Business process decision mining method and system based on causal inference

Abstract

The invention discloses a business process decision mining method and system based on causal inference. The method comprises the following steps: acquiring an event log hidden with business process decision information and a corresponding Petri network; the event log is preprocessed, and a selection structure in the Petri network is identified; extracting an adjacent relation data set corresponding to each selection structure from the processed event log; for each adjacent relation data set, constructing a causal model, and based on the causal model, adopting a causal inference method based on meta-learning to obtain an effective causal relation and a corresponding ATE; and processing the obtained effective causal relationship, and adding the processed effective causal relationship to a Petri net to obtain a process model capable of intuitively explaining a business process decision. According to the method, the integrity, the accuracy and the efficiency of business process decision mining are improved, the decision rule is clearly explained on the process model, an enterprise is helped to adjust the process design according to a result, unnecessary steps are avoided, possible losses are reduced, and the production efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of process mining, and in particular to a business process decision mining method, system, storage medium and computing device based on causal inference. Background Art

[0002] Business process decision means that some business processes are not simply carried out in sequence. At some decision points, it may be necessary to make a choice among multiple alternative activities, and these choices will affect the direction and results of the process. At present, business process decision mining usually adopts decision tree algorithm. This kind of algorithm constructs function estimator at each decision point and mines decision rules. Each decision condition is a set of logical expressions. When these conditions are true, the corresponding activities can be triggered. The traditional decision tree algorithm has the following limitations: First, the algorithm does not directly control the confounding variables. The model construction process may be affected by the confounding variables, resulting in inaccurate decision rules; second, the traditional decision tree algorithm often focuses on the overall prediction accuracy, which may cause the model to be biased towards the majority class, while ignoring those attribute values ​​that have a smaller number of samples but are more influential on the classification, affecting the integrity of the results; in addition, if the model under the current threshold fails to mine the rules at some decision points, the threshold needs to be lowered, which may cause the rules at some decision points to be extremely complex, and there is no primary and secondary relationship between each sub-condition, affecting the understanding of process execution. In response to the above-mentioned problems, this patent proposes a business process decision mining method and system based on causal inference, uses the LGBMClassifier based on the LightGBM algorithm as the basic classifier of the causal model, and adopts a meta-learning-based causal inference method based on the causal model to analyze the process decisions, which can better control confounding variables and handle unbalanced data. At the same time, thanks to the high efficiency of the LightGBM algorithm, the decision mining efficiency is higher, and the average treatment effect can be used as an evaluation criterion to measure the influence of each factor affecting the business process decision, so as to achieve a comprehensive, accurate and efficient analysis of the business process decision; finally, the decision mining results are used to construct a process model that can intuitively explain the business process decision, which can help enterprises improve and optimize their business processes. Summary of the invention

[0003] The first purpose of the present invention is to overcome the shortcomings and deficiencies of existing decision mining technology and provide a business process decision mining method based on causal inference, which breaks through the limitations of traditional methods that cannot directly control confounding variables, cannot handle unbalanced data and have low efficiency, improves the completeness and accuracy of business process decision mining, and can provide a process model that can intuitively explain business process decisions.

[0004] The second object of the present invention is to provide a business process decision mining system based on causal inference.

[0005] A third object of the present invention is to provide a storage medium.

[0006] A fourth object of the present invention is to provide a computing device.

[0007] The first object of the present invention is achieved by the following technical solution: a business process decision mining method based on causal inference comprises the following steps:

[0008] 1) Obtain basic data, namely, event logs and corresponding Petri nets that hide business process decision information;

[0009] 2) Preprocess the event log obtained in step 1) to obtain a standard event log that meets the requirements of subsequent steps, identify the selection structure in the Petri net corresponding to the event log, record all predecessor and successor relationship pairs in each selection structure, and obtain an npy file containing all selection structures;

[0010] 3) For each selected structure obtained in step 2), its predecessor transition corresponds to several events in the standard event log obtained in step 2), each event contains relevant attributes and attribute values, and the attribute data of all corresponding events are extracted to obtain the event-level attribute file corresponding to the selected structure, that is, the adjacent relationship data set of the selected structure; each selected structure is processed in turn to obtain several adjacent relationship data sets;

[0011] 4) For each adjacent relationship data set extracted in step 3), the relevant attributes are set as the relevant parameters of the causal model, including the outcome variable, treatment group, control group and covariate, and the basic classifier is selected to initialize the causal model; based on the causal model, the causal effect of all attribute values ​​on different "next activities" is analyzed by using the causal inference method based on meta-learning, and the average treatment effect (ATE) is calculated; ATE values ​​greater than a given threshold are regarded as valid causal relationships, and the corresponding predecessor-successor relationship pairs are recorded; all adjacent relationship data sets are analyzed in turn to obtain an npy file containing all valid causal relationships and their corresponding ATE;

[0012] 5) Sort the effective causal relationships of each predecessor-successor relationship pair obtained in step 4) according to their ATE values, and take the top three effective causal relationships with the highest ATE values ​​and add them to the corresponding transitions of the successor activities in the Petri net, as an explanation for the decision to select the successor transition to continue execution when the process executes to the selection structure instead of selecting other successor transitions, thereby obtaining a process model that can intuitively explain business process decisions.

