Business process prediction method based on multi-scale feature fusion and graph time sequence modeling

Abstract

This invention belongs to the field of business process prediction technology and discloses a business process prediction method based on multi-scale feature fusion and graph temporal modeling. This invention introduces edge features on top of the original node features, constructing a dedicated edge feature vector for each pair of events to address the relationships between events in a loan business process, enabling the model to accurately characterize the dynamic structure of the process. This invention constructs a multi-scale feature fusion model based on CRA-Unet to fully utilize the multi-dimensional feature information in the event log input. CRA, based on an attention mechanism, enhances the model's feature fusion and representation capabilities. This invention also designs a fusion model based on a combination of graph isomorphic networks and gated recurrent units, fully utilizing graph structure information and temporal features of the time log, enabling simultaneous modeling of dependencies and temporal evolution patterns between events, thus improving the accuracy and stability of future activity predictions.

Description

Technical Field

[0001] This invention belongs to the field of business process prediction technology, and specifically relates to a business process prediction method based on multi-scale feature fusion and graph time series modeling, which is particularly suitable for predicting future activities in loan business processes. Background Technology

[0002] With the advancement of digital transformation, process mining technology has been widely applied in various industries, especially in loan business processes. Predictive Process Monitoring (PPM) has become a core research direction in this field, aiming to predict the future development of business processes based on historical or ongoing events, such as the next activity, timestamps, and remaining time in a loan business process. Furthermore, in the financial sector, business processes are highly complex and involve multiple objects interacting. For example, an application event often involves multiple objects (customers, accounts, approvers, guarantee contracts, etc.), and the temporal sequences and dependencies between events are complex and diverse. Existing predictive process monitoring technologies are mainly based on two types of implementation schemes: one is a process model-centric approach, and the other is a sequence modeling method based on machine learning or deep learning.

[0003] While traditional process modeling methods can incorporate process trees, Petri nets, and other models for prediction, they struggle to guarantee accuracy in practical applications. With the development of deep learning, prediction models based on recurrent neural networks and graph convolutional networks have been widely applied and have achieved good results on traditional event logs. However, these methods generally only consider linear event logs and fail to adequately account for the multi-object and multi-interaction features present in real-world processes. Therefore, traditional prediction methods based on linear event logs struggle to effectively capture the multidimensional relationships between these objects, resulting in insufficient prediction accuracy.

[0004] To address the aforementioned issues, Object-Centric Process Mining (OCPM) has been proposed. This method, by incorporating interaction information between objects, can more realistically describe actual business processes. However, to apply OCPM event logs to predictive process monitoring, existing methods typically convert them into traditional event logs. This flattening process can easily lead to the loss of critical information. In recent years, some methods have attempted to use event graphs to represent object interactions to retain more semantic information. While these methods introduce object interaction information to some extent, the heterogeneity and complexity of OCPM event logs, involving multiple object types and their interaction relationships, mean that preserving both graph structure features and temporal characteristics during modeling remains a technical challenge in this field.

[0005] In summary, existing methods still have the following problems and shortcomings: 1. Insufficient utilization of edge features: Existing methods generally only focus on node or object attributes, ignoring edge features such as dependencies in the event graph (e.g., transaction delays, account transfer links, approval order), time intervals, and resource coordination), resulting in incomplete prediction information. 2. Limited feature fusion methods: Although some methods fuse different features, in the loan business process, the fusion of multi-source features (customer information, transaction time, account behavior, etc.) lacks a multi-scale mechanism, making it difficult to fully utilize their complementary information. 3. The loan business process includes both strict approval hierarchical relationships and non-linear time dynamics, and existing models have failed to effectively combine these two types of information, leading to unstable prediction results.

[0006] In summary, existing loan business process prediction models typically only utilize node features, neglecting the importance of edge features. Furthermore, financial business processes often exhibit strong time dependence and nonlinear characteristics, such as delays in fund transfers and bottlenecks in the approval process, limiting the effectiveness of existing methods in predicting the next transaction, approval time, or risk status. While some methods introduce feature fusion mechanisms, they suffer from limitations in fusion methods and insufficient utilization; others focus only on graph structure information or temporal information without effectively combining both. These shortcomings restrict further improvements in model predictive performance. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes a business process prediction method based on multi-scale feature fusion and graph temporal modeling, in order to achieve accurate prediction of future activity types in business processes.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A business process prediction method based on multi-scale feature fusion and graph temporal modeling includes the following steps:

[0010] Step 1. Obtain the object-centric raw event log from the historical loan business process, and transform the event log into a process execution graph by extracting the process execution; the process execution graph consists of nodes and edges between nodes; the attribute information of nodes and the attribute information of edges are stored in the node feature matrix and the edge feature matrix respectively, and are initially empty;

[0011] Step 2. Process the event logs using edge and node feature extraction functions respectively, and encode the obtained features into node features and edge features respectively, and then add them to the empty process execution graph;

[0012] We enhanced edge and node features using a multi-scale feature fusion model based on CRA-Unet and a k-means clustering algorithm, respectively; and used the process execution graph containing node and edge features, as well as the node feature matrix, as the dataset.

[0013] Step 3. Build a time graph prediction model GTPM that can simultaneously process graph structure information and log time series information. Its input is the process execution graph and the node feature matrix, i.e., sequence features. The model output is the next activity in the process execution.

[0014] The model is trained based on the dataset obtained in step 2, and the trained model is used to predict activities.

[0015] Furthermore, based on the business process prediction method based on multi-scale feature fusion and graph temporal modeling, this invention also proposes a corresponding business process prediction system based on multi-scale feature fusion and graph temporal modeling, the technical solution of which is as follows:

[0016] A business process prediction system based on multi-scale feature fusion and graph temporal modeling includes the following modules:

[0017] The preprocessing module is used to obtain object-centric raw event logs from historical loan business processes and transform the event logs into process execution graphs by extracting process execution. The process execution graph consists of nodes and edges between nodes. The attribute information of nodes and the attribute information of edges are stored in the node feature matrix and the edge feature matrix respectively, and are initially empty.

[0018] The feature extraction module is used to process event logs using edge and node feature extraction functions respectively, and encodes the obtained features into node features and edge features respectively, and then adds them to an empty process execution graph.

[0019] We enhanced edge and node features using a multi-scale feature fusion model based on CRA-Unet and a k-means clustering algorithm, respectively; and used the process execution graph containing node and edge features, as well as the node feature matrix, as the dataset.

[0020] The system also includes a prediction module, which is used to build a time graph prediction model (GTPM) that can simultaneously process graph structure information and log time series information. Its input is the process execution graph and the node feature matrix, i.e., sequence features. The model output is the next activity in the process execution. The model is trained based on the dataset and the trained model is used to predict activities.

