transparent and auditable decision support for disciplinary inspection and supervision personnel.

[0060] By implementing the above steps, the entire process of generating interpretable attribution reports from sparse supervisory data was successfully automated, significantly improving the intelligent evaluation efficiency and the accuracy of evidence tracing in discipline inspection and supervision work.

[0061] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0062] like Figure 2 As shown, an embodiment of the present invention provides a behavioral attribution assessment device for multi-source heterogeneous time-series data of discipline inspection and supervision, including: a data acquisition module, a semantic element processing module, a reward calculation module, and an attribution assessment module; The data acquisition module is used to acquire multi-source heterogeneous time-series data of discipline inspection and supervision, as well as the overall trajectory report; wherein, the overall trajectory report is used to characterize the comprehensive evaluation result of the supervision trajectory composed of multi-source heterogeneous time-series data of discipline inspection and supervision. The semantic element processing module is used to extract elements from multi-source heterogeneous time-series data of discipline inspection and supervision, generate multiple structured semantic elements, and perform vectorization mapping on each structured semantic element to generate each semantic feature vector. The reward calculation module is used to perform potential reward mapping calculation based on each semantic feature vector to generate each multidimensional potential reward vector; decode each multidimensional potential reward vector to generate each agent's instant reward; and perform allocation calculation based on each agent's instant reward and the overall trajectory reward to determine the contribution degree corresponding to each structured semantic element. The attribution assessment module is used to input each structured semantic element, each multidimensional potential reward vector, and the contribution degree corresponding to each structured semantic element into a preset large language model, so that the large language model generates an attribution assessment report of disciplinary inspection and supervision behavior based on each structured semantic element, each multidimensional potential reward vector, and the contribution degree corresponding to each structured semantic element.

[0063] In a preferred embodiment, the semantic element processing module extracts elements from multi-source heterogeneous time-series data of discipline inspection and supervision, generating multiple structured semantic elements, including: Multiple entities and events were extracted from the multi-source heterogeneous time-series data of discipline inspection and supervision, along with the roles, timestamps, and attributes corresponding to each entity and event. Each entity and each event is identified as a node in the semantic graph; Based on the roles, timestamps, and attributes of each entity and event, determine the associated edges connected to each node; For each node, the current node and the associated edges connected to the current node are combined to form the structured semantic elements corresponding to the current node.

[0064] In a preferred embodiment, the semantic element processing module performs vectorization mapping on each structured semantic element to generate semantic feature vectors, including: The nodes contained in each structured semantic element are initially embedded to generate the initial vector representation corresponding to each node; For each structured semantic element, the nodes contained in the current structured semantic element are determined as the current node; the nodes other than the current node corresponding to the associated edges connected to the current node are taken as the associated nodes of the current node; based on the associated edges contained in the current structured semantic element, feature aggregation calculation is performed on the initial vector representation corresponding to the current node and the initial vector representation corresponding to the associated nodes of the current node to generate the semantic feature vector corresponding to the current structured semantic element.

[0065] In a preferred embodiment, the reward calculation module performs potential reward mapping calculations based on each semantic feature vector to generate multi-dimensional potential reward vectors, including: For each semantic feature vector, the reward feature components of the current semantic feature vector under each preset supervision and evaluation dimension are calculated based on the preset reward mapping function. The reward feature components of the current semantic feature vector are concatenated and combined to generate a multidimensional potential reward vector corresponding to the current semantic feature vector.

[0066] In a preferred embodiment, the reward calculation module decodes each multi-dimensional potential reward vector to generate instant rewards for each agent, including: For each multidimensional potential reward vector, based on the preset evaluation weights corresponding to each supervised evaluation dimension, the reward feature components contained in the current multidimensional potential reward vector are weighted and fused to generate the fused reward value corresponding to the current multidimensional potential reward vector. The fusion reward value corresponding to the current multidimensional potential reward vector is normalized to generate the agent instant reward corresponding to the current multidimensional potential reward vector.