[0013] Further, in step 1), the event log includes multiple cases, each case includes multiple events, each event corresponds to an activity, and also includes related attributes and attribute values; the attributes refer to specific information fields associated with each event, and these attributes are related to the decisions made at the decision point of the business process;

[0014] The Petri net is represented by a triple (K, M, L), wherein K represents a place, which is represented by a circle in the Petri net; M represents a transition, which is represented by a rectangle; and L represents a flow relationship, which is represented by a directed arc. K, M, and L are three basic elements of the Petri net.

[0015] Furthermore, in step 2), the event log obtained in step 1) is processed, including attribute value mapping, discretization processing and trajectory separation, as follows:

[0016] Map the attribute values ​​of the event log text type to numerical indexes that can be directly recognized and used by machine learning algorithms, and record the corresponding relationship between the initial value and the index;

[0017] Numerical discretization processing: extract all values ​​of the attribute and sort them, calculate the first quartile num41 and the third quartile num43, and divide the values ​​of the numerical attribute into three categories: values ​​less than num41 are mapped to 1, values ​​greater than num41 and less than num43 are mapped to 2, and values ​​greater than num43 are mapped to 3. The corresponding relationship between the size range of the initial value and the index is recorded.

[0018] Separate the events in the event log by track: traverse each line in the event log, that is, each event, and determine whether the track identifier of the current event is the same as that of the previous event. If they are the same, they belong to the same track and are added to the current track temporary list. If they are different, it means that a new track is found. The current track temporary list is added to the preprocessed event log, and the current track temporary list is updated, with the current event as the beginning of the new track. Repeat the above operation until the entire event log is traversed to obtain the preprocessed standard event log.

[0019] Further, in step 2), the selection structure in the Petri net obtained in step 1) is identified, and the specific situation is as follows:

[0020] The selection structure in the Petri net refers to: according to the execution order of activities in the Petri net, when the business process executes to a certain library in the Petri net, according to the triggering rules of the Petri net, multiple successor transitions can be triggered. At this time, the business process needs to make a decision and choose one among multiple candidate transitions to continue execution. The structure composed of this library place and its predecessor transition and several successor transitions is called the selection structure in the Petri net. A selection structure in the Petri net contains a predecessor transition and several successor transitions. The structure composed of a predecessor transition and a successor transition is called the selection structure. A predecessor-successor relationship pair in a structure; the predecessor transition of the place refers to the transition directly connected to the input arc of the place, and the successor transition of the place refers to the transition directly connected to the output arc of the place; the trigger rule means that when all input places of a transition contain at least one token, the transition can be triggered, and triggering a transition will result in the consumption of a token in all its input places and the generation of a token in all its output places; the token refers to the abstract mark of the resource in the Petri net, which is stored in the place, and the existence of the token determines whether the transition in the Petri net can be triggered;

[0021] Traverse the entire Petri net, calculate the number of successor transitions of each place in the Petri net, thereby identifying all the selection structures in it, record all the predecessor-successor relationship pairs in all the selection structures, and save them as an npy file.

[0022] Further, in step 3), the adjacent relationship refers to the activities corresponding to a predecessor-successor relationship in the selection structure obtained in step 2), and there is a direct following order relationship in a certain case recorded in the event log; extract the adjacent relationship data set corresponding to each selection structure in the event log, and the specific situation is as follows:

[0023] For each predecessor-successor relationship pair in a selection structure, traverse all the tracks in the event log. If the track contains an event corresponding to the predecessor transition, check whether its next event is the event corresponding to the successor transition. If so, extract all the attribute data of the event corresponding to the predecessor transition, and use the name of the successor transition as the attribute value of an extended attribute "next activity" of this event as an event data of the adjacent relationship data set; all the event data corresponding to all the predecessor-successor relationship pairs of a selection structure in the event log constitute the adjacent relationship data set of this selection structure. Each adjacent relationship data set is an event-level attribute file corresponding to the selection structure one-to-one, saved in csv format, where the value of the concept:name attribute is unique, that is, the name of the predecessor transition in the selection structure, and the "next activity" attribute has several different values, which correspond one-to-one to the name of the successor transition in the selection structure.

[0024] Further, in step 4), there is a causal relationship, that is, the former has a causal effect on the latter, and the causal effect means: let R and N be two random variables, if there are at least two different values ​​of R, such as r and r′, so that when the variable R is intervened, that is, its value is changed from r to r′, the conditional probability distribution of the variable N P(N|do(R=r)) is different from P(N|do(R=r′)), then R is said to have a causal effect on N; the intervention refers to actively changing the value of a variable, rather than simply observing its natural changes;

[0025] Initialize the causal model and use the causal inference method based on meta-learning based on the causal model to analyze the different values ​​of the corresponding attributes of the predecessor activity and the causal effects of its different subsequent activities. The specific steps are as follows:

[0026] a. Set the result variable: The result variable W is a binary variable used to indicate whether each event in the adjacent relationship data set leads to a specific "next activity". Set the "next activity" attribute in each adjacent relationship data set as the result variable W. If the value of the "next activity" attribute is equal to the name of the successor transition being analyzed, W is 1, otherwise it is 0.

