Business process next event prediction method based on multi-dimensional attribute information fusion
The G-OL neural network, combining GCN and ONLSTM networks, addresses the lack of spatial correlation in existing methods by integrating time and spatial attributes, improving prediction accuracy and interpretability in business process forecasting.
Patent Information
- Application Number
- CN202411562729.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-05
- Publication Date
- 2024-12-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing methods for predicting the next event in a business process primarily focus on temporal correlation, neglecting spatial correlation, leading to reduced prediction accuracy and interpretability.
A method using a G-OL neural network that fuses time and spatial attribute information, incorporating a GCN graph convolutional neural network for spatial feature extraction and an ONLSTM neural network for temporal relationship extraction, to improve prediction accuracy and interpretability.
Enhances the accuracy and interpretability of next event prediction by integrating spatial and temporal attributes, achieving a prediction accuracy of 0.802 and demonstrating robustness across various business processes.
Smart Images

Figure CN119066622A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of process mining, and specifically relates to a method for predicting the next event of a business process based on multi-dimensional attribute information fusion. Background Art
[0002] Process mining can effectively optimize and improve the execution efficiency of business processes through systematic analysis of event data. With this technology, users can identify bottlenecks in the process, discover operational deviations, perform compliance diagnosis, and reduce failures and repeated operations. As an important branch of process mining, predictive process monitoring aims to predict the future development trend of the process being executed, and predict the process that has not yet ended by analyzing the historical complete process execution data. It has a wide range of applications, including predicting the results of process execution, the next event of the business process, and the remaining time of the process. In the fields of education, medical care, finance, etc., predictive process monitoring has demonstrated its practical application value, for example, helping enterprises reduce maintenance costs, provide support for decision-making, and prevent the occurrence of illegal tasks.
[0003] As one of the core tasks of predictive process monitoring, the prediction of the next event in a business process aims to predict the next possible event in the currently executed business process by analyzing historical event logs, so as to improve the transparency of the business process, optimize resource utilization, reduce operational risks, and provide intelligent support for enterprise decision-making. With the gradual development of artificial intelligence technology in the era of big data, the widespread application of deep learning has also provided new research ideas for the prediction of the next event in business processes. Traditional methods usually regard events, trajectories, and attributes in event logs as one-dimensional sequence data of unequal lengths. Researchers learn the relationship between events by inputting them into deep learning models, thereby improving the accuracy of the next event prediction. However, existing methods usually only focus on the temporal correlation between events, but ignore the spatial correlation between events, which leads to the model's lack of prediction accuracy and interpretability. Therefore, introducing the spatial characteristics of events into the prediction model is expected to further improve the accuracy of the next event prediction in the process and the interpretability of the model. Summary of the invention
[0004] In order to solve the above problems, the present invention proposes a method for predicting the next event in a business process based on the fusion of multi-dimensional attribute information, constructs a G-OL neural network that simultaneously fuses time and spatial attribute information, extracts the hierarchical relationship between events by analyzing the event time series, and uses the spatial attribute information of the events to improve the accuracy of the prediction of the next event in the business process, which is practical.
[0005] The technical solution of the present invention is as follows: A method for predicting the next event of a business process based on multi-dimensional attribute information fusion includes the following steps: Step 1: Obtain the business process event log, analyze the direct follow-up relationship of the activities in the event log, extract the time attribute information, construct a data set with the process trajectory prefix as the known information and the next event as the predicted target value, and divide it into a training set and a test set; Step 2: Construct the graph structure of the event log based on the direct follow-up relationship of the activities obtained in step 1; Step 3: construct a multi-dimensional attribute information fusion prediction model G-OL, which includes a spatial feature extraction module and a temporal relationship extraction module; Step 4: Use the training set to iteratively train the multi-dimensional attribute information fusion prediction model G-OL; Step 5: Input the test set into the multi-dimensional attribute fusion prediction model G-OL trained in step 4 to obtain the next event prediction result, and use the accuracy index to evaluate and analyze the prediction result.
