An effective low-frequency trajectory identification method based on multi-view compliance rules

Abstract

This application relates to the field of process mining, and specifically to a method for identifying effective low-frequency trajectories based on multi-perspective compliance rules. The method includes: acquiring raw event logs and dividing them into high-frequency and low-frequency event logs; evaluating control flow deviation, organizational deviation, and time deviation based on the high-frequency event logs; weighting and fusing the control flow compliance score, organizational compliance score, and time compliance score; and identifying effective low-frequency logs from the low-frequency event logs based on the fused score. This application addresses the problem of low identification accuracy of effective low-frequency trajectories in complex business scenarios using traditional data mining methods by introducing a multi-perspective joint evaluation of control flow deviation, organizational deviation, and time deviation.

Description

Technical Field

[0001] This application relates to the field of process mining, specifically to an effective low-frequency trajectory recognition method based on multi-perspective compliance rules. Background Technology

[0002] Process mining, a technology for process modeling and analysis based on event logs, has been widely applied in fields such as healthcare, manufacturing, finance, and government, helping to achieve automated process modeling, execution monitoring, and continuous improvement. As business processes become more complex, event logs contain not only frequently occurring routine trajectories but also a large number of low-frequency trajectories with low probability of occurrence. These low-frequency trajectories may involve special paths, abnormal situations, or execution steps under specific conditions within the process, such as "special treatment for rare diseases" in healthcare processes or "large-amount transfers" in financial scenarios. In practical applications, low-frequency trajectories are often treated as anomalies or noise and directly filtered out before modeling using frequency thresholds. While this strategy reduces model complexity, it may mistakenly delete process variations with business significance.

[0003] In existing technologies, traditional low-frequency trajectory identification relies solely on frequency for filtering, which has significant shortcomings: due to the lack of fixed judgment criteria, the methods often fail to identify low-frequency trajectories with specific value, such as special processes or abnormal compliance behaviors. For example, CN109190877A discloses an analysis method for effective low-frequency patterns based on Petri net behavior density. This scheme calculates behavior density using the minimum k-order inheritance relationship and behavior distance vector, and then uses a behavior density threshold for filtering to obtain denoised effective low-frequency trajectories. However, since it does not involve business data, this scheme cannot identify high-value low-frequency patterns that are special but structurally similar. Furthermore, the classification of simply considering structural similarity as valuable and deviation as noise is too simplistic and cannot effectively distinguish between effective low-frequency patterns and noise, easily missing key low-frequency behaviors.

[0004] Therefore, there is an urgent need for a more refined low-frequency trajectory recognition method to identify and retain low-frequency trajectories with higher value in the process. Summary of the Invention

[0005] In view of this, this application discloses an effective low-frequency trajectory recognition method based on multi-view compliance rules to solve the problem of low precision in low-frequency trajectory recognition in the existing process mining technology field, including:

[0006] S1. Obtain the raw event log and divide the raw event log into high-frequency event logs L. high and low-frequency event log L low ;

[0007] S2. Extract the sequential relationships and dependency strengths between activities from high-frequency event logs to construct compliance rules from a control flow perspective; evaluate the control flow deviation of trajectories in low-frequency event logs based on the control flow compliance rule set to obtain a control flow compliance score.

[0008] S3. Extract high-frequency resource-activity pairs from high-frequency event logs, construct compliance rules from an organizational perspective, and evaluate organizational deviations in the trajectories of low-frequency event logs to obtain organizational compliance scores.

[0009] S4. Extract the correlation time difference between activity pairs from the high-frequency event logs, calculate the reasonable time interval for each activity pair, construct compliance rules from a time perspective, and evaluate the time deviation of the trajectories in the low-frequency event logs to obtain a time compliance score.

[0010] S5. Perform weighted fusion of control flow compliance score, organizational compliance score, and time compliance score; identify valid low-frequency logs in the low-frequency event logs based on the fusion score.

[0011] The beneficial effects of this application include:

[0012] This paper presents an effective low-frequency trajectory recognition method based on multi-perspective compliance rules. It addresses the limitation of existing research, which focuses on a single perspective and cannot comprehensively depict the overall business situation. By introducing a multi-perspective joint evaluation of control flow deviation, organizational deviation and time deviation, it makes up for the shortcomings of traditional methods in complex business scenarios and achieves more three-dimensional and accurate recognition of low-frequency trajectories.