[0027] b. Set the treatment group and control group: the treatment group refers to the group that has received some kind of intervention or treatment, and the control group refers to the group that has not received any intervention; set one of the values ​​of the attribute currently being analyzed in each adjacent relationship data set as the treatment group, and the other different values ​​of the attribute as the control group;

[0028] c. Set covariates: Covariates S are variables that may affect the value of the outcome variable, but are not the direct object of treatment or experimental intervention. Covariates are used to control confounding factors. Set all attributes in each adjacent relationship data set except the attribute currently being analyzed as covariates S.

[0029] d. Initialize the causal model and perform causal inference: Based on the above settings and using the LGBMClassifier based on the LightGBM algorithm as the basic classifier, initialize the causal model and use the meta-learning-based causal inference method to analyze the causal effects of all attribute values ​​in each adjacent relationship data set on different "next activities";

[0030] The causal inference is an analytical method used to determine whether there is a causal relationship between variables and the strength of this relationship; the meta-learning is a strategy that uses pre-learned experience to guide the learning of new tasks; the LightGBM (Light Gradient Boosting Machine) is a lightweight gradient boosting algorithm based on the Leaf-wise decision tree growth strategy and the histogram optimization strategy.

[0031] Further, in step 4), the ATE value is calculated and the effective causal relationship is summarized. The effective causal relationship refers to the processing group whose calculated ATE value is greater than a given threshold, indicating that the attribute and value corresponding to this processing group have an impact on the "next activity" and are regarded as one of the influencing factors of the process decision of selecting the subsequent transition in the corresponding selection structure. The specific steps are as follows:

[0032] a. Calculate the ATE of each treatment group: ATE is an indicator to measure the average impact of the treatment group on the outcome variable. The calculation formula is as follows:

[0033] ATE=E[W a＝1 -W a＝0 ] (1)

[0034] In the formula, W represents the outcome variable, W a＝1 represents the result under treatment or intervention conditions, W a＝0 represents the result under untreated or non-intervention conditions, E[W a＝1 -W a＝0 ] represents the expected value of the difference between the treated and untreated conditions, so ATE can be used to measure the degree of influence of the treatment group on the outcome variable;

[0035] b. For the treatment groups whose calculated ATE values ​​are greater than a given threshold, they are considered as valid causal relationships leading to the corresponding “next activity” and their ATE is recorded;

[0036] Finally, a causal model is built for each adjacent relationship data set in turn, the causal effect is analyzed, and the ATE value is calculated, and finally an npy file containing the effective causal relationships of all predecessor and successor relationship pairs and their corresponding ATE is obtained.

[0037] Furthermore, in step 5), according to the causal relationship between the predecessor activities of all selection structures obtained in step 4) and all their successor activities, the top three causal relationships with the highest ATE values ​​are selected, and the ATE value is appended after each causal relationship. The processed causal relationship is added as an extended attribute to the successor transition in the corresponding selection structure of the Petri net, and a process model is obtained that can intuitively explain the decisions made by the business process on the selection structure.

[0038] The second object of the present invention is achieved through the following technical solution: a business process decision mining system based on causal inference, used to implement the above-mentioned business process decision mining method based on causal inference, which includes:

[0039] Data acquisition module, used to obtain event logs and corresponding Petri nets that hide business process decision information;

[0040] The event log and model preprocessing module is used to map attribute values, discretize and separate trajectories of event logs, and extract selection structures in Petri nets;

[0041] A close neighbor relationship data set extraction module is used to extract the close neighbor relationship data set corresponding to each selection structure from the preprocessed event log;

[0042] The causal inference module is used to build a causal model for each close-proximity relationship data set, and analyze the causal effects of all attributes based on the causal model using a meta-learning-based causal inference method to obtain effective causal relationships and corresponding ATEs;

[0043] The model building module is used to process the causal relationship and add it to the Petri net to obtain a process model that can intuitively explain the decisions made by the business process in the selection structure.

[0044] The third purpose of the present invention is achieved through the following technical solution: a storage medium storing a program, which, when executed by a processor, implements the above-mentioned business process decision mining method based on causal inference.

[0045] The fourth purpose of the present invention is achieved through the following technical solution: a computing device, comprising a processor and a memory for storing processor executable programs, when the processor executes the program stored in the memory, the above-mentioned business process decision mining method based on causal inference is implemented.

[0046] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0047] 1. This invention uses a causal inference method based on meta-learning for the first time to analyze the causal relationship between attribute data in event logs and process decisions, and realizes the mining of business process decisions. It can better control the confounding factors in event logs and help improve the accuracy of business process decision mining.

[0048] 2. The present invention uses the LGBMClassifier based on the LightGBM algorithm as the basic classifier of the causal model for the first time to mine business process decisions, which can better handle data imbalance and is conducive to improving the completeness of business process decision mining.

[0049] 3. The present invention uses the LightGBM algorithm as the core algorithm for business process decision mining for the first time. Since the LightGBM algorithm uses a Leaf-wise decision tree growth strategy and a histogram-based optimization strategy, it has a fast training speed and is beneficial to improving the efficiency of business process decision mining.

