Inductive process mining optimization method and system based on multi-agent cooperation

Abstract

The application discloses a kind of inductive process mining optimization method and system based on multi-agent cooperation.First, the low-frequency trajectory variant in event log is denoised and frequency is counted, and an inductive mining algorithm is used to mine a process tree model, an external knowledge base is retrieved in the context of business process, and structured information input is constructed.Second, a multi-agent group collaboration mechanism of "generation-review-optimization" is introduced, wherein the initial model generation group generates an initial process model based on structured information input, the expert review group reviews the initial process model from three dimensions of process tree operator, model structure robustness and activity event completeness, and the model optimization group refines and optimizes the model based on expert review.The method can effectively handle event logs with high complexity, noise interference and a large number of low-frequency behavior variants, understand the context information of business processes, and generate a process model with reasonable structure and accurate behavior.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of process mining, and particularly to an inductive process mining optimization method and system based on multi-agent collaboration. BACKGROUND

[0002] In recent years, under the drive of digital transformation, enterprises have regarded business process mining as one of the key technologies for optimizing processes and improving efficiency. Modern information systems support the execution and management of core business processes while generating massive event logs, providing a rich data basis for understanding the actual execution state of business processes. Process mining, as a new cross-disciplinary field, combines data science and process management technology. Its core goal is to automatically discover high-quality process models from event logs, which requires fine-tuning between multiple key dimensions of the model (such as fitness, accuracy, generalization ability, and complexity).

[0003] Multi-agent collaboration refers to multiple intelligent agents with autonomous decision-making capabilities working together to complete a complex task or achieve overall goals by sharing information, coordinating actions, and cooperating with each other. Multiple agents can work in parallel, handling different parts of the task, thereby greatly improving the efficiency of process mining. At the same time, the system can dynamically increase or decrease the number of agents to adapt to changes in task requirements, and different agents with different specialties can be combined to form a complementary whole to cope with the diverse challenges of process mining.

[0004] Inductive process mining is the mainstream approach in the field of process mining. It adopts a divide-and-conquer approach, detects structured cut points in event logs, splits the original event logs into sub-logs, and performs recursive mining. Finally, it generates a process model that can fit the behavior of the event logs by inductive integration. Compared to other process model mining methods such as heuristic mining and association rule mining, inductive process mining can split complex process structures such as loops, parallelism, and selection branches, ensuring that each sub-structure is simple and analyzable. Then, through inductive logic, the sub-structures are reorganized into a process model containing complete complex structures. In large-scale event logs and high-complexity process scenarios, these characteristics of inductive mining methods are of great significance.

[0005] Although many valuable researches have been done on process mining, there are still some limitations. First, when the business process contains a large number of variants and branches, the process model mined by the existing method may be highly complex, showing an "Italian pasta" structure, and lacking in understandability and optimization guidance value. On the contrary, the excessive pursuit of model simplicity may lead to "underfitting", ignoring important process behaviors and variants. Second, the traditional inductive mining algorithm mainly relies on the basic attributes of the event log such as activity name, timestamp and case identifier, and lacks the understanding ability of the process context information, which greatly limits its potential for deeper process analysis and optimization. Finally, the business process modeling based on large language model has been initially automated, but the input is mainly based on business process documents, and the process mining facing large-scale event logs is still difficult. SUMMARY

[0006] The present application aims at the problems of low quality of process model and inability to effectively utilize business process context in the complex process scenario of event log facing traditional inductive mining, and proposes an inductive process mining optimization method and system based on multi-agent cooperation, aiming to improve the model quality of process mining and reduce the complexity of model structure.

[0007] In the first aspect of the present application, an inductive process mining optimization method based on multi-agent cooperation comprises the following steps:

[0008] S1, constructing an external knowledge base;

[0009] S2, collecting business process event logs and business process background description, and preprocessing the event logs based on the external knowledge base;

[0010] S3, encapsulating the preprocessed information in a structured information input semantic label to construct a structured information input, wherein the structured information input semantic label specifically comprises a process background label, a process rule label, an event log information label, an output example label and a task instruction label;

[0011] S4, taking the structured information input as the prompt word of the initial model generation group, the initial model generation group is composed of a plurality of agents, each agent independently generates a plurality of candidate process tree sequences as a candidate set, and the candidate set is evaluated by a model evaluator, and the initial process model is selected from the candidate set according to the highest F-measure;

[0012] S5, constructing an expert review group, the expert review group is composed of three groups of agent teams playing different expert roles, each group contains a plurality of agents, and the three groups respectively evaluate and propose revision opinions for the initial process model from three dimensions of process tree operator logic correctness, model structure robustness and activity event completeness;

[0013] S6, according to the revision suggestions provided by the expert review group and the prompt words of the initial process model construction model optimization group, the model optimization group is composed of a plurality of agents, each agent independently optimizes the potential defects of the initial process model and generates a plurality of candidate process tree sequences as candidate models, and the candidate models are evaluated by a model evaluator, and the final process model is selected from the candidate models according to the highest F-measure;

[0014] S7, calculating the evaluation index of the final process model, if the evaluation index reaches the preset quality threshold requirement, outputting the final process model; otherwise, repeating steps S4-S7.