[0006] Furthermore, in step 1, the process trajectory prefix is the trajectory information before any activity node or time on a certain trajectory.
[0007] Furthermore, in step 2, constructing the graph structure of the event log refers to constructing all activities in the event log trajectory and the direct following activity relationships between activities as nodes and edges in the graph respectively, and constructing the relationships between activities in all trajectories into an adjacency matrix. The relationships between activities in all trajectories include sequential, selection and concurrent relationships.
[0008] Furthermore, in step 3, the spatial feature extraction module is composed of a two-layer GCN graph convolutional neural network, which is a neural network structure used to extract graph structure features; the temporal relationship extraction module adopts the ONLSTM neural network, which is a variant of the recurrent neural network. The core of the ONLSTM neural network is to allow the network to dynamically divide neurons into different levels of granularity to help capture long-term dependencies and hierarchical information in the sequence.
[0009] Furthermore, in step 3, the specific working process of the multi-dimensional attribute information fusion prediction model G-OL is as follows: Step 3.1: Input the graph structure of the event log into the spatial feature extraction module to extract spatial attribute information; the forward propagation calculation formula of the spatial feature extraction module is as follows: ; In the formula, It is the output matrix of the spatial feature extraction module; is the activation function; is the node feature matrix; is the adjacency matrix of the trajectory; is the softmax activation function; is the sum of the adjacency matrix and the identity matrix of the trajectory, that is , is the identity matrix; is the ReLU activation function; and They are the parameter matrices of the two-layer GCN graph convolutional neural network; ReLU and softmax are the activation functions of the two-layer GCN graph convolutional neural network; Step 3.2: Input the temporal attribute information and spatial attribute information into the temporal relationship extraction module to obtain the temporal information and hierarchical relationship between events; the specific process is as follows: The ONLSTM neural network controls the degree of forgetting historical information and retaining the current information through the gate structure, and distinguishes historical information from current information by dividing the level of network unit state information. In order to achieve the interval division, the model uses two integers and , which are used to represent the lowest level of historical information and the highest level of current information, respectively. In the process of model updating, when When the two intervals overlap, the overlapping part is updated using the original long short-term memory neural network LSTM. For the part below Only the currently entered information is retained. When the two intervals do not overlap, for the non-overlapping part, the current storage unit value is directly set to 0, and for the part greater than and less than of the components, respectively retaining only historical information and current input information.
[0010] Furthermore, in step 4, the Adam optimizer is used, the size of the regularization term coefficient and the size of the initial learning rate are pre-set, and the time and space attribute information of the events in the event log are integrated for model training. During each training, the trajectory prefix of a different event log is replaced for input, so as to predict the next event in the business process and realize real-time tracking and prediction in the actual business process.
[0011] Furthermore, in step 5, the calculation formula of the accuracy index is as follows: ; in, is the accuracy value; is the total number of samples; is the number of categories; For the The number of samples in the class; , , , Respectively The number of correctly classified positive examples, the number of incorrectly classified positive examples, the number of incorrectly classified negative examples, and the number of correctly classified negative examples in the class.
[0012] The beneficial technical effects brought by the present invention are as follows: the present invention extracts the spatial features between events for the next event prediction task of the process for the first time, and obtains the temporal information and hierarchical relationship between events for the next event prediction task of the process for the first time, which is innovative; the present invention integrates the GCN graph convolutional neural network and the ONLSTM neural network for the first time to construct a multi-dimensional attribute information fusion prediction model G-OL for predicting the next event of the process, and iteratively trains the prediction model, and replaces the trajectory prefix of different event logs for input each time during training, so as to predict the next event in the business process and realize real-time tracking and prediction in the actual business process; the present invention uses the accuracy rate Accuracy to evaluate the next event prediction method of the process based on G-OL, which effectively improves the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Figure 1 This is a flow chart of the method for predicting the next event in a business process based on multi-dimensional attribute information fusion of the present invention.