[0050] 4. For the first time, the present invention adds ATE as an evaluation standard for the degree of influence on the decision after each influencing factor in the decision explanation model, and adds it to the process model after sorting according to the size of the ATE value, which is conducive to improving the interpretability of the process model for path decisions and facilitating people to analyze and understand the process more efficiently.

[0051] 5. The present invention has a wide range of applications in business process decision mining, and supports the use of decision mining results to build a decision explanation model. It also has broad prospects in improving and enhancing business processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 The figure is a logical flow diagram of the method of the present invention.

[0053] Figure 2 Schematic diagram of the Petri net of the case of the present invention.

[0054] Figure 3 Schematic diagram of the process model of the case of the present invention.

[0055] Figure 4 Schematic diagram of the system of the present invention. DETAILED DESCRIPTION

[0056] The present invention will be further described below in conjunction with specific embodiments.

[0057] Example 1

[0058] like Figure 1 As shown, this embodiment discloses a business process decision mining method based on causal inference, firstly, basic data is obtained, namely, event logs and corresponding Petri nets that hide business process decision information; then, the event logs are preprocessed and the selection structure in the Petri net is identified; then, the adjacent relationship data set corresponding to each selection structure is extracted from the processed event logs; a causal model is constructed for each extracted adjacent relationship data set, and a causal inference method based on meta-learning is used based on the causal model to analyze the causal effects of all attribute values ​​on different "next activities" to obtain effective causal relationships and corresponding ATEs; finally, the processed causal relationships are added to the Petri net to obtain a process model that can intuitively explain business process decisions. Specifically, the following steps are included:

[0059] 1) Obtain basic data, namely, event logs that hide business process decision information and corresponding Petri nets.

[0060] Using the above steps, Figure 2Taking the case Petri net and the corresponding process shown in the figure as an example, there are 10 transitions in the Petri net process model, namely A, B, C, D, E, F, G, H, I, and J, which correspond to the activities in the business process respectively. p2, p3, p4, and p5 are places in the Petri net; the event log corresponding to this process contains a total of 200 traces and 5 data attributes. The details of each trace variant are shown in Table 1.

[0061] Table 1 Event log trace composition

[0062]

[0063] 2) Process the event log obtained in step 1) and identify the selection structure in the Petri net, as follows:

[0064] A. Preprocess the event logs obtained in step 1), including attribute value mapping, discretization and trajectory separation. The details are as follows:

[0065] A.1) Map the attribute values ​​of the event log text type to numerical indexes that can be directly recognized and used by machine learning algorithms, and record the corresponding relationship between the initial value and the index;

[0066] A.2) Numerical discretization processing: extract all values ​​of the attribute and sort them, calculate the first quartile num41 and the third quartile num43, and divide the values ​​of the numerical attribute into three categories: values ​​less than num41 are mapped to 1, values ​​greater than num41 and less than num43 are mapped to 2, and values ​​greater than num43 are mapped to 3. The corresponding relationship between the size range of the initial value and the index is recorded;

[0067] A.3) Separate the events in the event log by track: traverse each line in the event log, that is, each event, and determine whether the track identifier of the current event is the same as that of the previous event. If they are the same, they belong to the same track and are added to the current track temporary list. If they are different, it means that a new track is found. The current track temporary list is added to the preprocessed event log, and the current track temporary list is updated, with the current event as the beginning of the new track; repeat the above operation until the entire event log is traversed;

[0068] Using the above steps, Figure 2Taking the case Petri net shown in Figure 1 and the event log shown in Table 1 as examples, the result of mapping the text type attribute values ​​in the event log to numerical indexes is shown in Table 2. In the table, concept:name and P are two attributes in the event log. The first and second columns are the corresponding attribute values ​​before and after the mapping of the attribute concept:name. Similarly, the third and fourth columns are the corresponding attribute values ​​before and after the mapping of the attribute P. According to num41 and num43, the result of discretizing the continuous attribute values ​​in the event log is shown in Table 3. Q, X, Y, and Z are continuous attributes in the event log. The two columns under each attribute are the attribute values ​​before and after the discretization. The event log is separated by trajectory according to the trajectory identifier case:concept:name, and finally the processed event log data is obtained. The information of the first three trajectories in the event log is shown in Table 4.

[0069] Table 2 Mapping of text type attribute values ​​in event logs

[0070]

[0071] Table 3 Attribute value mapping for discretization in event logs

[0072]

[0073] Table 4 The first three trace information of the processed event log

[0074] case:concept:name concept:name StartTime CompleteTime P Q X Y Z trace1 1 00:00.0 00:00.0 1 2 2 3 2 trace1 2 00:01.0 00:01.0 1 3 1 3 2 trace1 5 00:01.0 00:01.0 1 3 1 3 2 trace2 1 00:01.0 00:01.0 2 2 1 3 3 trace2 4 00:02.0 00:02.0 1 2 2 2 3 trace2 9 00:03.0 00:03.0 1 2 1 2 3 trace3 1 00:04.0 00:04.0 2 3 1 3 2 trace3 4 00:04.0 00:04.0 2 3 1 3 3 trace3 10 00:05.0 00:05.0 2 3 1 3 2

[0075] B. Identify the selection structure in the Petri net obtained in step 1), as follows:

[0076] Traverse all transitions in the Petri net and check the number of successor transitions in the output location of each transition. If there is an output location with a number of successor transitions greater than 1, the transition is considered to be a predecessor transition in the selection structure, and all successor transitions of the output location are considered to be successor transitions in the selection structure. Add the transition name of this successor transition to the successor transition list of the predecessor transition as a selection structure; a selection structure in the Petri net contains a predecessor transition and several successor transitions. The structure composed of a predecessor transition and a successor transition is called a predecessor-successor relationship pair in the selection structure; record all predecessor-successor relationship pairs in all selection structures and save them as an npy file.