[0021] The present invention has the following advantages:

[0022] As described above, this invention discloses a business process prediction method based on multi-scale feature fusion and graph temporal modeling. This invention introduces an edge feature modeling mechanism on top of the original node features, further expanding the expressive power of flowcharts in loan business process scenarios. For the relationships between events such as transactions, approvals, and fund transfers in the loan business process, a dedicated edge feature vector is constructed for each pair of events. This enables the model to not only learn the context of a single event but also capture the dependency patterns and interaction rules between transactions, thereby more accurately depicting the dynamic structure in the loan business process. Furthermore, this invention constructs a multi-scale feature fusion model based on CRA-Unet to fully utilize the multi-dimensional feature information in the event log input. The CRA module, based on an attention mechanism, filters out financial feature channels with higher weights through channel reconstruction, focusing on key event patterns (such as high-risk transaction chains or abnormal approval sequences), effectively improving the model's feature fusion and representation capabilities. In addition, this invention designs a fusion model GTPM based on a combination of Graph Isomorphic Network (GIN) and Gated Recurrent Unit (GRU), fully utilizing the graph structure information of the loan application process model and the temporal features of the time log. This mechanism can simultaneously model the dependencies and temporal evolution patterns between events, thereby significantly improving the accuracy and stability of predictions in loan business process tasks. Attached Figure Description

[0023] Figure 1 This is a flowchart of the business process prediction method based on multi-scale feature fusion and graph temporal modeling in an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of the object diagram in an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of the process execution in an embodiment of the present invention;

[0026] Figure 4 This is a network structure diagram of CRA-UNet in an embodiment of the present invention;

[0027] Figure 5 This is a network structure diagram of the CRA module structure in an embodiment of the present invention;

[0028] Figure 6 This is a network structure diagram of the Time-Graph Prediction Model (GTPM) in an embodiment of the present invention.

[0029] Figure 7 This is a schematic diagram of the experimental results of the α hyperparameter in a specific example of the present invention;

[0030] Figure 8The following are prediction accuracy curves of datasets with different prefix lengths in specific examples of the present invention; (a), (b), (c), and (d) are prediction accuracy curves of datasets BPI-2017, BPI2017C, BPI2017E, and BPI2017H, respectively.

[0031] Figure 9 This is a schematic diagram of the significance test results in a specific example of the present invention;

[0032] Figure 10 This is a schematic diagram of an ablation experiment in a specific example of the present invention. Detailed Implementation

[0033] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0034] Example 1

[0035] To address the shortcomings of existing loan business processes in terms of process complexity, time dependence, and object interactivity, particularly the problems of information loss, insufficient utilization of edge features, limited feature fusion methods, and ineffective integration of graph structure and temporal information in object-centric event log modeling, this invention proposes a business process prediction method based on multi-scale feature fusion and graph temporal modeling. This method, by combining graph structure modeling with temporal feature fusion, can accurately predict future activity types in the loan business process, thereby providing data support for risk control and business optimization. After processing by this invention, the semantic information in the event log can be more comprehensively preserved and utilized, thus improving prediction performance.

[0036] like Figure 1 As shown, the business process prediction method based on multi-scale feature fusion and graph temporal modeling in this embodiment mainly includes the following three steps: 1. Event log processing: This step, as the starting step, mainly focuses on the acquisition and transformation of event logs. Event logs can come from the credit business process system in a financial scenario, and then the algorithm transforms them into the process execution graph required for subsequent steps. 2. Feature engineering: The input of this step is the process execution graph extracted in step 1 and the original event logs. Various feature extraction algorithms are used to extract and encode features from the event logs. The encoded features are added to the above process execution graph to form a feature-enhanced process execution graph used for subsequent training. Different methods are used to enhance the edge features and node features respectively. 3. Building the prediction model GTPM: The feature-enhanced process execution graph and the node feature matrix, i.e., the sequence features, are combined to form a dataset for the following model training. First, it is divided into a training set and a test set for model training and testing, and the trained model is used to predict future activities.

[0037] Specifically, the business process prediction method based on multi-scale feature fusion and graph temporal modeling includes the following steps:

[0038] Step 1. Obtain the original object-centric event logs from the historical loan business process, and transform the event logs into a process execution graph by extracting process execution. The process execution graph consists of nodes and edges between nodes; the attribute information of nodes and edges is stored in the node feature matrix and edge feature matrix respectively, and is initially empty.

[0039] This embodiment uses a real loan process log from a Dutch financial institution, BPI2017, which can be downloaded directly from https: / / data.4tu.nl / datasets. Step 1 retrieves the process execution graph from the object center event log. The input is the raw event log, and the following method is used to output a feature-free process execution graph.

[0040] In this embodiment, the object-centric event log is similar to the traditional event log, both consisting of events. The difference lies in that each event can be associated with one or more objects. As shown in Table 1, each line in the event log represents an event activity, and each event activity is associated with one or more of the following: applicant, loan offer, and approver.

[0041] Table 1 Event Log

[0042]

[0043] In this way, object-centric event logs can help analyze logs from three different perspectives.

[0044] From the applicant's perspective, the entire loan process only includes creating and submitting the completed application, as well as the subsequent return for modification. It does not include processes such as creating a loan offer. Similarly, from the approver's perspective, it only includes processes such as accepting the loan and creating a loan offer. Furthermore, some approval stages may involve multiple reviewers.

[0045] The method for converting event logs into a process execution graph by extracting process execution is as follows:

[0046] I. Based on the event log, an undirected graph, or object graph, is obtained, consisting of a set of interconnected objects. In the object graph, nodes represent different objects, and edges represent events connecting two objects.

[0047] Events that connect two objects, for example Figure 2 The events are e4, e7, etc. The event logs shown in Table 1 can be used to obtain the following results: Figure 2The object diagram shown contains five objects and five bridging events.

[0048] in Figure 2 In this code, o1, a1, w1, w2, and w3 represent objects, and e4, e5, e6, e7, and e8 represent bridging events.

[0049] II. In this embodiment, each process execution diagram is obtained from an object diagram.

[0050] Specifically, this involves identifying the events associated with each object in the object graph and the trajectories of these events. These events are then used as nodes in the execution flow diagram, and the order of these trajectories is used as edges to form the final execution flow diagram. For example... Figure 3 As shown.

[0051] Step 2. Process the event logs using edge and node feature extraction functions respectively, and encode the obtained features into node features and edge features respectively, and then add them to the empty process execution graph.

[0052] We enhanced edge and node features using a multi-scale feature fusion model based on the channel-refined attention network CRA-Unet and the k-means clustering algorithm, respectively. The execution graph containing node and edge features, along with the node feature matrix (i.e., sequence features), were used as the dataset for training the model. The sequence features include the corresponding real activity, appearing as the activity name of the current node; these sequence features only contain node features and lack edge features and graph structure.