[0013] To address the pain point that traditional methods are difficult to adapt to complex data distributions, a variety of dynamic screening rules are introduced in the multi-perspective evaluation. These rules can flexibly adjust the judgment boundary according to the data distribution characteristics, breaking the limitations of a single fixed standard and significantly enhancing the method's adaptability to processing multi-source heterogeneous data.

[0014] To address the differences in features across various business processes, a flexible weighting strategy was developed. This strategy optimizes the weights of control flow, organization, and time based on specific scenario requirements, significantly improving the overall discrimination capability of low-frequency trajectory recognition. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of an effective low-frequency trajectory recognition method based on multi-view compliance rules in an embodiment of this application;

[0016] Figure 2 This is a schematic diagram showing the sequence number and frequency of event log L in the embodiments of this application;

[0017] Figure 3 This is a schematic diagram of the medical treatment process in the embodiments of this application;

[0018] Figure 4 This is a schematic diagram of the clustering results of different clusters in the embodiments of this application;

[0019] Figure 5 This is a schematic diagram showing the comparison of F-measure values ​​tested in the embodiments of this application;

[0020] Figure 6 This is a schematic diagram showing the comparison of generalization tested in the embodiments of this application;

[0021] Figure 7 This is a schematic diagram showing the comparison of the simplicity of the tests in the embodiments of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, features, and advantages of this application clearer and to facilitate a better understanding of the technical solutions of this application by those skilled in the art, the following detailed description of this application is provided in conjunction with the accompanying drawings and embodiments.

[0023] This embodiment includes an effective low-frequency trajectory recognition method based on multi-view compliance rules, such as... Figure 1 As shown, it includes:

[0024] S1. Obtain the raw event log. Based on the frequency of occurrence of trajectory variants in the log, divide the raw event log into high-frequency event logs and low-frequency event logs, including:

[0025] S11. Obtain the original event log L; define the standard event trajectory process, and extract the trajectory variant V and trajectory variant frequency X from the event log L. The following will describe the low-frequency trajectory recognition method designed in this application in conjunction with the acute cardiac disease treatment process. The original event log L for acute cardiac disease treatment is obtained through the hospital information system. The sequence number and frequency of the event log L are as follows: Figure 2 As shown, Figure 3 The event log L displays the medical treatment process activities represented by each letter.

[0026] All trajectory variants in event log L constitute the trajectory variant set V={v1,v2,…,v n}; Statistical analysis of each trajectory variant v i Frequency c in the original event log L i The frequencies of all trajectory variants are then transformed into a set of feature vectors X={x1,x2,…,x...} n}, where x i =[c i [ ] is a one-dimensional feature vector representing the frequency of trajectory variants. Trajectory variants For example, then construct The feature vector x1 = [1 2 5 6] is used, and other trajectory variants are processed in the same way.

[0027] S12. The k-means clustering algorithm is used to cluster the trajectory variant frequency X by minimizing the sum of squared errors within the cluster; the centroid values ​​of each cluster are obtained.

[0028] The objective function of the k-means clustering algorithm is expressed as:

[0029]

[0030] Where k represents the number of preset clusters, C j Let x represent the j-th cluster. i Represents the feature vector within a cluster. j This represents the centroid value of the j-th cluster. The centroid values ​​of each cluster are obtained through iterative clustering calculations.

[0031] S13. The cluster with the largest centroid value is identified as the high-frequency cluster, and the set of trajectory variants contained in the high-frequency cluster is labeled as the high-frequency variant set V. high Set a centroid value threshold; clusters with values ​​below the threshold are classified as low-frequency clusters, and the set of trajectory variants contained within them is labeled as the low-frequency variant set V. low .

[0032] S14. Based on the high-frequency variant set V high With low-frequency variant set V low Filter the raw event log L to match the high-frequency variant set V. high The trajectory output is a high-frequency event log L high This will match the low-frequency variant set V low The trajectory output is a low-frequency event log L low .

[0033] In this embodiment, the number of clusters is set to 2. The input feature vector set X={

[1256] ,

[1164] ,

[1002] ,

[989] ,

[50] ,

[32] ,

[28] ,

[24] ,

[20] ,

[16] ,

[15] ,

[12] }. The frequencies of the two clusters are as follows: Figure 4 As shown, it will contain trajectory variants { , , , The cluster denoted by} is the high-frequency variant set V. high , will contain trajectory variants { , , , , , , , The cluster denoted as the low-frequency variant set V is denoted as} low .