[0067] In a preferred embodiment, the reward calculation module performs allocation calculations based on the immediate rewards of each agent and the overall trajectory reward to determine the contribution of each structured semantic element, including: Based on the timestamps corresponding to each structured semantic element, the structured semantic elements are arranged in chronological order to generate a global time series sequence. For each structured semantic element, based on a preset time discount factor, the agent's immediate reward corresponding to the current structured semantic element and the agent's immediate reward corresponding to each structured semantic element located after the current structured semantic element in the global time series are calculated to generate the cumulative return prediction value corresponding to the current structured semantic element. Calculate the difference between the cumulative return prediction value corresponding to the current structured semantic element and the overall trajectory return to determine the contribution of the current structured semantic element.

[0068] In a preferred embodiment, the attribution evaluation module trains a pre-defined large language model in the following manner: Obtain several training samples; each training sample includes historical structured semantic elements, historical multidimensional potential reward vectors, historical contribution, and corresponding historical attribution evaluation labels; the historical attribution evaluation labels are historical attribution evaluation reports; Each training sample is input into the large language model to train it until a preset number of training iterations are reached. Each time the large language model receives a training sample, it outputs a prediction attribution evaluation report corresponding to that training sample. Based on the prediction attribution evaluation report and the corresponding historical attribution evaluation labels, a loss function value is calculated. The large language model is then updated based on the loss function value.

[0069] Specifically, based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0070] Reference Figure 2 As shown, one embodiment of the present invention provides a behavioral attribution assessment device for multi-source heterogeneous time-series data in discipline inspection and supervision. The device includes a data acquisition module, a semantic element processing module, a reward calculation module, and an attribution assessment module.

[0071] The data acquisition module is used to acquire multi-source heterogeneous time-series data and overall trajectory reports from discipline inspection and supervision. This multi-source heterogeneous time-series data covers approval and processing records, meeting minutes, financial audit documents, and complaint texts with long time spans. The overall trajectory report represents the comprehensive evaluation result of the supervision trajectory composed of this multi-source heterogeneous time-series data. The supervision trajectory represents the complete business process from accepting a tip to closing a case. The overall trajectory report is an objective evaluation value determined based on domain knowledge. To comprehensively and objectively reflect the supervision effect, the following steps are taken: multiplying the preset result effectiveness weight value with the result effectiveness indicator score; multiplying the preset timeliness weight value with the timeliness indicator score; multiplying the preset compliance weight value with the compliance indicator score; and multiplying the preset rectification impact weight value with the rectification impact indicator score. These four products are then weighted and summed to accurately derive the overall trajectory report.

[0072] The semantic element processing module extracts elements from multi-source heterogeneous time-series data of discipline inspection and supervision, generating multiple structured semantic elements. This module extracts multiple entities and events from the data, along with the corresponding roles, timestamps, and attributes for each entity and event. A structured semantic element is defined as a data unit containing node features and local topological connections. The module identifies each entity and event as a node in the semantic graph and determines the associated edges connecting each node based on its corresponding role, timestamp, and attributes. For each node, the module combines the current node and its associated edges to form the corresponding structured semantic element.

[0073] The semantic element processing module is also used to vectorize and map each structured semantic element, generating semantic feature vectors. The module performs initial embedding processing on the nodes contained in each structured semantic element, generating initial vector representations for each node. For each structured semantic element, the module identifies the nodes contained in the current structured semantic element as the current node. The module also identifies the nodes other than the current node corresponding to the edges connected to the current node as its associated nodes. Based on the associated edges contained in the current structured semantic element, the module performs feature aggregation calculations on the initial vector representations of the current node and the initial vector representations of its associated nodes. In the specific feature aggregation operation, a preset graph feature aggregation operator is used to perform deep numerical fusion on the initial vector representations of the current node, the initial vector representations of the associated nodes, and the feature information corresponding to the edges connecting the current node and its associated nodes, thereby generating the semantic feature vector corresponding to the current structured semantic element.