[0053] In simple terms, a sequence feature is a sequence that contains events and their features, and it has a temporal order.

[0054] The input here is the empty process execution graph from step 1 and the original event log. Feature extraction is performed on the event log to obtain vector representations related to activities, times, objects, and resources in the event log. At the same time, the event relationship graph corresponding to the specified process instance is extracted from the original event log, and the edges in the process execution graph are generated and assigned corresponding feature values.

[0055] A process instance is a process, which is a transitional form before generating a process execution diagram. The process execution diagram is transformed from this sequence of process instances. The process instance is a part of the original event log, which is extracted to form the process execution diagram.

[0056] For the vector representation of the obtained activity-related features, the corresponding node features are obtained by using node feature encoding.

[0057] The features include activity-related features, such as the name of the previous activity; time-related features, such as the timestamp of the current activity; and resource-related features, such as the loan amount. For the vector representation of the node-pair features, the corresponding edge features are obtained using edge feature encoding. Here, node-pair features include, for example, whether two nodes fall on the same day.

[0058] The node features and edge features are assigned to the node feature matrix and the edge feature matrix respectively to form a flow execution graph with features. The edge features and node features are then enhanced using the CRA-Unet network and the k-means clustering algorithm respectively.

[0059] This invention designs a corresponding edge feature processing algorithm for edge features, aiming to extract the event relationship graph corresponding to a specified process instance from the event log, and generate and assign corresponding feature values ​​to the edges therein.

[0060] The process of processing event logs using edge feature extraction functions is as follows:

[0061] First, based on the given process instance number, extract the subgraph corresponding to the process instance from the event log. The subgraph contains all event nodes related to the process instance and the edges connecting them to each other.

[0062] For example, a loan log has 1-10, a total of ten events, among which:

[0063] Numbers 1-5 are the first process, the first example of a process, which is A's loan process; numbers 6-8 are the second example of a process, which is B's loan process; numbers 9-10 are the third, and finally C's loan process.

[0064] Then, feature calculations are performed on each edge in the subgraph sequentially.

[0065] For each edge, the predefined set of edge feature processing functions is traversed sequentially. With the event log as the input parameter, the corresponding feature function is automatically called to calculate different edge attributes. After the calculation is completed, the edge features of all edges are obtained.

[0066] For each edge, the predefined set of edge feature processing functions is traversed sequentially to calculate different edge attributes, such as the time interval between events, the difference in application amount, the change in credit score level, or whether the evaluation process has timed out.

[0067] By assigning the feature values ​​obtained from edge feature extraction to the corresponding edges, an edge data structure with attributes is formed. Through the aforementioned special edge feature processing algorithm, feature maps with rich structural semantics can be effectively constructed.

[0068] The edge feature processing function defined in this embodiment is as follows:

[0069] The first edge feature processing function is used to calculate the absolute time difference between two endpoint events in a loan business process, such as the time difference between the issuance of a loan application and the acceptance of the application in a loan process.

[0070] First, let e=(s,t) represent an edge, where s and t are the source node and the target node, respectively. Let T(s) and T(t) represent the event timestamps of the source node s and the target node t, respectively. The absolute time difference is obtained by subtracting them.

[0071] If the timestamp is of type datetime, you can directly subtract it and use the total_seconds() method to get the number of seconds; if it is of numeric type, you can directly subtract it to get the number of seconds.

[0072] The second edge feature processing function is used to determine whether the events of the two endpoints both fall on Saturday or Sunday, and returns a boolean value (true if the timestamps of both endpoints are both on the weekend), indicating whether the two events occur on Saturday or Sunday at the same time.

[0073] Similarly, in a loan business process, there are two events: applying for a loan and approving the application. The function of the second edge feature processing function is to determine whether these two events occur simultaneously on Saturday or Sunday.

[0074] Let e ​​= (s, t) represent an edge, where s and t are the source node and the target node, respectively. Let D(s) and D(t) represent the event dates of the source node s and the target node t, respectively. D(s) and D(t) are obtained using the day_of_week function, as shown in the following formula:

[0075] The third edge feature processing function is the function for determining the activity transition type. It is mainly used to determine the activity types of the two nodes associated with the current edge, and uses one-hot encoding to assign a vector to each association type.

[0076] For example, if the two node activities associated with the current edge are creating an order and accepting an offer, then the vector assigned to them is (1, 0, 0, 0). If they are accepting an offer and creating a loan offer, then the vector assigned to them is (0, 0, 1, 0).

[0077] The specific processing procedure is as follows: Let e=(s,t) represent an edge, where s and t are the source node and the target node, respectively. Check the activity type of s and t, and then match them with the one-hot dictionary to obtain the feature value of the edge.

[0078] The fourth edge feature processing function determines whether there is a timeout in the activities of the two nodes associated with the current edge.

[0079] The function returns a boolean value based on whether a timeout occurred (1 if the timeout occurred, 0 otherwise). Specifically, it calculates the time difference between the source and target nodes and compares it to a time threshold set by the function to obtain a feature value.

[0080] The fifth edge feature processing function determines whether the application amount has changed. It mainly uses the required loan amount field in the loan log. Similarly, it needs to judge the difference between two nodes. If the difference is 0, it returns a boolean value of 0, indicating that the loan amount has not changed. If the difference is not 0, it returns a boolean value of 1, indicating that the loan amount has changed.

[0081] The sixth feature processing function checks whether the credit score has changed. Its purpose is to detect and monitor fluctuations in the risk control score. The processing method is the same as the function above: it returns 1 if the score changes and 0 if it does not.

[0082] The processing of node features is similar to that of edge features. A set of node feature functions can be defined, and the node feature values ​​are calculated using these functions. These node feature functions are fairly standard and will not be elaborated upon further. For the subgraph obtained from the event log, the feature values ​​are calculated for each node sequentially using the set of feature functions.

[0083] The node features calculated using the node feature function include, for example, the following features: unique event identifier, event activity name, loan amount involved in the event, event service time, event execution duration, time elapsed from the start of the process to the current event, remaining time after the event execution, time difference between the current event and its direct predecessor event, maximum time interval between the predecessor events of the current event, waiting time of the current event, delay time of the current event, activity sequence before the event, number of times the block application activity has occurred before, and aggregation of loan amounts of predecessor events.

[0084] First, input the event log and the set of node feature functions. Then, iterate through each node feature function, calculate the feature value of each node feature and assign it to the node. Finally, return a process execution subgraph containing the node features.

[0085] Further, the results of edge feature and node feature extraction are encoded separately to obtain vector representations of the features. The node features and edge features are then assigned to the node feature matrix and the edge feature matrix, respectively.