[0014] Figure 2 This is an architecture diagram of the next event prediction system for a business process based on multi-dimensional attribute information fusion of the present invention. DETAILED DESCRIPTION
[0015] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: The flow chart of the present invention is as follows Figure 1 As shown, starting from the event log, the acquired event log is first preprocessed, the direct follow-up relationship of the activity is analyzed, the time attribute information is extracted, and the training set and the test set are divided. Then, the graph structure of the event log is constructed, and its spatial attribute information is extracted. Then, the temporal information and hierarchical relationship between events are obtained, and then the multi-dimensional attribute information fusion prediction model is trained, and finally the prediction model is evaluated and analyzed. The present invention also designs a business process next event prediction system based on multi-dimensional attribute information fusion according to the above process. The main functional modules of the system include: event log preprocessing module, process spatial feature extraction module, process temporal relationship extraction module, multi-dimensional attribute information fusion prediction model training module, and multi-dimensional attribute information fusion model prediction result evaluation module.
[0016] A method for predicting the next event of a business process based on multi-dimensional attribute information fusion includes the following steps: Step 1: Obtain the business process event log, analyze the direct follow-up relationship of activities in the event log, extract time attribute information, build a data set with the process trajectory prefix as the known information and the next event as the predicted target value, and divide it into a training set and a test set.
[0017] The process trajectory prefix is the trajectory information before any activity node or time on a certain trajectory.
[0018] In this step, the data set selects the open source real business process event log BPIC_2012_A. BPIC_2012_A is in .xes format. The record format of the event log is shown in Table 1.
[0019] Table 1 Contents of the BPIC_2012_A event log .
[0020] Step 2: Based on the direct follow-up relationship of the activities obtained in step 1, a graph structure of the event log is constructed.
[0021] The graph structure of constructing the event log refers to constructing all activities in the event log trajectory and the direct following activity relationships between activities as nodes and edges in the graph respectively, and constructing the relationships between activities in all trajectories into an adjacency matrix. The relationships between activities in all trajectories include sequential, selective and concurrent relationships.
[0022] Step 3: construct a multi-dimensional attribute information fusion prediction model G-OL, which includes a spatial feature extraction module and a temporal relationship extraction module; The spatial feature extraction module is composed of two layers of GCN graph convolutional neural network; the GCN graph convolutional neural network refers to a neural network structure specifically used to extract graph structure features; The temporal relationship extraction module adopts the ONLSTM neural network, which is a recurrent neural network variant. The core of the ONLSTM neural network is to allow the network to dynamically divide neurons into different levels of granularity to help capture long-term dependencies and hierarchical information in the sequence. The specific working process of the multi-dimensional attribute information fusion prediction model G-OL is as follows: Step 3.1: Input the graph structure of the event log into the spatial feature extraction module to extract spatial attribute information; the specific process is as follows: A two-layer GCN graph convolutional neural network is used to analyze the correlation between all events in the event log trajectory and extract the spatial features between all events. The forward propagation calculation formula of the spatial feature extraction module is as follows: ; In the formula, It is the output matrix of the spatial feature extraction module; is the activation function; is the node feature matrix; is the adjacency matrix of the trajectory; is the softmax activation function; is the sum of the adjacency matrix and the identity matrix of the trajectory, that is , is the identity matrix; is the ReLU activation function; and They are the parameter matrices of the two-layer GCN graph convolutional neural network; ReLU and softmax are the activation functions of the two-layer GCN graph convolutional neural network. Building a two-layer GCN graph convolutional network enables the model to better understand the hierarchical relationships and remote dependencies in complex graph structures, thereby better processing complex graph structures and improving the performance of the next event prediction task in the business process.