[0015] Preferably, in step S1, the external knowledge base is constructed based on LanceDB vector database, the retrieval granularity of the database is set to text block level, the text block size is set to 1024, and the text block overlap window is set to 100; the external knowledge base stores the business process document provided by the business personnel in the form of a vector, and the business process document includes a business execution process described in natural language.

[0016] Preferably, in step S2, the preprocessing specifically includes:

[0017] Extracting activity event entities and trajectory variants in the event log; counting the frequency of each trajectory variant, and removing low-frequency trajectory variants below a denoising threshold, the denoising threshold being set to 0.05% of the total number of trajectory variants;

[0018] Processing the event log using inductive mining and its variants algorithm, and converting the process model obtained by mining into a text-based process tree sequence representation; the inductive mining and its variants algorithm includes Inductive Miner, Inductive Miner Direct and Inductive Miner Infrequent;

[0019] Retrieving the external knowledge base, mapping the business process background description into an embedded vector via a text-embedding-3-large model, and selecting the most similar text blocks after reordering according to the cosine similarity, and then splicing the process tree rules to inject prompt words as enhanced process context, the process tree rules specifically include: sequential execution (→), which specifies that the child nodes are strictly executed in left-to-right order; exclusive selection (X), which specifies that one of its child branches is accurately executed in each process instance; parallel execution (+), which allows all child branches to be executed concurrently; and loop structure (*), which defines a repeatable process fragment.

[0020] Preferably, in step S3, the structured information is input into the semantic label,

[0021] The flow context tag <context>To mark the event log corresponding to the business process background description;

[0022] The process rule label <rules>The event log information tag

[0023] The event log information tag <inputs>to mark three categories of key features extracted from the event log: first log features, the initial sequence of process trees generated by inductive process mining and its variants algorithm; second log features; trace variants and their frequencies, which provide the process traces extracted from the event log in descending order of frequency and their frequencies; third log features, the list of activity events, which contains the complete set of activity events;

[0024] the output example label <outputs>In combination with the few-shot prompt, an example of the candidate process model set to be generated is given;

[0025] The task instruction label <tasks>In combination with the role prompt, the business process expert role indicates clear task instructions.

[0026] Preferably, in step S4, the model evaluator is configured to convert the process tree sequence in text form into a Petri net model, and evaluate the Fitness, Precision, and F-measure of the Petri net model; wherein the Fitness measures the ability of the process tree sequence to replay the observed behavior in the event log, the Precision measures the ability of the process tree sequence to avoid introducing redundant or irrelevant paths into the process tree sequence, and the F-measure is defined as the harmonic mean of the Fitness and the Precision, and the higher the F-measure value, the better the model quality.

[0027] Preferably, in step S5, the prompt words of the three expert review groups include:

[0028] The prompt words of the expert group for the logical correctness dimension of the process tree operator include the following four aspects: timing relationship verification, used to verify whether the execution order of each activity in the model conforms to the inherent logical constraints of the business logic, and output the activities whose output business logic does not conform to the inherent logical constraints; decision structure evaluation, used to evaluate whether the selection gateway in the model accurately maps the real decision points in the business process, judge the rationality, completeness of the branch condition setting and its coverage degree on the business scenario; parallel structure review, used to identify the activity combination in the parallel gateway that does not truly represent the tasks that can be executed concurrently in the business; loop structure analysis, used to verify whether the loop structure accurately reflects the links that need to be repeatedly executed in the business process, and ensure the logical correctness of the loop entry condition and the exit condition;

[0029] The prompt words of the expert group for the model structure robustness dimension include the following three aspects: structure complexity optimization, used to identify unnecessary loops, mergable branches or deadlocks that cause the process to stall in the model; business compliance verification, used to compare the process tree sequence with the predefined business rules, and give the nodes or paths in the model that have compliance deviations; abnormal structure detection, used to analyze the execution frequency and path probability distribution of each activity in the process model based on the event log, instruct the large language model to infer a threshold value based on the initial process model structure and the path probability distribution, and identify low-probability paths in the process model whose probability is lower than the threshold value;

[0030] The prompt words of the expert group for the activity event completeness dimension include the following two aspects: activity accuracy and integrity audit, used to verify whether all activity nodes in the process model accurately reflect the actual business tasks, identify and remove redundant activities, and check for missing business activities in the initial process model; implicit dependency relationship mining, used to identify implicit order constraints and business dependency relationships between activity nodes that are not directly expressed by the model structure.

[0031] Preferably, in step S7, the evaluation metrics include: Fitness, Precision, F-measure, and complexity metrics, wherein the complexity metrics include:

[0032] , , ;

[0033] Fitness measures the ability of a process model to replay observed behaviors in event logs; Precision measures the ability of a process tree sequence to avoid introducing redundant or irrelevant paths into the process tree sequence; F-measure is... and The harmonic mean of the model; a higher F-measure value indicates a better model quality.

[0034] Used to measure the structural complexity of a business process model based on node, edge, and gateway types. Used to measure the cyclic complexity of a business process model, focusing on the branching and looping structure of the process; model complexity metric. for and Harmonic mean The lower the value, the more streamlined the process model.

[0035] A second aspect of the present invention provides an inductive process mining system based on multi-agent collaboration, comprising:

[0036] The external knowledge base module is used to: build an external knowledge base to store business process documents provided by business personnel in vector form, wherein the business process documents include business execution processes described in natural language;

[0037] The preprocessing module is used to: collect business process event logs and business process background descriptions, and preprocess the event logs;

[0038] The structured information input module is used to: encapsulate preprocessed information within structured information input semantic tags to construct structured information input. The structured information input semantic tags specifically include process background tags, process rule tags, event log information tags, output example tags, and task instruction tags.