[0086] After completing the feature calculations for all edges, an edge feature matrix is ​​constructed based on the attribute information of each edge.

[0087] The specific approach is as follows: read all the feature values ​​of each edge in sequence, form them into a row vector, and add them to the edge feature matrix in sequence. This matrix can comprehensively reflect the multidimensional relationship information between events in the process instance.

[0088] Finally, a complete process execution feature graph structure containing edge feature information and node feature information is obtained. At this time, the shape of the process execution graph is (G, F), where G contains graph structure information and F contains a matrix of edge features and node features.

[0089] The feature map of this process execution not only preserves the structural connections between events, but also contains a variety of quantifiable interaction features, providing high-quality input data for subsequent graph neural network training and financial process prediction tasks.

[0090] This invention further enhances the features of edge features and node features. Based on the existing original features, feature enhancement generates more new features that are helpful for model learning or enhances the expressive power of the original features.

[0091] Considering that the node features are relatively distinct and can be roughly divided into three categories: activity-related, time-related, and resource-related features, we use clustering to perform a simple feature enhancement for the node features.

[0092] Here, we choose the k-means clustering algorithm with k=3 and use the clustering results as new features for subsequent training.

[0093] For the edge feature branches, the CRA-Unet model proposed in this invention was used for multi-scale feature fusion.

[0094] like Figure 4 As shown, the CRA-Unet network consists of four encoder stages and four decoder stages; the initial input of the CRA-Unet network is the edge features (B, C, L) in the process execution feature graph obtained by the previous feature encoding.

[0095] Where B, C, and L represent the batch size, number of channels, and number of edges of the feature graph executed in each process, respectively.

[0096] In the first and second encoder stages, the Channel Reconstruction Attention (CRA) module is used in conjunction with a downsampling operation to compress the length of the edges, so as to highlight the features that have a greater impact on the prediction task and suppress redundant information.

[0097] Considering that the CRA module has the function of channel reconstruction, and in order to keep the model lightweight, the number of channels remains consistent in each subsequent layer (referring to each layer of the encoder and decoder). After the first and second encoder stages, the number of edges becomes half of the original, and the overall sequence feature map structure becomes (B, C, L / 4).

[0098] In the third encoder stage, a simplified linear attention module (SLA) is used, and a downsampling operation is performed to further compress the side length, balancing computational efficiency and modeling capability. The SLA module uses a three-piece set of "ReLU instead of softmax + discarding normalization + using convolution to fill local gaps", thus having lower complexity than the traditional linear attention mechanism.

[0099] Therefore, for the input features, we use the ReLU dot product of Q and K to obtain the non-normalized attention weights, then use them to sum the weights over V to obtain the output, and finally use lightweight convolution to enhance locality.

[0100] After the third and fourth encoder stages, the number of edges is halved again, and the overall sequence feature map structure becomes (B, C, L / 8).

[0101] At the bottleneck layer, a multi-scale linear attention module is set up. Therefore, this layer can be used to extract higher-level semantic features, providing rich contextual information for the decoding stage, while the sequence feature map structure remains unchanged in this layer.

[0102] The multi-scale linear attention module uses Mamba linear attention (MLLA). MLLA and SLA are both linear attention mechanisms, but MLLA adds a dynamic gating mechanism to preserve long sequence information.

[0103] In the first and second decoder stages, the channel reconstruction attention module (CRA) is used in conjunction with a linear interpolation operation (LI) to upsample the edges and gradually restore their lengths to their original size.

[0104] In the third decoder stage, a simplified linear attention module (SLA) is used in conjunction with a linear interpolation operation (LI) to continue upsampling; in the fourth decoder stage, a simplified linear attention module (SLA) is used to gradually restore the complete structure of the feature map.

[0105] Set up skip connections between the corresponding stages of the encoder and the corresponding stages of the decoder.

[0106] This invention enhances the fine-grained representation of features and contextual consistency by setting a skip connection between the corresponding layers of the decoder and encoder, thereby fusing low-level local features and high-level semantic features.

[0107] Set up skip connections between the corresponding stages of the encoder and the corresponding stages of the decoder.

[0108] The above design achieves a balance between computational efficiency and model complexity, enhancing the model's generalization ability while maintaining high prediction accuracy. It is suitable for process execution log modeling scenarios involving long time spans and complex dependencies.

[0109] For an event log in a loan business process, whether it is a traditional event log or an object-centric event log, it exhibits obvious time-series and trend characteristics, and different characteristics have different impacts on different prediction tasks.

[0110] For example, in predicting the next activity, activity-related features have a significant impact on the final prediction result.

[0111] This invention employs a channel partitioning and reconstruction method that more effectively enhances channel representation and provides a higher-quality input for the subsequent attention mechanism. Through the channel attention mechanism, the model's ability to focus on key channels can be further improved without introducing additional parameters or computational overhead, leading to better selection of effective features.

[0112] Based on this, the present invention designs a Channel Reconstruction Attention Module (CRA), such as... Figure 5 As shown, the processing flow is as follows:

[0113] A dual-path structure was chosen at the beginning of this structure:

[0114] One line represents enriched features, and the other represents supplementary features, distinguished by a hyperparameter α. However, this structure focuses more on feature enhancement, but the fusion method is relatively simple and may not fully utilize all the information.

[0115] Therefore, in the subsequent process, this invention incorporates a local + global attention weighted fusion mechanism to guide the model to use the softmax function to calculate weights, dynamically balance the contributions of global and local information, and prevent some information from being completely overwhelmed.

[0116] Specifically, when the sequence feature map of input (B,C,L) enters the CRA module, it is processed as follows:

[0117] First, preliminary feature extraction is performed using one-dimensional convolution Conv1d. The extracted features are divided into two branches, with the number of channels divided into enriched channels according to ratios α and 1-α. and supplementary channels The number of channels are (1-α)C and αC, respectively.

[0118] During the initial feature extraction using the one-dimensional convolution Conv1d, the number of channels remains constant. Then, one-dimensional convolutions are performed on both types of channels to compress them, yielding key channel feature representations P1 and P2, which reduce the computational burden of subsequent convolutions.

[0119] Then, group convolution (GW Conv1d) and point convolution (PW Conv1d) are performed on the enriched channel key feature representation P1, respectively, and the number of channels is restored to C. Finally, the results are summed to obtain the enriched channel representation. .

[0120] On the other hand, point convolution is applied to the key feature representation P2 of the supplementary channel, and the number of channels is adjusted to αC. Then, it is concatenated with the input of the supplementary channel to obtain the supplementary channel representation. .