[0023] In this step, the input graph structure of the GCN graph convolutional neural network is a trajectory Take the example to show that the trajectory consists of multiple activities, as shown in Table 1. , , , , Represent the event IDs of five different activities respectively. The adjacency matrix and degree matrix are constructed based on these five activities. For example, in the adjacency matrix, if , If there is a direct follow-up relationship between the corresponding activities, the first row and second column of the adjacency matrix is 1. Since the trajectory contains five different activities, a 5×5 matrix can be obtained. Finally, the adjacency matrix of the trajectory is obtained. Sum degree matrix as follows: ; ; Step 3.2: Input the temporal attribute information and spatial attribute information into the temporal relationship extraction module to obtain the temporal information and hierarchical relationship between events; the specific process is as follows: The ONLSTM neural network controls the degree of forgetting historical information and retaining the current information through the gate structure, and distinguishes historical information from current information by dividing the level of network unit state information. In order to achieve the interval division, the model uses two integers and , which are used to represent the lowest level of historical information and the highest level of current information, respectively. In the process of model updating, when When the two intervals overlap, the overlapping part is updated using the original long short-term memory neural network LSTM. For the part below Only the currently entered information is retained. When the two intervals do not overlap, for the non-overlapping part, the current storage unit value is directly set to 0, that is, its state information is assigned an initial value, and for greater than and less than , only retaining historical information and current input information respectively. This updating process can better extract the temporal relationship between events.
[0024] Step 4: Use the training set to iteratively train the multi-dimensional attribute information fusion prediction model G-OL; the specific process is as follows: At the same time, the time and space attribute information of events in the event log are integrated for model training. Different event log trajectory prefixes are replaced for input each time training to predict the next event in the business process and realize real-time tracking and prediction in the actual business process.
[0025] In this step, the Adam optimizer is used and the regularization coefficient size is set to 5×10 -4 , the initial learning rate size is set to 1×10 -4 , thereby ensuring that the multi-dimensional attribute information fusion prediction model is trained to full convergence.
[0026] Step 5: Input the test set into the multi-dimensional attribute information fusion prediction model G-OL trained in step 4 to obtain the next event prediction result, and select the accuracy index to evaluate and analyze the accuracy of the prediction result. The accuracy index can calculate the proportion of correct classification, and the calculation formula is as follows: ; in, is the accuracy value; is the total number of samples; is the number of categories; For the The number of samples in the class; , , , Respectively The number of correctly classified positive examples, the number of incorrectly classified positive examples, the number of incorrectly classified negative examples, and the number of correctly classified negative examples in the class.
[0027] The prediction evaluation result value of the next event in the process of this method obtained in this step is 0.802.
[0028] Based on the above method, the present invention designs a business process next event prediction system based on multi-dimensional attribute information fusion. The event log is input into the system to obtain the next event prediction result as output, such as Figure 2 As shown, the system includes the following modules: The event log preprocessing module is used to analyze the direct follow-up relationship of activities in the event log, extract the time attribute, construct a data set with the process trajectory prefix as the known information and the next event as the predicted target value, and divide it into a training set and a test set; The process space feature extraction module is used to construct the graph structure of the event log based on the direct follow-up relationship of the activities. The input of the GCN graph convolutional neural network model is the graph structure, which extracts spatial attribute information. All activities in the event log trajectory and the direct follow-up activity relationships between activities are constructed as nodes and edges in the graph respectively. At the same time, the relationships between activities in all trajectories are constructed into an adjacency matrix. The relationships between activities in all trajectories include order, selection, and concurrency. The constructed adjacency matrix is used as the input of the GCN graph convolutional neural network to extract spatial attribute information. The process temporal relationship extraction module is used to extract the spatial features and time attributes of the event log as the input of the ONLSTM neural network model to extract the temporal information and hierarchical relationship between events. ONLSTM controls the degree of forgetting historical information and the degree of retaining current information through the gating structure, and distinguishes historical information from current information by dividing the hierarchy of network unit state information. In order to achieve the interval division, the model uses two integer and , which are used to represent the lowest level of historical information and the highest level of current information, respectively. In the process of model updating, when When the two intervals overlap, the overlapping part is updated using the original LSTM method. For the part below Only the currently entered information is retained. When the two intervals do not overlap, the non-overlapping part is directly set to 0, and the and less than The update process can better extract the temporal relationship between events. The multi-dimensional attribute information fusion model prediction model training module is used to iteratively train the prediction model according to the training set. The multi-dimensional attribute information fusion prediction model combines the GCN graph convolutional neural network that extracts spatial features and the ONLSTM neural network that extracts temporal relationships to predict the next event. During each training, the trajectory prefix of a different event log is replaced for input, thereby predicting the next event in the business process and realizing real-time tracking and prediction in the actual business process. The multi-dimensional attribute information fusion model prediction result evaluation module is used to input the test set into the trained multi-dimensional attribute information fusion prediction model to obtain the next event prediction result, and use the Accuracy indicator to evaluate and analyze the accuracy of the prediction result.