[0074] The reward calculation module performs potential reward mapping calculations based on various semantic feature vectors, generating multi-dimensional potential reward vectors. For each semantic feature vector, the module calculates the reward feature components of the current semantic feature vector under each preset supervision and evaluation dimension based on a preset reward mapping function. The module then concatenates and combines the reward feature components of the current semantic feature vector to generate the corresponding multi-dimensional potential reward vector. These multi-dimensional potential reward vectors possess clear business interpretability, covering dimensions such as integrity assessment, fund compliance quantification, and case handling timeliness verification.

[0075] The reward calculation module also decodes each multidimensional potential reward vector to generate immediate agent rewards. For each multidimensional potential reward vector, the reward calculation module performs weighted fusion calculations on each reward feature component contained in the current multidimensional potential reward vector based on the preset evaluation weights corresponding to each supervision evaluation dimension, generating a fused reward value corresponding to the current multidimensional potential reward vector. The reward calculation module normalizes the fused reward value corresponding to the current multidimensional potential reward vector to generate the immediate agent reward corresponding to the current multidimensional potential reward vector.

[0076] The reward calculation module is also used to allocate and calculate rewards based on the immediate rewards of each agent and the overall trajectory return, determining the contribution of each structured semantic element. The module arranges the structured semantic elements in chronological order according to their timestamps, generating a global time-series sequence. For each structured semantic element, the module performs time-series decay calculations on the immediate agent reward corresponding to the current structured semantic element and the immediate agent rewards corresponding to all structured semantic elements following the current structured semantic element in the global time-series sequence, based on a preset time discount factor, generating a cumulative return prediction value for the current structured semantic element. The module calculates the difference between the cumulative return prediction value and the overall trajectory return for the current structured semantic element, determining its contribution. This contribution value quantifies the intensity of the impact of specific case-handling actions on the final supervisory effectiveness.

[0077] The attribution assessment module is used to input each structured semantic element, each multidimensional potential reward vector, and the contribution degree corresponding to each structured semantic element into a preset large language model, so that the large language model can generate an attribution assessment report of discipline inspection and supervision behavior based on each structured semantic element, each multidimensional potential reward vector, and the contribution degree corresponding to each structured semantic element.

[0078] The attribution assessment module optimizes and updates the large language model using training samples. The module acquires several training samples. Each training sample includes historical structured semantic elements, historical multidimensional potential reward vectors, historical contribution, and corresponding historical attribution assessment labels. These historical attribution assessment labels are historical attribution assessment reports manually annotated and reviewed by case-handling experts. The attribution assessment module inputs each training sample into the large language model for training until a preset number of training iterations is reached. Upon receiving each training sample, the large language model outputs a predicted attribution assessment report corresponding to that training sample.

[0079] The attribution evaluation module calculates and generates a loss function value based on the predicted attribution evaluation report and the corresponding historical attribution evaluation labels. Specifically, the loss calculation mechanism involves inputting the input feature set, composed of historical structured semantic elements, historical multidimensional potential reward vectors, and historical contribution values, into the large language model. By rigorously comparing the predicted text generated by the model with the standard text sequence contained in the historical attribution evaluation report, the negative log-likelihood sum of the conditional probabilities is calculated. This sum is used as the generated loss function value to measure the semantic deviation between the model's predicted output and the expert annotation standards. The attribution evaluation module then performs backpropagation gradient updates on the network training parameters of the large language model based on the obtained generated loss function value.

[0080] By implementing the above-described device embodiments, the present invention achieves end-to-end automated output from raw, fragmented monitoring materials to high-level interpretable assessment reports, significantly enhancing the quantitative accuracy of attribution assessment of disciplinary and supervisory behavior and the transparency of business review.

[0081] It should be noted that the embodiments of the device described above correspond to the embodiments of the present invention described above, and can realize the behavioral attribution assessment method for multi-source heterogeneous time-series data of discipline inspection and supervision as described in any one of the above embodiments of the present invention. Furthermore, the embodiments of the device described above are merely illustrative. The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without creative effort.