[0121] Then and The reconstructed features are obtained by adding elements one by one. The input sequence has C channels; in this step, the channels of the input sequence are structurally reconstructed, enhancing cross-channel expressive power. The above process can be represented as:

[0122] ;

[0123] ;

[0124] ;

[0125] ;

[0126] in , , 1×1 convolution, 1×1 group convolution, and 1×1 point convolution, respectively.

[0127] Next, the reconstructed features obtained from the channel reconstruction... (With C channels), the following further processing is performed:

[0128] On the one hand, the reconstructed features Global average pooling is performed to obtain the aggregated representation for each channel. The result is then processed by two 1×1 point convolutions and a ReLU activation function to extract global context information. .

[0129] On the other hand, regarding the reconstructed features Local context information can be extracted directly using two 1×1 point convolutions. .

[0130] Finally, the information extracted from the two aspects above is integrated to obtain... The corresponding attention weights are obtained using the softmax function. During this process, the number of channels remains unchanged, as shown in the following formula:

[0131] ;

[0132] ;

[0133] ;

[0134] ;

[0135] in Indicates a fully connected layer. Indicates the channel index. This indicates the global information fusion path after channel enhancement. This indicates the local information fusion path after channel enhancement. Presentation layer normalization method.

[0136] Finally, assign each attention weight Represented by the corresponding channels Multiply them, then add them together.

[0137] Finally, a 1×1 convolution is used to further promote feature fusion, resulting in fused features.

[0138] The final output of CRA can be expressed as: .

[0139] Therefore, by first enhancing the channels and then fusing global and local dual-path information, CRA achieves more effective feature selection and fusion in temporal prediction compared to standard convolution or simple attention, providing higher quality input for subsequent training.

[0140] Step 3. Build a time graph prediction model GTPM that can simultaneously process graph structure information and log time series information. Its input is the process execution graph and the node feature matrix, i.e., sequence features. The model output is the next activity in the process execution.

[0141] like Figure 6 As shown, the Time Map Prediction Model (GTPM) in this embodiment consists of two parallel structures:

[0142] One is used to process graph structure information, based on Graph Isomorphism Network (GIN); the other is used to process temporal information, employing Bidirectional Gated Recurrent Unit (Bi-GRU).

[0143] Specifically, the first branch structure is a graph structure processing branch, whose input is a process execution graph containing node features and edge features. It processes structural information based on graph isomorphic networks and obtains graph structure information features.

[0144] Graph structure processing branches consist of multiple graph convolutional blocks, such as... Figure 6 As shown.

[0145] Each graph convolution module contains a graph neural network, a normalization method, and an activation function.

[0146] In this branch, the input graph structure F_g(G,F) will pass through two graph convolution modules.

[0147] In the feature update process of graph neural networks, edge features are fused with the features of adjacent nodes. The fused node and edge features are then aggregated by averaging, and finally transformed nonlinearly through a multilayer perceptron to generate new node features. This path extracts information about the topology, nodes, and edges of the graph, enriching the expressive power of the graph structure.

[0148] ; ; ; .

[0149] Norm(i) represents the four normalization operations. Represents the graph features of the input. This invention employs a learnable weighted graph normalization method that balances the weights of four normalization methods. To improve training stability and feature representation capabilities, this invention utilizes a hybrid weighted graph normalization method.

[0150] This method comprehensively considers four normalization schemes: node, adjacency, graph level, and batch, and performs weighted fusion through learned shared weights, thereby enhancing the ability to capture local and global structural information while ensuring stability.

[0151] Graph structure processing branches effectively extract graph structure and node feature information from the output.

[0152] Of course, in this embodiment, the graph structure information extraction module can not only use graph convolutional networks (GCN), but also graph attention networks (GAT) or GraphSAGE, to adapt to the graph modeling needs in different scenarios.

[0153] The second branch structure is the time-series structure processing branch, whose input is the node feature matrix, i.e., the sequence features, which is used to process the time information using a bidirectional gated recurrent unit (Bi-GRU) to obtain the time-series information features.

[0154] The timing structure processing branch uses a bidirectional gated cyclic unit (Bi-GRU) to model the timing information of the event log. Considering that the bidirectional structure would double the output dimension, a linear transformation layer is added to this branch to restore the output dimension to its original length.

[0155] The input of the temporal structure processing branch only utilizes node features (B,C,L) because the temporal nature of node features is the most significant. The input first passes through a Bi-GRU, which doubles the number of channels, and then passes through a linear layer to restore the number of channels to C.

[0156] Among them, the number of stacked channels for graph structure processing branches and time sequence structure processing branches becomes 2×C.

[0157] This design effectively compensates for the shortcomings of graph neural networks in capturing temporal features.

[0158] ;

[0159] ;

[0160] ;

[0161] ;

[0162] Where z represents the gate weight matrix, ⊙ represents element-wise multiplication, and GC(·) represents the graph convolution operation.

[0163] T represents the input sequence features. The inverse form representing sequence features; This represents the result of GRU processing of sequence features; This represents the sigmoid function, used to obtain α, which represents the attention weights.

[0164] F represents the final feature after the weighted fusion of the two paths.

[0165] After feature extraction is completed in the two branch structures, the output features of the two branch structures are first concatenated using a Concat function. Then, the concatenated features are input into the gated fusion block. This gated fusion block consists of a linear layer and a sigmoid function, which is used to dynamically adjust the flow ratio of graph structure information and temporal information to highlight key information.

[0166] Of course, in addition to the gating mechanism consisting of linear layers and the softmax function, it can also be replaced by a multi-head self-attention mechanism, gated convolutional units, or weighted averaging methods to achieve effective integration of multi-source feature information.

[0167] The obtained graph structure information and time series information flow ratios are multiplied by the corresponding branch structure outputs to obtain the first multiplication result and the second multiplication result. Then, the two multiplication results are added together and the prediction result is output through the dense connection layer.

[0168] Of course, in this embodiment, in addition to using a bidirectional gated recurrent unit (Bi-GRU), the temporal modeling module can also use a long short-term memory network (LSTM), a temporal convolutional network (TCN), or a sequence modeling method based on Transformer.

[0169] These time series models can also further enhance the ability to capture long sequence dependencies.

[0170] The GTPM model proposed in this invention, through the aforementioned parallel structure and fusion mechanism, can simultaneously capture the dependencies of the graph structure and the temporal characteristics of the event sequence, thereby improving the accuracy and robustness of future activity prediction in the loan business process.

[0171] In the loan business process, if the next activity can be accurately predicted, banks can arrange resources in advance, reduce delays, monitor risks, improve customer experience, and achieve intelligent and controllable processes.

[0172] The model is trained based on the dataset obtained in step 2, and the trained model is used to predict activities.