[0039] An initial model generation module is configured to input structured information as a prompt for an initial model generation group, the initial model generation group is composed of a plurality of agents constructed using a large language model, each agent independently generates a plurality of candidate process tree sequences as a candidate set, and the candidate process tree sequences are evaluated by a model evaluator, and an initial process model is selected from the candidate set according to the highest F-measure;

[0040] An expert review module is configured to construct a multi-agent-based expert review group, the expert review group is composed of three groups of agent teams playing different expert roles, each group contains a plurality of agents constructed using a large language model, the three groups respectively review the initial process model from three dimensions of process tree operator logic correctness, model structure robustness and activity event completeness and provide revision opinions, and finally the multi-dimensional revision opinions are summarized in the order of process tree operator-model structure robustness-activity event completeness.

[0041] A model optimization module is configured to construct a prompt for a model optimization group according to the revision opinions provided by the expert review group and the initial process model, the model optimization group is composed of a plurality of agents constructed using a large language model, each agent independently optimizes potential defects of the initial process model and generates a plurality of candidate process tree sequences as candidate models, and the candidate models are evaluated by a model evaluator, and a final process model is selected from the candidate models according to the highest F-measure.

[0042] A model verification module is configured to calculate an evaluation index of the final process model, and if the evaluation index meets a preset quality threshold requirement, the final process model is output; otherwise, iteration is performed.

[0043] In a third aspect, the application provides a computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the multi-agent cooperative inductive process mining optimization method.

[0044] Compared with the prior art, the application has the following beneficial effects:

[0045] The application provides an inductive process mining optimization method based on multi-agent cooperation. The method first uses event logs to perform low-frequency trajectory denoising and trajectory frequency statistics to extract core trajectories to shorten the context length of the input large language model, constructs a basic process model based on three inductive mining algorithms, uses business process background to search external knowledge base to splice business information as enhanced context, and finally constructs structured information. Subsequently, a "generation-review-optimization" multi-agent group cooperation mechanism is introduced: the generation group generates a candidate process model based on the basic process model and structured context information; the review group reviews the process model from three dimensions of process tree operator, model structure robustness and activity event completeness; the optimization group optimizes the model structure of the process model based on expert review opinions. The method of the application effectively combines traditional inductive mining algorithms and business process information, and in the complex business process scenario of large-scale event logs, the method can mine process models with higher model quality and lower structural complexity. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 A flowchart of the method of the application.

[0047] Figure 2 A framework schematic diagram of the inductive process mining optimization method based on multi-agent cooperation.

[0048] Figure 3 A module composition schematic diagram of the inductive process mining system based on multi-agent cooperation. DETAILED DESCRIPTION

[0049] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the specific embodiments of the application will be described in detail below with reference to the drawings. In the following description, many specific details are set forth in order to provide a thorough understanding of the application. However, the application can be implemented in many different ways other than those described herein, and those skilled in the art can make similar improvements without departing from the scope of the application, therefore the application is not limited to the specific embodiments disclosed below. The technical features in each embodiment of the application can be combined accordingly without conflict.

[0050] As shown in Figure 1 The inductive process mining optimization method based on multi-agent cooperation described in the embodiment includes the following steps:

[0051] S1. Construct an external knowledge base based on the LanceDB vector database. The database retrieval granularity is set to the text block level, the text block size to 1024, and the text block overlap window to 100. Store business process documents provided by business personnel in vector form, including business execution processes described in natural language (for example, when an employee submits an application, it is sent to the travel management department for approval. After approval, the application is transferred to the budget manager, and then to the supervisor; if the budget manager and supervisor are the same person, only one approval step needs to be completed; in some cases, the application also needs to be approved by the director).

[0052] S2. Collect business process event logs, as shown in Table 1. Event e is a specific action that occurs during process execution, containing attributes such as case number, event number, activity name, and timestamp, denoted as... ,in It is the case identifier of the event. It is the identifier of the event. It is the name of the event. These four attributes—timestamp of the event execution, and event duration—constitute one event. The required attributes are listed; other information, such as resources and execution roles, is optional for the event. The case study is a complete process instance within a business workflow, while the trajectory... For all events in the same process instance An ordered set is denoted as ,in , The set to which the event belongs. for Length of the event log. for The set of, denoted as ,in This represents the number of event logs.

[0053] Table 1 Example of an event log

[0054]

[0055] Preprocessing the event log includes the following sub-steps:

[0056] S2.1 Extract the set of event activity names and trajectory variants from the event log, count the frequency of each trajectory variant, and remove low-frequency trajectory variants below the denoising threshold, which is set to 0.05% of the total number of trajectory variants. This denoising process can reduce the context length of the input large language model, allowing it to focus on the core trajectory variants in the event log.

[0057] S2.2, process the event log using traditional inductive mining algorithms and their variants (such as Inductive Miner, Inductive Miner Direct, Inductive Miner Infrequent), and convert the mined process model into a text-based process tree sequence representation, as shown in Table 2. The Inductive Miner method can effectively generate a high-fitting model that matches the traces in the event log. The Inductive Miner Direct method builds a model based on a direct follow-up graph, which can more intuitively reflect the sequence relationship of activities in the event log. The Inductive Miner Infrequent method considers the frequency of traces and events, and can reflect the key behavior model.