[0034] S2. Extract the sequence relationships and dependency strengths between activities from high-frequency event logs to construct compliance rules from a control flow perspective; evaluate the control flow deviation of trajectories in low-frequency event logs based on the control flow compliance rule set to obtain a control flow compliance score; including:

[0035] S21. Extract L high All included activities constitute the activity set; iterate through L. high The total trajectory, and the direct following frequency of each pair of adjacent activities (a, b) in the statistical activity set. Construct the direct following relationship matrix M dep The formula for directly following frequency is:

[0036]

[0037] in, This represents the number of times that a is directly followed by b in L, and is a measure of the dependency relationship between a and b, where a and b represent two activities in the activity set. Indicates direct following frequency. Represents a trajectory sequence. It is a sequence number. The expression represents a logical AND, where L represents the current event log, i.e., the high-frequency event log; the constructed M... dep The rows and columns correspond to the activities in the activity set.

[0038] S22. Based on the direct following relation matrix M dep Calculate the dependency measure between activity pairs; the formula is:

[0039]

[0040] in, This represents the dependency measure of activity a on activity b.

[0041] S23. Iterate through the dependency metrics of all activity pairs and construct the causal dependency probability matrix M. causal .

[0042] The causal dependency probability matrix M constructed in this embodiment causal As shown below:

[0043]

[0044] S24. Set the direct follow frequency threshold Th dep and the causal dependency probability threshold Th causal ; Traverse directly following the relation matrix Mdep and the causal dependency probability matrix M causal Filter out those that simultaneously meet the direct following frequency requirement >Th dep And dependent on metrics >Th causal The activity pair (a', b') is considered a strongly correlated activity pair with a clear unidirectional dependency and is included in the control flow compliance rule set R. flow .

[0045] S25. Traverse the low-frequency event log L low For each low-frequency trajectory, extract the activity sequence of the trajectory and calculate the total number N of adjacent activity pairs contained in the sequence. total ;Assign adjacent activity pairs to the control flow compliance rule set R flow Compare each one individually, and count the number N of matching active pairs included in the rule set. match ; Calculate the ratio of the number of matching activity pairs to the total number of adjacent activity pairs (Score). flow = This serves as a compliance score for the control flow of current low-frequency trajectories.

[0046] The adjacent activity pair refers to a combination of two activities that occur in sequence. This is consistent with the description of each adjacent activity pair in S21, which states "the direct following frequency of each pair of adjacent activities (a,b) in the statistical activity set".

[0047] S3. Extract high-frequency resource-activity pairs from high-frequency event logs, construct compliance rules from an organizational perspective, and assess organizational deviations in the trajectories of low-frequency event logs to obtain an organizational compliance score; including:

[0048] S31. Extract high-frequency event logs L high All execution resources contained therein constitute the global resource set; for L high For any activity 'a' in the global resource set and any execution resource 'r' in the global resource set, count their performance in L. high The common frequency C(a,r) of all trajectory events is used to construct the resource-activity frequency matrix M from an organizational perspective. org ;

[0049] S32. For any activity a, obtain the set R of all its execution resources. a The resources in the set are sorted according to their frequency C(a,r) to obtain the resource sequence L. r Computing resources r i The execution percentage P(a, r) i ); Set the Pareto cumulative threshold Th pareto Determine whether it meets the requirements. The smallest positive integer m, let {r1,…, r m} constitute the initial core resource set, and denoted as r m For boundary resources.

[0050] In this embodiment, the sorted resource sequence L r ={r1,r2,…,r i}, satisfying C(a, r1) C(a, r2) … C(a,r i Computing resources r i The execution percentage formula is P(a, r) i )= , i represents the sorted resource sequence L r The total number of elements.

[0051] S33. Set the peer tolerance threshold Th peer Based on boundary resources r m The execution percentage P(a,r) m Trigger peer rules to perform similarity resource clustering checks, and obtain a compliant resource-activity pair rule set R. org .

[0052] The similarity resource clustering check includes: sequentially verifying resource sequence L r For the remaining resources in index j>m, if the percentage difference P(a,r) is satisfied... m )-P(a,r j ) Th peer Then the edge resource r j Add to the initial core resource set; after traversing all activities, obtain the final determined core resource set Core(a) for each activity, thus obtaining the compliant resource-activity pair rule set R. org ={(a, r)|a A, r Core(a)}, A represents the high-frequency event log L high The total set of all unique activity names appearing in the list, i.e., the complete set of activities.