[0173] The training process of the GTPM model can use cross-entropy loss, mean squared error loss, or contrastive learning loss as the objective function, and optimizers such as Adam, SGD, or RMSProp can be selected to improve the convergence speed and stability of the model.

[0174] After the model is trained, it is deployed in practice. The specific reasoning process is as follows:

[0175] The collected loan process event logs are first transformed into a process execution graph containing only structural information using an algorithm. Then, information such as activities and resources in the event logs are extracted as features and added to the process execution graph. A targeted edge feature construction algorithm is also designed to extract edge-related features using an edge feature function and add them to the process execution graph.

[0176] The generated execution graph at this point already contains structural information, node feature matrices, and edge feature matrices. The model then fuses and enhances these features, using the enhanced features as input for the model's inference phase.

[0177] In actual deployment, when the model receives a new process prefix or event sequence, it will calculate the probability of occurrence of each candidate next event based on the current features and output the event with the highest probability as the prediction result.

[0178] The results can be used for real-time monitoring, process decision-making, or risk warning without relying on labels or manual comparison.

[0179] Compared with conventional methods, the present invention has advantages in at least the following aspects:

[0180] (1) Feature space expansion and edge feature utilization:

[0181] This invention applies edge features to business process event log modeling, supplementing the traditional methods that do not consider edge features as a factor, and node features cannot reflect transaction relationships and dependency semantics. This enables the model to capture structural information such as account interactions and approval chains, significantly improving the accuracy of process prediction compared to object-centered prediction methods.

[0182] (2) Enhanced multi-scale feature fusion capability:

[0183] This invention proposes a CRA-Unet model, which utilizes an encoder-decoder structure to extract multi-scale features of local transaction patterns and global dependencies in financial transactions. Furthermore, by reconstructing the attention module CRA through a designed channel, the CRA-Unet model can integrate key financial features, thereby improving its robustness in modeling complex business processes.

[0184] (3) Effective supplementation of time information:

[0185] The GTPM model of this invention uses a two-way GRU to capture the temporal dependencies in the transaction and approval process, which makes up for the shortcomings of relying solely on structural or static features, and enables the model to more accurately reflect the dynamic changes in capital flow and risk indicators.

[0186] (4) Balancing local and global information:

[0187] The CRA module dynamically balances local and global features through a channel attention mechanism, enabling the model to identify local anomalies (such as anomalies in a single transaction) while maintaining a global awareness of the overall financial process trend, thus improving the stability of predictions.

[0188] (5) Computational efficiency and scalability:

[0189] This invention improves prediction accuracy while controlling computational costs through modular design and lightweight operations (such as 1×1 convolution and residual connection), and can efficiently process large-scale financial logs and is suitable for real-time monitoring and risk warning scenarios.

[0190] Furthermore, to verify the effectiveness of the method of this invention, experiments were conducted on seven object-centered datasets to predict the next activity, and the results were analyzed. All experiments were performed in Python 3.8 and the PyTorch 2.1.0 framework.

[0191] The five datasets used in the experiment included a real loan process event log (BPI2017) and four datasets (BPI2017C, BPI2017E, BPI2017H, and BPI2017L) split from the log.

[0192] The basic information about these logs is shown in Table 2.

[0193] Table 2 Basic Log Information

[0194]

[0195] The dataset for each event log was divided into a 70% training set, a 20% validation set, and a 10% test set. Training was performed using the Adam optimizer (learning rate 0.001, weight decay 1×10⁻⁴), with a batch size of 64 for all logs. The training epochs were set to 60. To ensure time consistency and avoid information leakage, a prefix-based training strategy was employed.

[0196] Specifically, the raw event logs are first converted into an object-centered execution graph, where nodes represent events and edges represent the time, dependency, or resource relationships between events, with each node and edge accompanied by a corresponding feature vector. Based on this, for each complete trajectory in the training set... Multiple training samples are generated by progressively truncating the prefix: .

[0197] Each prefix subsequence corresponds to a subgraph structure containing the currently occurring nodes and their associated edge features. The model uses this prefix subgraph as input to learn how to predict the next event and its occurrence time.

[0198] This sample generation method ensures that the model input depends only on events that have already occurred, making full use of the structural and semantic information described by edge features and node features, thus better conforming to the time logic of financial process execution and avoiding the leakage of future information.

[0199] Each prefix serves as input, and the corresponding next event (or time-related attribute) serves as the prediction target. This strategy ensures that the model observes only historical information when predicting future events, thus preserving the causal structure of the process.

[0200] To explore the optimal value of the hyperparameter α in the CRA module proposed in this invention, experiments were conducted by varying α from 1 / 6 to 5 / 6, and the corresponding training time and accuracy were compared. The results are attached. Figure 7 As shown.

[0201] A larger α value can improve the overall prediction accuracy of the model; however, the training time also increases accordingly. Therefore, α=1 / 2 achieves the best trade-off. Similarly, the hyperparameters for the SLA and MLLA modules are shown in Tables 3 and 4:

[0202] Table 3 Hyperparameters of the SLA module

[0203]

[0204] Table 4 Hyperparameters of the MLLA module

[0205]

[0206] To verify the effectiveness of the method of the present invention, comparative experiments were conducted on five datasets with various baseline models.

[0207] Baseline models include the model proposed by Gherissi et al. (OCPM-LSTM), convolutional graph neural networks (GCN), attention graph neural networks (GAT), and semi-supervised convolutional graph neural networks (SGAE). Experiments were conducted with uniform feature count, loss function, and batch size to ensure fairness. Table 5 shows the comparative experimental results of various methods.

[0208] Table 5 Comparative Experimental Results

[0209]

[0210] The results show that our invention (Our) achieves the best performance on all five datasets, verifying the effectiveness of edge feature modeling.

[0211] Figure 8 Figures (a), (b), (c), and (d) show the prediction accuracy curves for the BPI-2017, BPI2017C, BPI2017E, and BPI2017H datasets with different prefix lengths, respectively. The experiment selected trajectory prefixes with lengths from 2 to 22, as prefixes exceeding 22 were no longer statistically significant and were therefore excluded.

[0212] The results show that, among the five datasets selected, the baseline model has low accuracy with shorter prefixes, improves with increasing prefix length, but declines after reaching a local maximum.

[0213] In comparison, the method of this invention performs similarly to these methods in the initial stages, but maintains high and stable prediction performance in long prefix scenarios. This advantage mainly stems from three key factors:

[0214] 1. Innovations in the CAR-Unet architecture:

[0215] A multi-scale feature fusion mechanism is introduced, significantly enhancing the model's ability to encode contextual information in long sequences. Through a progressive downsampling and upsampling encoder-decoder design, CAR-Unet can extract features at multiple time scales, thereby alleviating the feature dilution problem commonly found in long sequences. Furthermore, the Channel Attention Enhancement (CRA) module dynamically emphasizes highly important channels through channel reconstruction and weight allocation, ensuring that key semantic information is preserved over long spans.