[0029] The present invention also designs a storage medium storing a program, which, when executed by a processor, implements the above-mentioned method for predicting the next event of a business process based on multi-dimensional attribute information fusion.
[0030] The present invention also designs a computing device, including a processor and a memory for storing a program executable by the processor, and when the processor executes the program stored in the memory, the above-mentioned model mining method for business processes containing repeated tasks is implemented. The computing device can be a desktop computer, a laptop computer, a smart phone, a PDA handheld terminal, a tablet computer, a programmable logic controller (PLC, Programmable Logic Controller), or other terminal devices with processor functions.
[0031] The present invention can make full use of the temporal and spatial feature information contained in the event log, effectively improve the accuracy of predicting the next event in the business process, and at the same time provide decision-making for process executors and prevent illegal tasks from occurring. In addition, it can be applied in multiple fields and scenarios such as finance, medical care, and education, and has practical promotion value and is worthy of promotion.
[0032] In order to prove the feasibility and superiority of the present invention, the following comparative experiment is given.
[0033] The data sets for the comparative experiment use three open source real event logs, namely: BPIC_2012_A, HospitalBilling and Prepaid_Travel_Costs. The comparison results are shown in Table 2: Table 2 Prediction comparison experimental results of the method of the present invention and the baseline method .
[0034] The prediction comparison experiment of the prediction model of the method of the present invention and three baseline methods is carried out. The three baseline methods are ATT-BiLSTM, ATT-BiGRU, and ONLSTM. The numbers in Table 2 are prediction accuracy. By comparing the results of the four methods on three event logs, it can be seen that the accuracy of the G-OL prediction model of the present invention is higher than that of the ONLSTM prediction model. This shows that in the prediction task, the fusion of multi-dimensional attribute information, i.e., time and space feature information, can effectively improve the performance of the prediction model, and the G-OL prediction model has the best prediction effect on all data sets, proving that the present invention can ensure the robustness of the prediction results of the next event in the business process. The results also show that GCN can better extract the order between the events in the business process, and can construct the spatial attribute information between all events in the event log track, thereby improving the prediction results of the next event. This proves the feasibility and superiority of the present invention.
[0035] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.
Claims
1. A business process next event prediction method based on multi-dimensional attribute information fusion, characterized in that: The steps include: Step 1: Obtain the business process event log, analyze the direct follow-up relationship of the activities in the event log, extract the time attribute information, construct a data set with the process trajectory prefix as the known information and the next event as the predicted target value, and divide it into a training set and a test set; Step 2: Construct the graph structure of the event log based on the direct follow-up relationship of the activities obtained in step 1; Step 3: construct a multi-dimensional attribute information fusion prediction model G-OL, which includes a spatial feature extraction module and a temporal relationship extraction module; Step 4: Use the training set to iteratively train the multi-dimensional attribute information fusion prediction model G-OL; Step 5: Input the test set into the multi-dimensional attribute fusion prediction model G-OL trained in step 4 to obtain the next event prediction result, and use the accuracy index to evaluate and analyze the prediction result.