[0053] S34. Traverse the low-frequency event log L low Each low-frequency trajectory is compared with the rule set R. org Compare the data and calculate the organizational compliance score for low-frequency trajectories.

[0054] Calculating the organizational compliance score for low-frequency trajectories includes: extracting the total number N of events containing activity and resource attributes. total ; Extract the event resource pairs from the trajectory (ai ,r i ) and rule set R org Perform a comparison and count the number N of matching event resource pairs contained in the rule set. match ;Use the formula Score org = The organizational compliance score for this low-frequency trajectory was calculated.

[0055] S4. Extract the correlation time difference between activity pairs from high-frequency event logs, calculate the reasonable time interval for each activity pair, construct compliance rules from a time perspective, and evaluate the time deviation of trajectories in low-frequency event logs to obtain a time compliance score; including:

[0056] S41. Traverse L high The trajectory is used to arrange events within the trajectory in chronological order by timestamp; the rules belonging to the control flow compliance set R are extracted. flow For adjacent activity pairs p=(a, b), calculate the timestamp difference between activity b and activity a, and use it as L. high time difference of trajectory association ;statistics Average time difference p with standard deviation p Construct a time feature rule set R based on the control flow compliance rule set. time ;

[0057] S42. Traverse L low The trajectory is used to arrange the events within the trajectory in chronological order by timestamp; all adjacent activity pairs within the trajectory are traversed, and L is calculated. low time difference of trajectory association ;

[0058] S43. Calculate L low Temporal deviation of adjacent activity pairs in the trajectory (Dev) p The formula is:

[0059]

[0060] in, This represents the average time difference between activities and p. This represents the preset fixed time deviation penalty value.

[0061] S44. L low Temporal deviation Dev of all adjacent activity pairs in the trajectory p The arithmetic mean of the values ​​is used as the time compliance score for the current low-frequency trajectory.

[0062] S5. Weighted fusion of control flow compliance score, organizational compliance score, and time compliance score; identification of valid low-frequency logs in the low-frequency event logs based on the fusion score; including:

[0063] S51. Obtain each low-frequency trajectory from the low-frequency event log. Control flow compliance score flow Organizational compliance score org Time compliance score time Since the time compliance score is based on time deviation, before weighted fusion, minimax normalization and inverse mapping are performed to obtain a positively normalized time score. * time =1 - Normalize(Score) time To ensure that the scores from each perspective fall within the [0,1] range and are positively correlated with the level of compliance; the weighting coefficients for control flow, organization, and time perspectives are set as w, respectively. flow w org and w time The low-frequency trajectory was obtained by weighted summation. Weighted fusion score final The calculation formula is:

[0064]

[0065] S52. Set validity determination threshold Traversing all low-frequency trajectories after evaluation will satisfy the weighted fusion score. final Low-frequency trajectories are identified as valid low-frequency trajectories containing business value, retained, and stored in the valid low-frequency log Lvalid; the Score is... final < The low-frequency trajectory was identified as abnormal noise and filtered out.

[0066] Furthermore, effective low-frequency logs and high-frequency event logs are merged to construct a complete business process model. In this embodiment, the high-frequency event log L... high Compared with the filtered effective low-frequency log L valid Perform a set union operation to merge and generate a complete fusion event log L. merged =L high L valid The fusion event log L merged As input, the inductive mining algorithm is used to discover processes and generate a final business process model that retains effective low-frequency variants while filtering out structural noise.

[0067] Furthermore, to verify the effectiveness of the method proposed in this invention, the effective low-frequency trajectory recognition method in this embodiment was tested, and the test metrics were based on the following three model quality metrics:

[0068] 1. F-measure value; The F-measure is the harmonic mean of fitness and precision, defined as:

[0069]

[0070] Fitness is used to evaluate the extent to which the generated process model can reproduce the behavior in the event log. A high fitness score indicates that the model can cover more of the actual events. It is defined as:

[0071]

[0072] in, This indicates the number of tokens that needed to be manually added during the model replay process due to insufficient tokens in the logs. This indicates the number of tokens consumed during the log replay process. This represents the number of unconsumed tokens that remain in the various libraries of the model during the model replay process. This indicates the number of tokens generated during the model replay process.