[0058] Table 2 Process tree sequence

[0059]

[0060] S2.3, obtain the business process background description, map the business process background description to an embedding vector via the text-embedding-3-large model, retrieve relevant business process information from an external knowledge base, and calculate the cosine similarity between the business process background description and the stored information. According to the cosine similarity in descending order, select the most similar text blocks and concatenate the process tree rules to inject into the prompt words as enhanced process context. The process tree rules specifically include: sequential execution (→), which specifies that the child nodes are strictly executed from left to right in sequence; exclusive selection (X), which specifies that in each process instance, exactly one of its child branches is executed; parallel execution (+), which allows all child branches to be executed concurrently; and loop structure (*), which defines a repeatable execution of a process fragment.

[0061] S3, construct structured information input, encapsulate the preprocessed information obtained in step S2 into a structured information input semantic label, which specifically includes a process background label <context>, flow rule label <rules>, event log information tag <inputs>, output example label <outputs>With task instruction tag <tasks>As shown in Table 3, in particular:

[0062] S3.1, Flow Background Tag <context>The natural language context description of the marked event log corresponding to the business process is used to assist the large language model to understand the context background of the business process.

[0063] S3.2, process rule label <rules>The process tree rule enhanced by the business process rule obtained by marking the external knowledge base and the PM4PY semantic constraint is helpful for large language models to understand and reason about the business process constraints of the entire business process.

[0064] S3.3, event log information label <inputs>For marking three categories of key features extracted from event logs, the first log feature, the basic inductive process model, the initial process tree sequence generated by the traditional process mining and variant algorithm; the second log feature; the trajectory variant and its frequency, which provides the key process trajectory and its frequency sorted in descending order of frequency extracted from the event log; the third log feature, the activity event list, containing the complete set of activity events.

[0065] S3.4, output example label <outputs>In combination with the few-shot prompt, a candidate process model set example is provided.

[0066] S3.5, task instruction label <tasks>In combination with the role prompt, the business process expert role indicates clear task instructions.

[0067] Table 3 Structured information input instance

[0068]

[0069]

[0070] S4, the structured information input obtained in the S3 step is used as a prompt word for generating an initial model generation group, the initial model generation group is composed of a plurality of agents constructed using a large language model, each agent independently generates a plurality of candidate process tree sequences as a candidate set, and the candidate process tree sequences are evaluated by a model evaluator. The model evaluator converts the text form process tree sequence into a Petri net, and evaluates the Fitness, Precision, and F-measure of the Petri net model; wherein Fitness measures the ability of the process model to replay the observed behavior in the event log, Precision measures the degree to which the process model only contains the observed behavior in the log and avoids introducing redundant or irrelevant paths in the log, and F-measure is defined as the harmonic mean of Fitness and Precision, and the higher the F-measure value, the better the model quality. The initial process model is selected from the candidate set according to the highest F-measure.

[0071] S5, an expert review group based on multi-agent is constructed, the expert review group is composed of three groups of agent teams playing different expert roles, each group contains a plurality of agents constructed using a large language model, and the three groups respectively review and propose revision opinions on the initial process model obtained in the S4 step from three dimensions of process tree operator logic correctness, model structure robustness and activity event completeness, and finally the multi-dimensional revision opinions are summarized in the order of "process tree operator-model structure robustness-activity event completeness", specifically:

[0072] S5.1, the prompt words for the process tree operator logic correctness dimension are shown in Table 4, which include the following four aspects: timing relationship verification, used to verify whether the execution order of each activity in the model conforms to the inherent logical constraints of the business logic; decision structure evaluation, used to evaluate whether the selection gateway in the model accurately maps the real decision point in the business process, judge the rationality, completeness of the branch condition setting and its coverage degree to the business scenario; parallel structure review, used to review whether the activity combination in the parallel gateway truly represents the task set that can be executed concurrently or in any order in the business; loop structure analysis, used to verify whether the loop structure accurately reflects the repeated execution link in the business process, and ensure the logical correctness of the loop entry condition and exit condition.

[0073] Table 4 Flow tree operator prompt words

[0074]

[0075] S5.2, the prompt words of the model structure robustness dimension expert group, as shown in Table 5, include the following three aspects: structure complexity optimization, used to identify unnecessary loops, mergable branches or deadlocks that cause process stagnation in the model; business compliance verification, used to compare the flow tree model with the pre-defined business rules and business standards, and give the nodes or paths in the model that exist compliance deviation; abnormal structure detection, used to analyze the execution frequency and path probability distribution of each activity in the process model based on event logs, and the large language model will infer the threshold value based on the initial process model structure and path probability distribution, and identify the low probability paths in the model whose probability is lower than the threshold value.

[0076] Table 5 Flow model structure robustness prompt words

[0077]

[0078] S5.3, the prompt words of the activity event completeness dimension expert group, as shown in Table 6, include the following two aspects: activity accuracy and integrity audit, used to verify whether all activity nodes in the process model accurately reflect the actual business tasks, identify and remove redundant activities, and check for missing business activities in the initial process model; implicit dependency mining, used to identify implicit order constraints and business dependency relationships between activity nodes that are not directly expressed by the model structure.