[0082] Based on the above-described method embodiments of the present invention, a corresponding embodiment of an electronic device is provided.

[0083] An embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the behavioral attribution evaluation method for multi-source heterogeneous time-series data of discipline inspection and supervision as described in any one of the present invention, or, when the processor executes the computer program, it implements the functions of each module in the above-described device embodiments.

[0084] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the terminal device.

[0085] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0086] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0087] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD card), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0088] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments; Another embodiment of the present invention provides a storage medium including a stored computer program, wherein, when the computer program is running, it controls the device where the storage medium is located to execute the behavioral attribution assessment method for any of the above-described discipline inspection and supervision multi-source heterogeneous time-series data.

[0089] The aforementioned storage medium is a computer-readable storage medium, and the computer program includes computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0090] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0091] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for behavioral attribution assessment of multi-source heterogeneous time-series data in discipline inspection and supervision, characterized in that, include: Acquire multi-source heterogeneous time-series data of discipline inspection and supervision and the overall trajectory report; wherein, the overall trajectory report is used to characterize the comprehensive evaluation result of the supervision trajectory composed of multi-source heterogeneous time-series data of discipline inspection and supervision; Element extraction is performed on multi-source heterogeneous time-series data of discipline inspection and supervision to generate multiple structured semantic elements; vectorization mapping is performed on each structured semantic element to generate semantic feature vectors. The potential reward mapping is calculated based on each semantic feature vector to generate each multidimensional potential reward vector; the multidimensional potential reward vector is then decoded to generate each agent's instant reward. The contribution of each structured semantic element is determined by allocating and calculating the immediate rewards of each agent and the overall return of the trajectory. Each structured semantic element, each multidimensional potential reward vector, and the contribution degree corresponding to each structured semantic element are input into a preset large language model, so that the large language model generates an attribution assessment report of disciplinary inspection and supervision behavior based on each structured semantic element, each multidimensional potential reward vector, and the contribution degree corresponding to each structured semantic element.

2. The behavioral attribution assessment method for multi-source heterogeneous time-series data in discipline inspection and supervision as described in claim 1, characterized in that, Element extraction was performed on multi-source heterogeneous time-series data from discipline inspection and supervision departments to generate multiple structured semantic elements, including: Multiple entities and events were extracted from the multi-source heterogeneous time-series data of discipline inspection and supervision, along with the roles, timestamps, and attributes corresponding to each entity and event. Each entity and each event is identified as a node in the semantic graph; Based on the roles, timestamps, and attributes of each entity and event, determine the associated edges connected to each node; For each node, the current node and the associated edges connected to the current node are combined to form the structured semantic elements corresponding to the current node.

3. The behavioral attribution assessment method for multi-source heterogeneous time-series data in discipline inspection and supervision as described in claim 2, characterized in that, Each structured semantic element is vectorized and mapped to generate semantic feature vectors, including: The nodes contained in each structured semantic element are initially embedded to generate the initial vector representation corresponding to each node; For each structured semantic element, the nodes contained in the current structured semantic element are determined as the current node; the nodes other than the current node corresponding to the associated edges connected to the current node are taken as the associated nodes of the current node; based on the associated edges contained in the current structured semantic element, feature aggregation calculation is performed on the initial vector representation corresponding to the current node and the initial vector representation corresponding to the associated nodes of the current node to generate the semantic feature vector corresponding to the current structured semantic element.

4. The behavioral attribution assessment method for multi-source heterogeneous time-series data in discipline inspection and supervision as described in claim 3, characterized in that, Based on each semantic feature vector, a potential reward mapping is calculated to generate various multi-dimensional potential reward vectors, including: For each semantic feature vector, the reward feature components of the current semantic feature vector under each preset supervision and evaluation dimension are calculated based on the preset reward mapping function. The reward feature components of the current semantic feature vector are concatenated and combined to generate a multidimensional potential reward vector corresponding to the current semantic feature vector.