[0073] Precision measures the extent to which a model avoids overfitting, i.e., it doesn't produce additional behavior not present in the logs. Higher precision means the model accurately depicts log behavior. It is defined as:

[0074]

[0075] Where S represents the state generated by the log track prefix. The weights of state s are represented by their values. This represents the set of subsequent activities observed after log state s. This represents the set of all possible subsequent activities at model state s.

[0076] 2. Generalization: This evaluates the extent to which a model maintains its predictive ability on unseen trajectories. Higher generalization means the model is better able to adapt to plausible behaviors not fully covered in the logs. It is defined as:

[0077]

[0078] Where N represents the total number of times all transitions are triggered in the event log, and M represents the total number of times a choice needs to be made at any location. This represents the number of tokens consumed in the j-th decision. This represents the number of possible choices in the j-th decision.

[0079] 3. Simplicity; used to evaluate the simplicity of a model's structure. While overly complex models may cover more behaviors, they are less conducive to understanding and practical application. It is defined as:

[0080]

[0081] Where d is the average degree of all nodes in the model, and k is the baseline degree.

[0082] During the testing process, the method designed in this application (ITI-MCR) was compared with five existing process mining methods. The first method is a heuristic mining algorithm-based method (referred to as HM); the second method is a frequency-based inductive mining method (referred to as IM-i); the third method is a data-driven heuristic mining method (referred to as DHM); the fourth method is an effective low-frequency pattern analysis method based on Petri net behavior density (referred to as LFB); and the fifth method is an effective low-frequency sequence analysis method based on the maximum probability path (referred to as MP).

[0083] The tests used publicly available datasets, including: Request For Payment, Domestic Declarations, Prepaid Travel Costs, International Declarations, and Travel Permit Data.

[0084] Figure 5The report presents the F-measure values ​​of six methods across five datasets. It shows that the ITI-MCR method proposed in this application achieves the highest F-measure values ​​on the Request For Payment, Prepaid Travel Costs, and International Declarations datasets, significantly outperforming the other five methods. This indicates that the method of this invention can effectively filter random noise while preserving low-frequency variations with business value to the greatest extent possible in such complex process scenarios, thereby reconstructing a high-fidelity business process. On the Domestic Declarations and TravelPermit Data datasets, although the HM method has a slightly higher F-measure value, the value obtained by the ITI-MCR method is extremely close to the highest value and significantly outperforms the other four methods, including IM-i, DHM, LFB, and MP. The performance of the HM method fluctuates significantly across different datasets; the F-measure value of the IM-i method performs poorly on multiple datasets. In summary, compared to the other five methods, the ITI-MCR method proposed in this invention maintains extremely high F-measure values ​​on datasets from different domains, demonstrating the strongest robustness and comprehensive performance, and effectively improving the overall quality of the final process model.

[0085] Figure 6 This section describes the generalization performance of six methods across five datasets. Generalization reflects the adaptability and predictive ability of the discovered process model to new trajectories not included in the event logs but consistent with the underlying business logic. Generally, a higher value indicates better model scalability. As shown in the figure, the ITI-MCR method proposed in this invention achieved the highest generalization on the Prepaid Travel Costs, International Declarations, and Travel Permit Data datasets. This indicates that the proposed method accurately removes low-frequency noise while effectively extracting and preserving the core skeletal features of the process, enabling the generated model to fully accommodate unknown and reasonable business process variations. On the Request For Payment dataset, the generalization performance of the ITI-MCR method is very close to the best-performing IM-i method and significantly better than the other four comparison methods: HM, DHM, LFB, and MP. On the Domestic Declarations dataset, the generalization performance of the ITI-MCR method is relatively low. However, combined with… Figure 2The F-measure value of this method reveals that other methods often fall into the trap of "overgeneralization," sacrificing model accuracy and allowing a large number of illegal or non-compliant abnormal behaviors to pass through in exchange for high generalization. The ITI-MCR method of this invention, by introducing rigorous multi-perspective compliance rule evaluation, intercepts "pseudo-compliance" noise such as unauthorized operations and time delay anomalies, thus exhibiting a more rigorous model boundary on this specific dataset. This proves that this method achieves the best balance between generalization and accuracy while ensuring the rigor of the model structure and business compliance. In summary, compared with the other five comparative methods, the ITI-MCR method of this invention not only avoids unrestrained overgeneralization but also demonstrates superior model expansion and adaptability on the vast majority of datasets.