[0079] Table 6 Activity event completeness prompt words

[0080]

[0081] S6, according to the revision suggestions provided by the S5 step expert review group and the initial process model as the prompt words of the model optimization group, as shown in Table 7, the model optimization group is composed of several intelligent agents constructed by large language models, each intelligent agent independently optimizes the potential defects of the initial process model and generates multiple candidate flow tree sequences as candidate models, and the candidate models are evaluated by a model evaluator, and the final process model is selected from the candidate models according to the highest F-measure.

[0082] Table 7 Example of expert review revision suggestions

[0083]

[0084]

[0085] S7, calculate the evaluation index of the final process model, if the evaluation index reaches the preset quality threshold requirement, output the final process model; otherwise, repeat steps S4-S7.

[0086] To show the advantages of the above-mentioned S1-S7 based on the multi-agent cooperation inductive process mining optimization method of the application, the following will be applied to specific examples to show its technical effect, the specific framework diagram is as shown in Figure 2

[0087] Embodiment

[0088] The steps of this embodiment are as shown in the above-mentioned S1-S7, which will not be repeated here. The implementation process and implementation results are shown as follows:

[0089] This embodiment selects eight public event logs from real business scenarios as experimental data sets. All event logs are in the standard format XES in the field of process mining.

[0090] Among them, five logs are selected from the authoritative BPI Challenge 2020 (BPIC2020) data set, which records the business travel expense reimbursement process of a university in the Netherlands within two years, including Domestic Declarations, International Declarations, Prepaid Travel Cost, Request For Payment and independent Permit Log process. In this embodiment, three public data sets widely used in the field of process mining are additionally introduced, which are KULeuvenICTManagement, HelpdeskProcess and CoSeLoGProject, to enhance the universality of the experimental results. These data sets cover processes of different complexity, size and business field, and the detailed information of the experimental data sets is shown in Table 8.

[0091] Table 8 Detailed information of experimental data sets

[0092]

[0093] ​The method of the present application is compared with several other mining methods in this embodiment, and the prediction methods used for comparison are: (1) Inductive Miner (IM): an inductive mining algorithm, and the mining result is a process tree; (2) Inductive Miner-infrequent (IMf): an inductive mining algorithm based on frequency filtering, and the mining result is a process tree; (3) Optimized Process Discovery based on Inductive Miner (IMopt): an inductive mining algorithm based on local process structure optimization, and the mining result is a process tree; (4) Inductive miner maximal order petri metric (IMpc): a mining algorithm based on maximum activity partitioning of partial order, and the mining result is a partial order workflow language; (5) Inductive miner dynamic clustering (IMdc): a mining algorithm based on dynamic filtering and clustering mechanism, and the mining result is a partial order workflow language; (6) LLM-based Business Process Optimization Method For Inductive Miner (LLM-IM): an inductive process mining optimization method based on multi-agent collaboration, i.e. the method of the present application, and the mining result is a process tree.

[0094] In this embodiment, the agents in the initial model generation, expert review and model optimization parts are all based on the DeepSeek-V3 model, and the number of agents in each group is set to 5, and in addition, the number of candidate sets in the initial model generation and model optimization generation is set to 10.

[0095] It should be noted that, in order to comprehensively evaluate the mined process model, Fitness, Precision and F-measure are selected as model quality indicators, and , , are introduced to evaluate the complexity of the model structure.

[0096] The ability of the process model to replay the observed behavior in the event log is measured, the degree to which the process model only contains the observed behavior in the log, avoiding the introduction of redundant or irrelevant paths in the log. is defined as the harmonic mean of and , the higher the value, the better the model quality.

[0097] set up It is a source location of ,in The set representing the treasury, A set representing change, Represents the flow relationship. It is an auxiliary function. For any ,here Indicates change The collection of successor libraries, Indicates from the warehouse The set of transitions from which to begin. Then the extended Kaldorso metric. The definition is as follows:

[0098]

[0099] set up yes The corresponding reachability graph has , From the source location A set of identifiers that can be reached through several transitions, and , It is the set of edges caused by transitions between markers in the reachable graph. Then the extended cycle metric is... The definition is as follows:

[0100]

[0101] in It is a picture Number of middle sides, It is a picture The number of mid-vertexes, It is a picture The number of strongly connected components. To measure the structural complexity of a business process model, consider the types of nodes, edges, and gateways. To measure the cyclic complexity of a business process model, focus on the branching and looping structures of the process. Model complexity metrics. Defined as and The harmonic mean, The lower the value, the more streamlined the process model.

[0102] In this embodiment, "Fit" in the table represents "Prec" represents "F" represents "ECAM" stands for "ECYM" stands for "Com" represents , and "-" represents that the relevant indicators of the model cannot be quantified due to excessive model complexity, and the optimal value of each model evaluation indicator of the corresponding event log is marked in bold.

[0103] As shown in Table 9, the model quality results of the method of the application on each data set compared with the comparative methods, the Precision and F-measure of the process model obtained by the method of the application are better than all the comparative methods (taking F-measure as the core evaluation indicator), the process model obtained by the method of the application has good behavior coverage completeness, and the Fitness is stably higher than 0.85.