[0216] 2. Introduction of edge features:

[0217] The modeling of structural dependencies between events is enhanced, providing additional relational context and compensating for the gradual weakening of sequential dependencies in long sequences. The combination of structural information and multi-scale fusion enables the model to exhibit significant robustness and generalization ability when handling long prefix trajectories.

[0218] 3. Combination of edge features and CAR-Unet architecture:

[0219] Edge features provide structural dependencies and interaction semantics between events. The CRA-Unet module encodes, fuses, and weights edge features through multi-scale convolution and attention mechanisms, thereby reconstructing and enhancing the feature space. Through this linkage mechanism, the model can dynamically adjust the weights of different edge feature channels, aggregating multi-scale information across time and objects.

[0220] In addition, the present invention also performed a significance test on the experimental results.

[0221] With a fixed 7 / 2 / 1 data partition, the training was repeated 5 times using different random seeds. OCPM-LSTM, SGAE, and GAT were selected as baseline models on the BPI2017 and Order datasets for pairwise t-tests.

[0222] The results are as follows Figure 9 As shown, on the loan log, the p-values ​​of the method of the present invention with SGAE and GAT are all less than 0.001, and the p-value with OCPM-LSTM is less than 0.01, indicating that the method of the present invention is significantly better than the baseline model.

[0223] In this context, * indicates superior, ** indicates significantly superior, and NS indicates no significant difference.

[0224] Furthermore, to verify the effectiveness of the method of this invention, an ablation experiment was also conducted on the loan event log (BPI2017). The experimental settings were kept consistent, with only different modules removed to evaluate their effects. The ablation experiment results are attached. Figure 10 As shown.

[0225] After introducing edge features, the model accuracy improved by 4.1%, indicating that edge features can provide richer graph structure information and effectively enhance the expressive power of node features. However, the improvement was not very significant. After adding CRA-Net on the basis of edge feature enhancement, the accuracy improved by 7.7%. This improvement mainly comes from the feature space expansion and multi-scale feature fusion architecture.

[0226] The CRA module effectively extracts key features and achieves full fusion. The attention mechanisms of the MLLA and SLA modules further enhance this effect, achieving a balance between local and global features. It can be seen that the collaborative work of edge features and the CRA-Unet module has an advantage in improving prediction accuracy. When used alone, the GTPM module improves accuracy by 6.0% compared to the baseline model, indicating that temporal information plays an indispensable role in prediction. Through the bidirectional GRU module, GTPM compensates for the shortcomings of graph neural networks in processing temporal information, playing a significant supplementary role in overall performance. The final results show that the accuracy of Baseline + edge features + CRA-Unet is significantly higher than that of Baseline + GTPM. This is because edge features expand the feature space, allowing the model to not only utilize node features but also capture relationship patterns and semantic dependencies between events. CRA-Net further ensures the full integration of edge features, while the encoder-decoder architecture can simultaneously capture local patterns and long-range dependencies.

[0227] Example 2

[0228] This embodiment 2 describes a business process prediction system based on multi-scale feature fusion and graph temporal modeling, which is based on the same inventive concept as the business process prediction method based on multi-scale feature fusion and graph temporal modeling in the above embodiment 1.

[0229] The business process prediction system based on multi-scale feature fusion and graph temporal modeling in this embodiment includes the following modules:

[0230] The preprocessing module is used to obtain object-centric raw event logs from historical loan business processes and transform the event logs into process execution graphs by extracting process execution. The process execution graph consists of nodes and edges between nodes. The attribute information of nodes and the attribute information of edges are stored in the node feature matrix and the edge feature matrix respectively, and are initially empty.

[0231] The feature extraction module is used to process event logs using edge and node feature extraction functions respectively, and encodes the obtained features into node features and edge features respectively, and then adds them to an empty process execution graph.

[0232] We enhanced edge and node features using a multi-scale feature fusion model based on CRA-Unet and a k-means clustering algorithm, respectively; and used the process execution graph containing node and edge features, as well as the node feature matrix, as the dataset.

[0233] The system also includes a prediction module for building a time graph prediction model (GTPM) that can simultaneously process graph structure information and log time series information. Its inputs are the process execution graph and the node feature matrix, i.e., sequence features, and the model output is the next activity in the process execution. The system trains the model based on the dataset and uses the trained model to predict activities.

[0234] It should be noted that the implementation process of the functions and roles of each functional module in the business process prediction system described in this embodiment 2 is detailed in the implementation process of the corresponding steps in the method of embodiment 1 above, and will not be repeated here.

[0235] Example 3

[0236] This embodiment 3 describes a computer device including a memory and one or more processors. Executable code is stored in the memory. When the processor executes the executable code, it implements the steps of the business process prediction method based on multi-scale feature fusion and graph temporal modeling described in embodiment 1 above.

[0237] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0238] Example 4

[0239] This embodiment 4 describes a computer-readable storage medium storing a program that, when executed by a processor, is used to implement the steps of the business process prediction method based on multi-scale feature fusion and graph temporal modeling in embodiment 1 above.

[0240] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.