[0077] Using the above steps, Figure 2Taking the case Petri net shown in the figure as an example, the selection structure in the Petri net is identified and saved as an npy file. The file content is displayed in a table as shown in Table 5. This Petri net includes a total of four selection structures. The selection structure with A as the predecessor transition includes a total of three successor transitions, namely B, C, and D. Therefore, this selection structure includes a total of three predecessor and successor relationship pairs, namely AB, AC, and AD. The same is true for the other three selection structures.

[0078] Table 5 Selection structure in Petri network

[0079]

[0080] 3) Extract the adjacent relationship dataset corresponding to each selection structure in the event log, as follows:

[0081] For each predecessor-successor relationship pair in each selection structure obtained in step 2), traverse all tracks in the event log. If the track contains an event corresponding to the predecessor transition, check whether its next event is the event corresponding to the successor transition. If so, extract all attribute data of the event corresponding to the predecessor transition, and use the name of the successor transition as the attribute value of an extended attribute "next activity" of this event as an event data of the adjacent relationship data set; all event data corresponding to all predecessor-successor relationship pairs of a selection structure in the event log constitute the adjacent relationship data set of this selection structure, and each adjacent relationship data set is an event-level attribute file corresponding to the selection structure one-to-one, saved in csv format; the value of the concept:name attribute is unique, that is, the name of the predecessor transition in the selection structure, and the "next activity" attribute has several different values, which correspond one-to-one to the name of the successor transition in the selection structure.

[0082] Using the above steps, Figure 2 Taking the case Petri net and the corresponding process shown in Table 5 as an example, for each selection structure in Table 5, the event data related to it is retrieved in the event log shown in Table 4, and the continuous relationship data set is extracted to obtain four data sets A.csv, B.csv, C.csv and D.csv. Taking the A.csv data set as an example, the first ten event data are shown in Table 6.

[0083] Table 6A.csv first ten event data

[0084]

[0085] 4) According to the adjacent relationship data set of each selection structure obtained in step 3), the relevant attributes therein are set as the relevant parameters of the causal model, and the basic classifier is selected to initialize the causal model; based on the causal model, the causal effect of all attribute values ​​on different "next activities" is analyzed by using the causal inference method based on meta-learning, and the effective causal relationship and ATE of each corresponding predecessor-successor relationship pair are obtained. The specific steps are as follows:

[0086] 1) Initialize the causal model. Based on the causal model, use the meta-learning-based causal inference method to analyze the causal effects of different values ​​of the corresponding attributes of the predecessor activity on its different subsequent activities. The specific situation is as follows:

[0087] 1.1) Set the result variable: The result variable W is a binary variable used to indicate whether each event in the adjacent relationship data set leads to a specific "next activity"; set the "next activity" attribute in each adjacent relationship data set to the result variable W. If the value of the "next activity" attribute is equal to the name of the successor transition being analyzed, W is 1, otherwise it is 0.

[0088] 1.2) Set the treatment group and control group: the treatment group refers to the group that received a certain intervention or treatment, and the control group refers to the group that did not receive the intervention; in each adjacent relationship data set, one of the values ​​of the attribute currently being analyzed is set as the treatment group, and the other different values ​​of the attribute are set as the control group;

[0089] 1.3) Set covariates: Covariates S are variables that may affect the value of the outcome variable, but are not the direct object of treatment or experimental intervention. Covariates are used to control confounding factors. In each adjacent relationship data set, all attributes except the attribute currently being analyzed are set as covariates S;

[0090] 1.4) Initialize the causal model and perform causal inference: Based on the above settings and using the LGBMClassifier based on the LightGBM algorithm as the basic classifier, initialize the causal model and use the meta-learning-based causal inference method to analyze the causal effects of all attribute values ​​in each adjacent relationship data set on different "next activities".

[0091] 2) Calculate the ATE value and summarize the effective causal relationship. The effective causal relationship refers to the processing group whose calculated ATE value is greater than the given threshold, indicating that the attribute and value corresponding to this processing group have an impact on the "next activity" and can be regarded as one of the influencing factors of the process decision of selecting the subsequent transition in the corresponding selection structure. The specific steps are as follows:

[0092] 2.1) Calculate ATE to evaluate causal effects. ATE refers to the average treatment effect, which is an indicator to measure the degree of influence of the treatment group on the overall average result. Its calculation formula is as follows:

[0093] ATE=E[W a＝1 -W a＝0 ] (1)

[0094] In the formula, W represents the outcome variable, W a＝1 represents the result under treatment or intervention conditions, W a＝0 represents the result under untreated or non-intervention conditions, E[W a＝1 -W a＝0 ] represents the expected value of the difference between the treated and untreated conditions, so ATE can be used to measure the degree of influence of the treatment group on the outcome variable;

[0095] 2.2) For treatment groups whose calculated ATE values ​​are greater than a given threshold, they are considered to be effective causal relationships leading to the corresponding predecessor and successor activities, and their ATE is recorded. According to expert experience, the ATE threshold is usually set to 0.05, that is, when the ATE value is greater than 0.05, the treatment group is considered to have an effect on the outcome variable.