[0104] Among them, the IM, IMpc and IMdc methods have higher Fitness of the process model due to excessive generalization ability, but introduce a large number of unobserved behavior paths, which reduces the Precision of the process model, and further seriously affects the performance of F-measure. The IMf method reduces the low-frequency activity behavior in the process model by filtering method, and only retains the high-frequency activity behavior. Even if the Fitness of the process model has a slight decrease, the Precision of the process model is significantly improved compared with IM, IMpc and IMdc, so that the F-measure can be improved. The IMopt method is based on IMf, and further optimizes the inaccurate structure in IMf to obtain a process model with higher Precision, but in complex business process event logs such as DD, PL and RFP logs, the performance is poor, and there is still a large optimization space.

[0105] The method of the application further improves the mining effect of the process model by preprocessing the event log and combining the semantic understanding of the business process context of the large language model and the multi-agent collaboration mechanism, which shows that the method of the application can mine a process model with higher quality.

[0106] Table 9 Comparison of process model quality results mined by each method

[0107]

[0108] As shown in Table 10, the ECAM, ECYM and Complexity of the process model obtained by the method of the application are better than those of other baseline methods in the whole, which shows that the structure of the mined model effectively suppresses redundant paths, makes the overall structure more concise and easy to understand, and better reflects the core behavior model in the actual business process.

[0109] The IM, IMpc and IMdc methods have a large number of redundant behavior paths due to overgeneralization, resulting in a significantly higher complexity of the process model than the method of the application. The IMf method effectively reduces unnecessary activity paths in the process model through a path filtering mechanism, resulting in a lower complexity of the process model compared to IM, IMpc and IMdc. The IMopt method further optimizes the inaccurate structure in the process model based on IMf to obtain a more concise process model, but in complex business process event logs such as PL, it still shows high complexity in mining the process model.

[0110] The method of the application filters out low-frequency and abnormal paths by preprocessing complex event log information, and then constructs structured information input by means of preprocessed event logs, so that the large language model can efficiently understand high-frequency process trajectories and process models, and further improve the quality of the process model and simplify the complexity of the process model in combination with the expert review group reviewing the model structure and optimizing it in collaboration.

[0111] Table 10 Comparison of process model complexity results mined by each method

[0112]

[0113] In addition, the present embodiment also carries out ablation research to verify the effectiveness of each step of the method of the application. Each ablation variant experiment is named as follows:

[0114] LLM-IM w / o Review & Optimize represents that the expert review and model optimization steps are not included, and only preprocessing and initial model generation are used.

[0115] LLM-IM w / o Review represents that only the expert review is removed, and only the preprocessing, initial model generation, model optimization and other steps are used.

[0116] The present embodiment carries out experiments on different ablation variants, and the performance on multiple data sets is shown in Table 11.

[0117] The results of LLM-IM w / o Review & Optimize show that the initial model generation has a F-measure higher than 0.8 on each event log, and can generate an initial process model with relatively high quality, providing a good base point for subsequent expert review and model optimization.

[0118] Comparing the experimental data of LLM-IM w / o Review & Optimize and LLM-IM w / o Review, it can be seen that the process model presents improvement in the three indicators of Fitness, Precision and F-measure, especially on the KIM, RFP, HP and PTC event logs, the F-measure is more significantly improved. This shows that the model optimization step can effectively correct the incorrect constraint relationship in the process model and eliminate the redundant process structure, thereby improving the quality of the final process model.

[0119] Comparing the experimental data of LLM-IM w / o Review and LLM-IM, it can also be observed that the process model presents overall improvement in the three indicators of Fitness, Precision and F-measure, and the F-measure is significantly improved in the DD, HP and CP event logs. After removing the expert review step, there is no unreasonable structure in the initial model to correct, which leads to a decline in the quality of the process model.

[0120] Comparing different ablation experiments, it is confirmed that the expert review can effectively analyze and annotate the unreasonable structure of the process model from multiple dimensions, which can effectively improve the quality of the final process model. This ablation study shows the importance of the synergy of expert review and model optimization, which can effectively simplify and correct the unreasonable structure of the process model and supplement the missing path, thereby enhancing the integrity and accuracy of the process model.

[0121] Table 11 Comparison of process model quality mining results after ablation of different steps

[0122]

[0123] In addition, it should be noted that the inductive process mining optimization method based on multi-agent collaboration in the above embodiment can be essentially executed by a computer program. Based on the same method concept as described above, the present embodiment further provides an inductive process mining system based on multi-agent collaboration, as shown in Figure 3 which comprises:

[0124] An external knowledge base module for building a vector database to store the business process documents provided by the business personnel in the form of vectors, which are described in natural language and describe the business execution process;

[0125] A preprocessing module for extracting an inductive process model from the event log according to the inductive mining algorithm and its variants; extracting all trajectory variants, trajectory frequency and each activity event in the event log, and filtering low-frequency behavior trajectories; retrieving high-similarity business process rules from the external knowledge base through the business process background description to form an enhanced business process context;

[0126] A structured information input module encapsulates the information obtained by the preprocessing module into a semantic label to obtain structured information input;

[0127] An initial model generation module constructs an initial model generation group composed of a plurality of agents, takes the structured information input obtained by the preprocessing module as a prompt word of the initial model generation group, and each agent independently generates a plurality of candidate process tree sequences as a candidate set, and evaluates the candidate process tree sequences through a model evaluator, and selects an initial process model from the candidate set according to the highest F-measure.