5. The behavioral attribution assessment method for multi-source heterogeneous time-series data in discipline inspection and supervision as described in claim 4, characterized in that, The multidimensional potential reward vectors are decoded to generate instant rewards for each agent, including: For each multidimensional potential reward vector, based on the preset evaluation weights corresponding to each supervised evaluation dimension, the reward feature components contained in the current multidimensional potential reward vector are weighted and fused to generate the fused reward value corresponding to the current multidimensional potential reward vector. The fusion reward value corresponding to the current multidimensional potential reward vector is normalized to generate the agent instant reward corresponding to the current multidimensional potential reward vector.

6. The behavioral attribution assessment method for multi-source heterogeneous time-series data in discipline inspection and supervision as described in claim 5, characterized in that, The contribution of each structured semantic element is determined by allocating and calculating the immediate rewards for each agent and the overall trajectory reward, including: Based on the timestamps corresponding to each structured semantic element, the structured semantic elements are arranged in chronological order to generate a global time series sequence. For each structured semantic element, based on a preset time discount factor, the agent's immediate reward corresponding to the current structured semantic element and the agent's immediate reward corresponding to each structured semantic element located after the current structured semantic element in the global time series are calculated to generate the cumulative return prediction value corresponding to the current structured semantic element. Calculate the difference between the cumulative return prediction value corresponding to the current structured semantic element and the overall trajectory return to determine the contribution of the current structured semantic element.

7. The behavioral attribution assessment method for multi-source heterogeneous time-series data in discipline inspection and supervision as described in claim 6, characterized in that, The pre-defined large language model is trained using the following methods: Obtain several training samples; each training sample includes historical structured semantic elements, historical multidimensional potential reward vectors, historical contribution, and corresponding historical attribution evaluation labels; the historical attribution evaluation labels are historical attribution evaluation reports; Each training sample is input into the large language model to train it until a preset number of training iterations are reached. Each time the large language model receives a training sample, it outputs a prediction attribution evaluation report corresponding to that training sample. Based on the prediction attribution evaluation report and the corresponding historical attribution evaluation labels, a loss function value is calculated. The large language model is then updated based on the loss function value.

8. A behavioral attribution assessment device for multi-source heterogeneous time-series data in discipline inspection and supervision, characterized in that, include: The module includes a data acquisition module, a semantic element processing module, a reward calculation module, and an attribution evaluation module. The data acquisition module is used to acquire multi-source heterogeneous time-series data of discipline inspection and supervision, as well as the overall trajectory report; wherein, the overall trajectory report is used to characterize the comprehensive evaluation result of the supervision trajectory composed of multi-source heterogeneous time-series data of discipline inspection and supervision. The semantic element processing module is used to extract elements from multi-source heterogeneous time-series data of discipline inspection and supervision, generate multiple structured semantic elements, and perform vectorization mapping on each structured semantic element to generate each semantic feature vector. The reward calculation module is used to perform potential reward mapping calculation based on each semantic feature vector to generate each multidimensional potential reward vector; decode each multidimensional potential reward vector to generate each agent's instant reward; and perform allocation calculation based on each agent's instant reward and the overall trajectory reward to determine the contribution degree corresponding to each structured semantic element. The attribution assessment module is used to input each structured semantic element, each multidimensional potential reward vector, and the contribution degree corresponding to each structured semantic element into a preset large language model, so that the large language model generates an attribution assessment report of disciplinary inspection and supervision behavior based on each structured semantic element, each multidimensional potential reward vector, and the contribution degree corresponding to each structured semantic element.

9. An electronic device, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the behavioral attribution assessment method for multi-source heterogeneous time-series data of discipline inspection and supervision as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device where the storage medium is located to perform the behavioral attribution assessment method for multi-source heterogeneous time-series data of discipline inspection and supervision as described in any one of claims 1 to 7.

Images