[0096] Using the above steps, Figure 2 Taking the case Petri net and the corresponding process shown in Table 6 as an example, a causal model is constructed for each adjacent relationship data set shown in Table 6, the causal effect is analyzed, and the ATE value is calculated. The effective causal relationship of each predecessor-successor relationship pair is finally obtained as shown in Table 7. Taking A as the selection structure of the predecessor as an example, according to the results, after the process executes transition A, if the value of attribute Q is greater than 75, the value of attribute Z is greater than 74, the value of attribute X is greater than 83, and the value of attribute P is FALSE, the process is more inclined to execute transition B; similarly, if the value of attribute Z is less than 35, the value of attribute X is less than 38, and the value of attribute Q is between 29 and 75, the process is more inclined to execute transition C; if the value of attribute X is between 38 and 83, the value of attribute Z is between 35 and 74, the value of attribute Q is less than 29, and the value of attribute P is TRUE, the process is more inclined to execute transition D; this information can provide a strong basis for business process decision-making.

[0097] Table 7 Summary of cause-effect relationships in case flow

[0098]

[0099]

[0100] 5) According to the causal relationship between the predecessor activities of all selection structures obtained in step 4) and all their successor activities, select the top three causal relationships with the highest ATE values, and append the ATE value after each causal relationship. Add the processed causal relationship as an extended attribute to the successor transition in the corresponding selection structure of the Petri net, and obtain a process model that can intuitively explain the decisions made by the business process on the selection structure.

[0101] Using the above steps, Figure 2 Taking the case Petri net and the corresponding process as an example, the process model that explains the business process decision is as follows Figure 3 shown.

[0102] Example 2

[0103] This embodiment discloses a business process decision mining system based on causal inference, which is used to implement the business process decision mining method based on causal inference described in Example 1. Figure 4 As shown, the system includes the following functional modules:

[0104] Data acquisition module, used to obtain event logs and corresponding Petri nets that hide business process decision information;

[0105] The event log and model preprocessing module is used to map attribute values, discretize and separate trajectories of event logs, and extract selection structures in Petri nets;

[0106] A close neighbor relationship data set extraction module is used to extract the close neighbor relationship data set corresponding to each selection structure from the preprocessed event log;

[0107] The causal inference module is used to build a causal model for each close-proximity relationship data set, and analyze the causal effects of all attributes based on the causal model using a meta-learning-based causal inference method to obtain effective causal relationships and corresponding ATEs;

[0108] The model building module is used to process the causal relationship and add it to the Petri net to obtain a process model that can intuitively explain the decisions made by the business process in the selection structure.

[0109] Example 3

[0110] This embodiment discloses a storage medium storing a program. When the program is executed by a processor, the business process decision mining method based on causal inference described in Embodiment 1 is implemented.

[0111] The storage medium in this embodiment can be a disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), a USB flash drive, a mobile hard disk, or other media.

[0112] Example 4

[0113] This embodiment discloses a computing device, including a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the business process decision mining method based on causal inference described in Example 1 is implemented.

[0114] The computing device described in this embodiment may be a desktop computer, a laptop computer, a smart phone, a PDA handheld terminal, a tablet computer, a programmable logic controller (PLC), or other terminal devices with a processor function.

[0115] To sum up, after adopting the above scheme, the present invention provides a new method and system for business process decision mining, using the LGBMClassifier based on the LightGBM algorithm as the basic classifier of the causal model, and adopting the causal inference method based on meta-learning based on the causal model to analyze the process decision, which can better control the confounding variables and process unbalanced data; at the same time, thanks to the high efficiency of the LightGBM algorithm, the decision mining efficiency is higher; and the average treatment effect can be used as the evaluation criterion to measure the influence of each factor affecting the business process decision; finally, the results of decision mining are used to construct a process model that can intuitively explain the business process decision, which can help enterprises improve and optimize the business process, has practical promotion value, and is worthy of promotion.