[0128] An expert review module constructs three teams of agents playing different expert roles, each team including a plurality of agents constructed by a large language model, and the three teams respectively review and provide revision opinions for the initial process model from three dimensions of logical correctness of process tree operator, model structure robustness, and activity event completeness, and finally, according to the hierarchical order of "process tree operator-model structure robustness-activity event completeness", the multi-dimensional revision opinions are summarized.

[0129] A model optimization module constructs a model optimization group composed of a plurality of agents, takes the multi-dimensional revision opinions provided by the expert review group as a prompt word of the model optimization group, each agent independently optimizes potential defects in the initial process model and generates a plurality of candidate process tree sequences as candidate models, and evaluates the candidates through a model evaluator, and selects a final process model from the candidate models according to the highest F-measure.

[0130] A model verification module calculates the evaluation indicators of the final process model, detects whether the evaluation indicators meet the preset quality threshold requirements, and if the threshold is higher, outputs the final process model, otherwise needs to return to the initial model generation module for reiteration.

[0131] Based on the same method concept as above, the embodiment also provides a computer readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the above inductive mining method.

[0132] The above-described embodiments are only a preferred scheme of the present application, and are not intended to limit the present application. Any modification, equivalent replacement, improvement, and technical solution obtained by modifying the present application within the spirit and principles of the present application should be included in the protection scope of the present application.< / tasks> < / outputs> < / inputs> < / rules> < / context> < / tasks> < / outputs> < / inputs> < / rules> < / context> < / tasks> < / outputs> < / inputs> < / rules> < / context>

Claims

1. An inductive process mining and optimization method based on multi-agent collaboration, characterized in that, Includes the following steps: S1. Construct an external knowledge base; S2. Collect business process event logs and business process background descriptions, and preprocess the event logs based on the external knowledge base; S3. Encapsulate the preprocessed information within structured information input semantic tags to construct structured information input. The structured information input semantic tags specifically include process background tags, process rule tags, event log information tags, output example tags, and task instruction tags. S4. The structured information input is used as the prompt words for the initial model generation group. The initial model generation group consists of several agents. Each agent independently generates several candidate process tree sequences as a candidate set. The candidate set is evaluated by the model evaluator. The initial process model is selected from the candidate set based on the highest F-measure. S5. Construct an expert review group, which consists of three teams of intelligent agents playing different expert roles. Each team contains several intelligent agents. The three teams review the initial process model and propose revisions from three dimensions: the correctness of the process tree operator logic, the robustness of the model structure, and the completeness of the activity events. S6. Based on the revision opinions provided by the expert review group and the prompts for building the model optimization group of the initial process model, the model optimization group consists of several agents. Each agent independently optimizes the potential defects of the initial process model and generates multiple candidate process tree sequences as candidate models. The candidate models are evaluated by the model evaluator, and the final process model is selected from the candidate models based on the highest F-measure. S7. Calculate the evaluation index of the final process model. If the evaluation index meets the preset quality threshold requirements, output the final process model. Otherwise, repeat steps S4 to S7.

2. The inductive process mining and optimization method based on multi-agent cooperation as described in claim 1, characterized in that, In step S1, the external knowledge base stores business process documents provided by business personnel in vector form. The business process documents include business execution processes described in natural language.

3. The inductive process mining and optimization method based on multi-agent collaboration according to claim 1, characterized in that, In step S2, the preprocessing specifically includes: Extract active event entities and trajectory variants from the event log; count the frequency of each trajectory variant, and remove low-frequency trajectory variants below the denoising threshold, which is set to 0.05% of the total number of trajectory variants; Event logs are processed using inductive mining and its variants, and the mined process models are uniformly converted into text-based process tree sequence representations; the inductive mining and its variants include: Inductive Miner, Inductive Miner Direct, and Inductive Miner Infrequent; External knowledge bases are retrieved, and the business process background description is mapped into embedding vectors using the text-embedding-3-large model. After reordering based on cosine similarity, the most similar text blocks are selected and concatenated with process tree rules, and prompt words are injected as the enhanced process context. The process tree rules specifically include: sequential execution (→), which stipulates that child nodes are executed strictly in left-to-right order; exclusive selection (X), which stipulates that one sub-branch is executed precisely in each process instance; parallel execution (+), which allows all sub-branches to be executed concurrently; and loop structure (*), which defines process segments that can be executed repeatedly.

4. The inductive process mining and optimization method based on multi-agent cooperation as described in claim 1, characterized in that, In step S3, the structured information is input into the semantic tags. The process background label <context> Used to mark the business process background description corresponding to the event log;< / context> The process rule label <rules> Process tree rules are used to label and retrieve business process mining knowledge enhanced by external knowledge bases, and to constrain PM4PY semantics.< / rules> The event log information label <inputs> Three key features extracted from the event log were used to label the first log feature, which is the initial sequence of process trees generated by inductive process mining and its variant algorithms; the second log feature;< / inputs> Trajectory variants and their frequencies provide process trajectories and their frequencies extracted from the event log, sorted in descending order of frequency; the third log feature is the list of active events, which contains the complete set of active events. The output example label <outputs> Based on the few-sample hints, an example of the set of candidate process models to be generated is provided;< / outputs> The task instruction label <tasks> Combine role prompts with clear task instructions from the perspective of a business process expert.< / tasks> 5. The inductive process mining and optimization method based on multi-agent cooperation according to claim 1, characterized in that, In step S4, the model evaluator is used to convert the process tree sequence in text form into a Petri net model and evaluate the fitness, precision, and F-measure of the Petri net model. Fitness measures the ability of the process tree sequence to replay observed behaviors in the event log, precision measures the ability of the process tree sequence to avoid introducing redundant or irrelevant paths into the process tree sequence, and F-measure is defined as the harmonic mean of fitness and precision. The higher the F-measure value, the better the model quality.