2. The business process next event prediction method based on multi-dimensional attribute information fusion according to claim 1 is characterized in that: In step 1, the process trajectory prefix is the trajectory information before any activity node or time on a certain trajectory.
3. The business process next event prediction method based on multi-dimensional attribute information fusion according to claim 1 is characterized in that: In step 2, constructing the graph structure of the event log refers to constructing all activities in the event log trajectory and the direct following activity relationships between activities as nodes and edges in the graph respectively, and constructing the relationships between activities in all trajectories into an adjacency matrix. The relationships between activities in all trajectories include sequential, selection and concurrent relationships.
4. The business process next event prediction method based on multi-dimensional attribute information fusion according to claim 1 is characterized in that: In the step 3, the spatial feature extraction module is composed of a two-layer GCN graph convolutional neural network, which is a neural network structure used to extract graph structure features; the temporal relationship extraction module adopts the ONLSTM neural network, which is a variant of the recurrent neural network. The core of the ONLSTM neural network is to allow the network to dynamically divide neurons into different levels of granularity to help capture long-term dependencies and hierarchical information in the sequence.
5. The business process next event prediction method based on multi-dimensional attribute information fusion according to claim 1 is characterized in that: In step 3, the specific working process of the multi-dimensional attribute information fusion prediction model G-OL is as follows: Step 3.1: Input the graph structure of the event log into the spatial feature extraction module to extract spatial attribute information; the forward propagation calculation formula of the spatial feature extraction module is as follows: ; In the formula, It is the output matrix of the spatial feature extraction module; is the activation function; is the node feature matrix; is the adjacency matrix of the trajectory; is the softmax activation function; is the sum of the adjacency matrix and the identity matrix of the trajectory, that is , is the identity matrix; is the ReLU activation function; and They are the parameter matrices of the two-layer GCN graph convolutional neural network; ReLU and softmax are the activation functions of the two-layer GCN graph convolutional neural network; Step 3.2: Input the temporal attribute information and spatial attribute information into the temporal relationship extraction module to obtain the temporal information and hierarchical relationship between events; the specific process is as follows: The ONLSTM neural network controls the degree of forgetting historical information and retaining the current information through the gate structure, and distinguishes historical information from current information by dividing the level of network unit state information. In order to achieve the interval division, the model uses two integers and , which are used to represent the lowest level of historical information and the highest level of current information, respectively. In the process of model updating, when When the two intervals overlap, the overlapping part is updated using the original long short-term memory neural network LSTM. For the part below Only the currently entered information is retained. When the two intervals do not overlap, for the non-overlapping part, the current storage unit value is directly set to 0, and for the part greater than and less than of the components, respectively retaining only historical information and current input information.
6. The method for predicting the next event of a business process based on multi-dimensional attribute information fusion according to claim 5, characterized in that: In step 4, the Adam optimizer is used, the size of the regularization term coefficient and the size of the initial learning rate are pre-set, and the time and space attribute information of the events in the event log are integrated for model training. During each training, the trajectory prefix of a different event log is replaced for input, so as to predict the next event in the business process and realize real-time tracking and prediction in the actual business process.
7. The method for predicting the next event of a business process based on multi-dimensional attribute information fusion according to claim 1, characterized in that: In step 5, the calculation formula of the accuracy index is as follows: ; in, is the accuracy value; is the total number of samples; is the number of categories; For the The number of samples in the class; , , , Respectively The number of correctly classified positive examples, the number of incorrectly classified positive examples, the number of incorrectly classified negative examples, and the number of correctly classified negative examples in the class.
Citation Information
Patent Citations
Method and system for predicting next event of process instance based on log sampling
CN115328866A
Multi-modal business process prediction method and system fusing flow chart and process log
CN118134232A
Graph convolutional neural network and knowledge base-based paper correlation analysis method
WO2023155508A1
Cited By
Cooperative monitoring system and method for digital transformation process
CN119398720A
A collaborative monitoring system and method for digital transformation process
CN119398720B