[0116] The above-described embodiments are only preferred embodiments of the present invention and are not intended to limit the scope of implementation of the present invention. Therefore, all changes made according to the shape and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A business process decision mining method based on causal inference, characterized in that: The following steps are involved: 1) Obtain basic data, namely, event logs and corresponding Petri nets that hide business process decision information; 2) Preprocess the event log obtained in step 1) to obtain a standard event log that meets the requirements of subsequent steps, identify the selection structure in the Petri net corresponding to the event log, record all predecessor and successor relationship pairs in each selection structure, and obtain an npy file containing all selection structures; 3) For each selected structure obtained in step 2), its predecessor transition corresponds to several events in the standard event log obtained in step 2), each event contains relevant attributes and attribute values, and the attribute data of all corresponding events are extracted to obtain the event-level attribute file corresponding to the selected structure, that is, the adjacent relationship data set of the selected structure; Process each selection structure in turn to obtain several adjacent relationship data sets; 4) For each adjacent relationship data set extracted in step 3), the relevant attributes therein are set as relevant parameters of the causal model, including the outcome variable, the treatment group, the control group and the covariate, and a basic classifier is selected to initialize the causal model; Based on the causal model, a causal inference method based on meta-learning is used to analyze the causal effects of all attribute values ​​on different "next activities" and calculate their average treatment effect ATE; ATE values ​​greater than a given threshold are considered valid causal relationships, and their corresponding predecessor-successor relationship pairs are recorded; all adjacent relationship data sets are analyzed in turn to obtain an npy file containing all valid causal relationships and their corresponding ATEs; 5) Sort the effective causal relationships of each predecessor-successor relationship pair obtained in step 4) according to their ATE values, and take the top three effective causal relationships with the highest ATE values ​​and add them to the corresponding transitions of the successor activities in the Petri net, as an explanation for the decision to select the successor transition to continue execution when the process executes to the selection structure instead of selecting other successor transitions, thereby obtaining a process model that can intuitively explain business process decisions.

2. The business process decision mining method based on causal inference according to claim 1 is characterized by: In step 1), the event log contains multiple cases, each case contains multiple events, each event corresponds to an activity, and also contains related attributes and attribute values; the attributes refer to specific information fields associated with each event, and these attributes are related to the decisions made at the decision point of the business process; The Petri net is represented by a triple (K, M, L), wherein K represents a place, which is represented by a circle in the Petri net; M represents a transition, which is represented by a rectangle; and L represents a flow relationship, which is represented by a directed arc. K, M, and L are three basic elements of the Petri net.

3. The business process decision mining method based on causal inference according to claim 1 is characterized by: In step 2), the event log obtained in step 1) is preprocessed, including attribute value mapping, discretization and trajectory separation, as follows: Map the attribute values ​​of the event log text type to numerical indexes that can be directly recognized and used by machine learning algorithms, and record the corresponding relationship between the initial value and the index; Numerical discretization processing: extract all values ​​of the attribute and sort them, calculate the first quartile num41 and the third quartile num43, and divide the values ​​of the numerical attribute into three categories: values ​​less than num41 are mapped to 1, values ​​greater than num41 and less than num43 are mapped to 2, and values ​​greater than num43 are mapped to 3. The corresponding relationship between the size range of the initial value and the index is recorded. Separate the events in the event log by trajectory: traverse each line in the event log, that is, each event, and determine whether the trajectory identifier of the current event is the same as that of the previous event. If they are the same, they belong to the same trajectory and are added to the current trajectory temporary list. If they are different, it means that a new trajectory has been found. The current trajectory temporary list is added to the preprocessed event log, and the current trajectory temporary list is updated, with the current event as the beginning of the new trajectory. Repeat the above operations until the entire event log is traversed to obtain the preprocessed standard event log.

4. The business process decision mining method based on causal inference according to claim 1 is characterized by: In step 2), the selection structure in the Petri net obtained in step 1) is identified, as follows: The selection structure in the Petri net refers to: according to the execution order of activities in the Petri net, when the business process executes to a certain library in the Petri net, according to the triggering rules of the Petri net, multiple successor transitions can be triggered. At this time, the business process needs to make a decision and choose one among multiple candidate transitions to continue execution. The structure composed of this library place and its predecessor transition and several successor transitions is called the selection structure in the Petri net. A selection structure in the Petri net contains a predecessor transition and several successor transitions. The structure composed of a predecessor transition and a successor transition is called the selection structure. A predecessor-successor relationship pair in a structure; the predecessor transition of the place refers to the transition directly connected to the input arc of the place, and the successor transition of the place refers to the transition directly connected to the output arc of the place; the trigger rule means that when all input places of a transition contain at least one token, the transition can be triggered, and triggering a transition will result in the consumption of a token in all its input places and the generation of a token in all its output places; the token refers to the abstract mark of the resource in the Petri net, which is stored in the place, and the existence of the token determines whether the transition in the Petri net can be triggered; Traverse the entire Petri net, calculate the number of successor transitions of each place in the Petri net, thereby identifying all the selection structures in it, record all the predecessor-successor relationship pairs in all the selection structures, and save them as an npy file.

5. The business process decision mining method based on causal inference according to claim 1 is characterized by: In step 3), the adjacent relationship refers to the activities corresponding to a predecessor-successor relationship in the selection structure obtained in step 2), and there is a direct sequence relationship in a certain case recorded in the event log; extract the adjacent relationship data set corresponding to each selection structure in the event log, the specific situation is as follows: For each predecessor-successor relationship pair in a selection structure, traverse all tracks in the event log. If the track contains an event corresponding to the predecessor transition, check whether its next event is the event corresponding to the successor transition. If so, extract all attribute data of the event corresponding to the predecessor transition, and use the name of the successor transition as the attribute value of an extended attribute "next activity" of this event as an event data of the adjacent relationship data set; all event data corresponding to all predecessor-successor relationship pairs of a selection structure in the event log constitute the adjacent relationship data set of this selection structure. Each adjacent relationship data set is an event-level attribute file corresponding to the selection structure one-to-one, saved in csv format, where the value of the concept:name attribute is unique, that is, the name of the predecessor transition in the selection structure, and the "next activity" attribute has several different values, corresponding to the name of the successor transition in the selection structure one-to-one.