[0086] Figure 7 This paper presents six methods for assessing the simplicity of models across five datasets. Simplicity evaluates the conciseness of the process model structure; a higher value indicates fewer redundant nodes, complex cross-connections, and incomprehensible deadlock branches, effectively avoiding the generation of messy and redundant "spaghetti" models. As shown in the figure, the ITI-MCR method proposed in this invention achieves the highest simplicity score on the Domestic Declarations dataset. Furthermore, the ITI-MCR method also demonstrates excellent simplicity performance on the International Declarations, Request For Payment, and Prepaid Travel Costs datasets, closely approaching the MP method's highest score on these datasets and outperforming conventional heuristic mining methods such as HM and DHM. This indicates that this invention, through multi-perspective compliance assessment, successfully eliminates a large amount of abnormal noise that causes model structural chaos, effectively simplifying the model's topology. In conclusion, compared to the other five comparative methods, the business process model generated by the ITI-MCR method of this invention is both clear and easy to read, and possesses extremely high practical business guidance value.

[0087] Finally, it should be noted that the above description only depicts some embodiments of this application. For those skilled in the art, various changes, modifications, substitutions, and variations can be conceived of these embodiments without departing from the principles and spirit of this application. The scope of protection of this application is defined by the appended claims and their equivalents, and all the above-mentioned behaviors should be covered within the scope of protection of this application.

Claims

1. An effective low-frequency trajectory recognition method based on multi-view compliance rules, characterized in that, include: S1. Obtain the raw event log and divide the raw event log into high-frequency event logs L. high and low-frequency event log L low ; S2. Extract the sequential relationships and dependency strengths between activities from high-frequency event logs to construct compliance rules from a control flow perspective; evaluate the control flow deviation of trajectories in low-frequency event logs based on the control flow compliance rule set to obtain a control flow compliance score. S3. Extract high-frequency resource-activity pairs from high-frequency event logs, construct compliance rules from an organizational perspective, and evaluate organizational deviations in the trajectories of low-frequency event logs to obtain organizational compliance scores. S4. Extract the correlation time difference between activity pairs from the high-frequency event logs, calculate the reasonable time interval for each activity pair, construct compliance rules from a time perspective, and evaluate the time deviation of the trajectories in the low-frequency event logs to obtain a time compliance score. S5. Perform weighted fusion of control flow compliance score, organizational compliance score, and time compliance score; identify valid low-frequency logs in the low-frequency event logs based on the fusion score.

2. The effective low-frequency trajectory recognition method based on multi-view compliance rules according to claim 1, characterized in that, The original event log is divided into high-frequency event logs L. high and low-frequency event log L low ,include: S11. Obtain the original event log L; define the standard event trajectory process, and extract the trajectory variant V and trajectory variant frequency X from the event log L; S12. The k-means clustering algorithm is used to cluster the trajectory variant frequency X by minimizing the sum of squared errors within the cluster, and the centroid values ​​of each cluster are obtained. S13. The cluster with the largest centroid value is identified as the high-frequency cluster, and the set of trajectory variants contained in the high-frequency cluster is labeled as the high-frequency variant set V. high Set a centroid value threshold; clusters with values ​​below the threshold are classified as low-frequency clusters, and the set of trajectory variants contained within them is labeled as the low-frequency variant set V. low ; S14. Based on the high-frequency variant set V high With low-frequency variant set V low Filter the raw event log L to match the high-frequency variant set V. high The trajectory output is a high-frequency event log L high This will match the low-frequency variant set V low The trajectory output is a low-frequency event log L low .

3. The effective low-frequency trajectory recognition method based on multi-view compliance rules according to claim 1, characterized in that, S2 include: S21. Extract all activities from the high-frequency event log to form an activity set; traverse all trajectories in the activity set and count the direct following frequency of each pair of adjacent activities (a, b) in the activity set. Construct the direct following relationship matrix M dep ; S22. Based on the direct following relation matrix M dep Calculate the dependency measure between activity pairs ; S23. Iterate through the dependency metrics of all activity pairs and construct the causal dependency probability matrix M. causal ; S24. Set the direct follow frequency threshold Th dep and the causal dependency probability threshold Th causal Traversing directly follows the relation matrix M dep and the causal dependency probability matrix M causal Filter out those that simultaneously meet the direct following frequency requirement >Th dep And dependent on metrics >Th causal The activity pair (a', b') is considered a strongly correlated activity pair with a clear unidirectional dependency and is included in the control flow compliance rule set R. flow ; S25. Traverse the low-frequency event log L low For each low-frequency trajectory, extract the activity sequence of the trajectory and calculate the total number N of adjacent activity pairs contained in the sequence. total ;Assign adjacent activity pairs to the control flow compliance rule set R flow Compare each one individually, and count the number N of matching active pairs included in the rule set. match ; Calculate the ratio of the number of matching activity pairs to the total number of adjacent activity pairs (Score). flow = This serves as a compliance score for the control flow of current low-frequency trajectories.