6. The inductive process mining and optimization method based on multi-agent cooperation according to claim 1, characterized in that, In step S5, the prompts for the three expert review groups include: The expert group's prompts for the logical correctness dimension of the process tree operator include the following four aspects: Sequence verification, used to check whether the execution order of various activities in the model conforms to the inherent logical constraints of the business logic, and to output activities whose business logic does not conform to the inherent logical constraints; Decision structure evaluation, used to evaluate whether the selection gateway in the model accurately maps the real decision points in the business process, and to judge the rationality, completeness, and coverage of the business scenario of the branch condition settings; Parallel structure review, used to identify activity combinations within the parallel gateway that do not truly represent tasks that can be executed concurrently in the business; and Loop structure analysis, used to verify whether the loop structure accurately reflects the links that need to be repeatedly executed in the business process, and to ensure the logical correctness of the loop entry and exit conditions. The expert group's prompts for the model structural robustness dimension include the following three aspects: structural complexity optimization, used to identify unnecessary loops, mergeable branches, or deadlocks that cause process stagnation in the model; business compliance verification, used to compare the process tree sequence with predefined business rules and identify nodes or paths in the model that have compliance deviations; and abnormal structure detection, used to analyze the execution frequency and probability distribution of each activity and path in the process model based on event logs, instructing the large language model to infer a threshold based on the probability distribution of the initial process model structure and paths, and identify low-probability paths in the process model with probabilities below the threshold. The expert group's prompts for the activity event completeness dimension include the following two aspects: activity accuracy and completeness audit, which is used to verify whether all activity nodes in the process model accurately reflect the actual business tasks, identify and remove redundant activities, and check for missing business activities in the initial process model; and implicit dependency mining, which is used to identify implicit order constraints and business dependencies between activity nodes that are not directly expressed by the model structure.

7. The inductive process mining and optimization method based on multi-agent cooperation according to claim 1, characterized in that, In step S5, the revision opinions from multiple dimensions are summarized in the order of process tree operators, model structure robustness, and activity event completeness.

8. The inductive process mining and optimization method based on multi-agent cooperation according to claim 1, characterized in that, In step S1, the external knowledge base is constructed based on the LanceDB vector database. The retrieval granularity of the database is set to the text block level, the text block size is set to 1024, and the text block overlap window is set to 100. In step S7, the evaluation metrics include: Fitness, Precision, F-measure, and complexity metrics, wherein the complexity metrics include: 、 、 ； Fitness measures the ability of a process model to replay observed behaviors in event logs; Precision measures the ability of a process tree sequence to avoid introducing redundant or irrelevant paths into the process tree sequence; F-measure is... and The harmonic mean of the model; a higher F-measure value indicates a better model quality. Used to measure the structural complexity of a business process model based on node, edge, and gateway types. Used to measure the cyclic complexity of a business process model, focusing on the branching and looping structure of the process; model complexity metric. for and Harmonic mean The lower the value, the more streamlined the process model.

9. An inductive process mining system based on multi-agent collaboration, characterized in that, include: The external knowledge base module is used to: build an external knowledge base to store business process documents provided by business personnel in vector form, wherein the business process documents include business execution processes described in natural language; The preprocessing module is used to: collect business process event logs and business process background descriptions, and preprocess the event logs based on an external knowledge base; The structured information input module is used to: encapsulate preprocessed information within structured information input semantic tags to construct structured information input. The structured information input semantic tags specifically include process background tags, process rule tags, event log information tags, output example tags, and task instruction tags. The initial model generation module is used to: take structured information input as prompt words for the initial model generation group. The initial model generation group consists of several agents built using a large language model. Each agent independently generates several candidate process tree sequences as a candidate set. The candidate process tree sequences are evaluated by the model evaluator, and the initial process model is selected from the candidate set based on the highest F-measure. The expert review module is used to: construct a multi-agent-based expert review group, which consists of three teams of agents playing different expert roles. Each team contains several agents constructed using a large language model. The three teams review the initial process model and propose revision opinions from three dimensions: the logical correctness of the process tree operator, the robustness of the model structure, and the completeness of the activity events. Finally, the revision opinions from multiple dimensions are summarized in the hierarchical order of "process tree operator - model structure robustness - activity event completeness". The model optimization module is used to: construct prompts for the model optimization group based on the revision opinions provided by the expert review group and the initial process model. The model optimization group consists of agents constructed from several large language models. Each agent independently optimizes the potential defects of the initial process model and generates multiple candidate process tree sequences as candidate models. The candidate models are evaluated by the model evaluator, and the final process model is selected from the candidate models based on the highest F-measure. The model validation module is used to calculate the evaluation index of the final process model. If the evaluation index meets the preset quality threshold requirements, the final process model is output. Otherwise, proceed with the iteration.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the inductive process mining optimization method based on multi-agent cooperation as described in any one of claims 1 to 8.

Images