6. The business process decision mining method based on causal inference according to claim 1 is characterized by: In step 4), there is a causal relationship, that is, the former has a causal effect on the latter. The causal effect means: let R and N be two random variables. If there are at least two different values ​​of R, such as r and r′, so that when the variable R is intervened, that is, its value is changed from r to r′, the conditional probability distribution of the variable N is different from P(N|do(R=r)) and P(N|do(R=r′)), then R is said to have a causal effect on N; the intervention means actively changing the value of a variable, rather than simply observing its natural changes; Initialize the causal model and use the causal inference method based on meta-learning based on the causal model to analyze the different values ​​of the corresponding attributes of the predecessor activity and the causal effects of its different subsequent activities. The specific steps are as follows: a. Set the result variable: The result variable W is a binary variable used to indicate whether each event in the adjacent relationship data set leads to a specific "next activity"; Set the "next activity" attribute in each adjacent relationship data set to the result variable W. If the value of the "next activity" attribute is equal to the name of the successor transition being analyzed, then W is 1, otherwise it is 0; b. Set the treatment group and control group: the treatment group refers to the group that has received some kind of intervention or treatment, and the control group refers to the group that has not received any intervention; set one of the values ​​of the attribute currently being analyzed in each adjacent relationship data set as the treatment group, and the other different values ​​of the attribute as the control group; c. Set covariates: Covariates S are variables that may affect the value of the outcome variable, but are not the direct object of treatment or experimental intervention. Covariates are used to control confounding factors. Set all attributes in each adjacent relationship data set except the attribute currently being analyzed as covariates S. d. Initialize the causal model and perform causal inference: Based on the above settings and using the LGBMClassifier based on the LightGBM algorithm as the basic classifier, initialize the causal model and use the causal inference method based on meta-learning to analyze the causal effects of all attribute values ​​in each adjacent relationship data set on different "next activities"; The causal inference is an analytical method used to determine whether there is a causal relationship between variables and the strength of this relationship; the meta-learning is a strategy that uses pre-learned experience to guide the learning of new tasks; the LightGBM is a lightweight gradient boosting algorithm based on the Leaf-wise decision tree growth strategy and the histogram optimization strategy.

7. The business process decision mining method based on causal inference according to claim 1 is characterized by: In step 4), the ATE value is calculated and the effective causal relationship is summarized. The effective causal relationship refers to the processing group whose calculated ATE value is greater than the given threshold, indicating that the attribute and value corresponding to this processing group have an impact on the "next activity" and are regarded as one of the influencing factors of the process decision of selecting the subsequent transition in the corresponding selection structure. The specific steps are as follows: a. Calculate the ATE of each treatment group: ATE is an indicator to measure the average impact of the treatment group on the outcome variable. The calculation formula is as follows: ATE=E[W a＝1 -W a＝0 ] (1) In the formula, W represents the outcome variable, W a＝1 represents the result under treatment or intervention conditions, W a＝0 represents the result under untreated or unintervention conditions, E[W a＝1 -W a＝0 ] represents the expected value of the difference between the treated and untreated conditions, so ATE can be used to measure the degree of influence of the treatment group on the outcome variable; b. For the treatment groups whose calculated ATE values ​​are greater than a given threshold, they are considered as valid causal relationships leading to the corresponding "next activity" and their ATE is recorded; Finally, a causal model is built for each adjacent relationship data set in turn, the causal effect is analyzed, and the ATE value is calculated, and finally an npy file containing the effective causal relationships of all predecessor and successor relationship pairs and their corresponding ATE is obtained.

8. The business process decision mining method based on causal inference according to claim 1 is characterized by: In step 5), according to the causal relationship between the predecessor activities of all selection structures obtained in step 4) and all their successor activities, the top three causal relationships with the highest ATE values ​​are selected, and the ATE value is appended after each causal relationship. The processed causal relationship is added as an extended attribute to the successor transition in the corresponding selection structure of the Petri net, and a process model is obtained that can intuitively explain the decisions made by the business process on the selection structure.

9. Business process decision mining system based on causal inference, characterized by: A business process decision mining method based on causal inference for implementing any one of claims 1 to 8, comprising: Data acquisition module, used to obtain event logs and corresponding Petri nets that hide business process decision information; The event log and model preprocessing module is used to map attribute values, discretize and separate trajectories of event logs, and extract selection structures in Petri nets; A close neighbor relationship data set extraction module is used to extract the close neighbor relationship data set corresponding to each selection structure from the preprocessed event log; The causal inference module is used to build a causal model for each close-proximity relationship data set, and analyze the causal effects of all attributes based on the causal model using a meta-learning-based causal inference method to obtain effective causal relationships and corresponding ATEs; The model building module is used to process the causal relationship and add it to the Petri net to obtain a process model that can intuitively explain the decisions made by the business process in the selection structure.

Images