4. The effective low-frequency trajectory recognition method based on multi-view compliance rules according to claim 3, characterized in that, The direct following frequency The formula is: ； Where a and b represent two activities in the activity set, Indicates direct following frequency. Represents a trajectory sequence. It is the sequence number, and L indicates the current event log.

5. The effective low-frequency trajectory recognition method based on multi-view compliance rules according to claim 3, characterized in that, The dependency metric The formula is: 。 6. The effective low-frequency trajectory recognition method based on multi-view compliance rules according to claim 1, characterized in that, S3 include: S31. Extract high-frequency event logs L high All execution resources contained therein constitute the global resource set; for L high Given any activity 'a' and any execution resource 'r' in the global resource set, calculate the relationship between 'a' and 'r' in L. high The common frequency C(a,r) of all trajectory events is used to construct the resource-activity frequency matrix M from an organizational perspective. org ; S32. For any activity a, obtain the set R of all its execution resources. a The resources in the set are sorted according to their frequency C(a,r) to obtain the resource sequence L. r ={r1,r2,…,r i }, satisfying C(a, r1) C(a, r2) … C(a,r i Computing resources r i The execution percentage P(a, r) i ); Set the Pareto cumulative threshold Th pareto Determine whether it meets the requirements. The smallest positive integer m, let {r1,…, r m Let r be the initial core resource set. m For border resources; S33. Set the peer tolerance threshold Th peer For boundary resources r m The execution percentage P(a,r) m Perform similarity resource clustering checks to obtain a set of compliant resource-activity pair rules; The similarity resource clustering check includes: sequentially verifying resource sequence L r For the remaining resources in index j>m, if the percentage difference P(a,r) is satisfied... m )-P(a,r j ) Th peer Then the edge resource r j Add to the initial core resource set; after traversing all activities, obtain the determined core resource set Core(a), and construct the compliant resource-activity pair rule set R. org ={(a, r)|a A, r Core(a)}; S34. Traverse the low-frequency event log L low For each low-frequency trajectory, an organizational compliance score is calculated by comparing it with the resource-activity rule set.

7. The effective low-frequency trajectory recognition method based on multi-view compliance rules according to claim 6, characterized in that, The computing resources r i The execution percentage P(a, r) i The formula is: ； Where k represents the total number of elements in the sorted resource sequence r.

8. The effective low-frequency trajectory recognition method based on multi-view compliance rules according to claim 1, characterized in that, S4 include: S41. Traverse L high The trajectory is used to arrange events within the trajectory in chronological order by timestamp; the rules belonging to the control flow compliance set R are extracted. flow For adjacent activity pairs p=(a, b), calculate the timestamp difference between activity b and activity a, and use it as L. high time difference of trajectory association ;statistics Average time difference p with standard deviation p Construct a time feature rule set R based on the control flow compliance rule set. time ; S42. Traverse L low The trajectory is used to arrange the events within the trajectory in chronological order by timestamp; all adjacent activity pairs within the trajectory are traversed, and L is calculated. low time difference of trajectory association ; S43. Calculate L low Temporal deviation of adjacent activity pairs in the trajectory (Dev) p The formula is: ； in, This represents the preset fixed time deviation penalty value; S44. Dev p The arithmetic mean is used as the time compliance score for the current low-frequency trajectory.

9. The effective low-frequency trajectory recognition method based on multi-view compliance rules according to claim 1, characterized in that, The weighted fusion process involves normalizing the time compliance score using a minimax normalization and inverse mapping before weighted fusion. The formula is: Score * time =1-Normalize(Score time ).

10. The effective low-frequency trajectory recognition method based on multi-view compliance rules according to claim 1, characterized in that, After identifying the valid low-frequency logs in the low-frequency event logs, the valid low-frequency logs are merged with the high-frequency event logs to build a complete business process model.

Images