[0241] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A business process prediction method based on multi-scale feature fusion and graph temporal modeling, characterized in that, Includes the following steps: Step 1. Obtain the object-centric raw event log from the historical loan business process, and transform the event log into a process execution graph by extracting the process execution; the process execution graph consists of nodes and edges between nodes; the attribute information of nodes and the attribute information of edges are stored in the node feature matrix and the edge feature matrix respectively, and are initially empty; Step 2. Process the event logs using edge and node feature extraction functions respectively, and encode the obtained features into node features and edge features respectively, and then add them to the empty process execution graph; We enhanced edge and node features using a multi-scale feature fusion model based on CRA-Unet and a k-means clustering algorithm, respectively; and used the process execution graph containing node and edge features, as well as the node feature matrix, as the dataset. The CRA-Unet network comprises four encoder stages and four decoder stages; The initial input to the CRA-Unet network is the edge feature graph (B, C, L) obtained from the previous feature encoding process, where B, C, and L represent the batch size, number of channels, and number of edges of each process execution feature graph, respectively. In the first and second encoder stages, the channel reconstruction attention module (CRA) is used and combined with a downsampling operation to compress the length of the edge, respectively. In the third encoder stage, a simplified linear attention module (SLA) is used and a downsampling operation is performed to continue compressing the side length; After the first and second encoder stages, the number of edges is halved, and the overall sequence feature map structure becomes (B, C, L / 4); after the third and fourth encoder stages, the number of edges is halved again, and the overall sequence feature map structure becomes (B, C, L / 8). A multi-scale linear attention module (MLLA) is set up in the bottleneck layer to extract higher-level semantic features; In the first and second decoder stages, the channel reconstruction attention module (CRA) is used in conjunction with a linear interpolation operation (LI) to perform upsampling, gradually restoring the length of the edge to its original size. In the third decoder stage, a simplified linear attention module (SLA) is used in conjunction with a linear interpolation operation (LI) to continue upsampling and gradually restore the complete structure of the feature map; the fourth decoder stage uses a simplified linear attention module (SLA). Set up skip connections between the corresponding stages of the encoder and the corresponding stages of the decoder; Step 3. Build a time graph prediction model GTPM that can simultaneously process graph structure information and log time series information. Its input is the process execution graph and the node feature matrix, i.e., sequence features. The model output is the next activity in the process execution. The model is trained based on the dataset obtained in step 2, and the trained model is used to predict activities. 2.The business process prediction method based on multi-scale feature fusion and graph temporal modeling according to claim 1, characterized in that, In step 1, the process of converting the event log into a process execution graph by extracting the process execution is as follows: First, an object graph consisting of a set of interrelated objects is obtained based on the event log. In the object graph, nodes represent different objects, and edges represent events that connect two objects. Next, find the events involved in each object in the object graph and the trajectory of each of these events. Use these events as nodes in the process execution graph and the order of the trajectories as edges to form the final process execution graph. 3.The business process prediction method based on multi-scale feature fusion and graph temporal modeling according to claim 1, characterized in that, Step 2 specifically involves: Feature extraction is performed on the original event log to obtain vector representations related to activities, time, objects, and resources in the event log; at the same time, the event relationship graph corresponding to the specified process instance is extracted from the original event log, and the edges of the process execution graph are generated and assigned corresponding feature values. For the vector representation of the obtained activity-related features, the corresponding node features are obtained by using node feature encoding; for the vector representation of the obtained node pair features, the corresponding edge features are obtained by using edge feature encoding. The node features and edge features are assigned to the node feature matrix and the edge feature matrix respectively to form a flow execution graph with features. The edge features and node features are then enhanced using the CRA-Unet network and the k-means clustering algorithm respectively.

4. The business process prediction method based on multi-scale feature fusion and graph temporal modeling according to claim 3, characterized in that, In step 2, the process of processing the event log using the edge feature extraction function is as follows: First, based on the given process instance number, extract the subgraph corresponding to the process instance from the event log. The subgraph contains all event nodes related to the process instance and their interconnecting edges. Then, perform feature calculations on each edge in the subgraph in sequence. For each edge, iterate through the predefined set of edge feature processing functions, using the event log as input parameters, and automatically call the corresponding feature functions to calculate different edge attributes. After the calculation is completed, obtain the edge features of all edges.

5. The business process prediction method based on multi-scale feature fusion and graph temporal modeling according to claim 4, characterized in that, In step 2, the defined edge feature processing function is as follows: The first edge feature processing function is used to calculate the absolute time difference between the occurrence of two endpoint events in the loan business process; First, let e=(s,t) represent an edge, where s and t are the source node and the target node, respectively. Let T(s) and T(t) represent the event timestamps of the source node s and the target node t, respectively. The absolute time difference is obtained by subtracting them. The second edge feature processing function is used to determine whether two endpoint events fall on Saturday or Sunday at the same time, and returns a boolean value indicating whether the two events occur on Saturday or Sunday at the same time. Let e ​​= (s, t) represent an edge, where s and t are the source node and the target node, respectively. Let D(s) and D(t) represent the event dates of the source node s and the target node t, respectively. D(s) and D(t) are obtained using the day_of_week function. The third edge feature processing function is a function to determine the activity transition type. It is used to determine the activity types of the two nodes associated with the current edge, and assigns a vector to each association type using one-hot encoding. The specific processing procedure is as follows: Let e=(s,t) represent an edge, where s and t are the source node and the target node, respectively. Check the activity type of s and t, and then match them with the one-hot dictionary to obtain the feature value of the edge. The fourth edge feature processing function is used to determine whether there is a timeout in the activities of the two nodes associated with the current edge; Based on whether a timeout has occurred, a boolean value is returned: 1 if the timeout occurs, and 0 otherwise. The specific implementation method is to calculate the time difference between the source node and the target node, and then compare it with the set time threshold to obtain the feature value. The fifth edge feature processing function is used to determine whether the application amount has changed, which requires judging the difference between the two nodes; If the difference is 0, a boolean value of 0 is returned, indicating that the loan amount has not changed; If the difference is not 0, the return value is a boolean value of 1, indicating that the loan amount has changed; The sixth edge feature processing function is used to determine whether the credit score has changed and returns a Boolean value of 0 or 1; specifically, it returns 1 when the credit score has changed and 0 when the credit score has not changed.

6. The business process prediction method based on multi-scale feature fusion and graph temporal modeling according to claim 1, characterized in that, The processing flow of the Channel Reconstruction Attention Module (CRA) is as follows: When the sequence feature map of input (B,C,L) enters the CRA module, it is processed as follows: First, preliminary feature extraction is performed through one-dimensional convolution. The extracted features are divided into two branches, and the number of channels is divided into enriched channels F_up and supplementary channels F_down according to the ratios α and 1-α, with the number of channels being (1-α)C and αC, respectively. Next, one-dimensional convolution is performed on the two channels to compress the channels and obtain the key feature representations P1 and P2 of the channels; Then, group convolution and point convolution are performed on the enriched channel key feature representation P1, respectively, and the number of channels is restored to C. Finally, the results are summed to obtain the enriched channel representation. On the other hand, point convolution is applied to the key feature representation P2 of the supplementary channel, and the number of channels is adjusted to αC. Then, it is concatenated with the input of the supplementary channel to obtain the supplementary channel representation. ; Then and The reconstructed features are obtained by adding elements one by one. Its number of channels is C; next, for Perform the following processing: On the one hand, the reconstructed features Global average pooling is performed to obtain the aggregated representation for each channel. The result is then processed by two 1×1 point convolutions and a ReLU activation function to extract global context information. ; On the other hand, regarding the reconstructed features Local context information can be extracted directly using two 1×1 point convolutions. ; The information extracted from the above two aspects is then integrated to obtain... The corresponding attention weights are obtained using the softmax function. ; assign each attention weight Represented by the corresponding channels Multiply them, then add them together; Finally, a 1×1 convolution is used to further promote feature fusion, resulting in fused features.

7. The business process prediction method based on multi-scale feature fusion and graph temporal modeling according to claim 1, characterized in that, In step 3, the time-map prediction model GTPM consists of two parallel branch structures: The first branch structure is a graph structure processing branch. Its input is a process execution graph containing node features and edge features. It processes structural information based on graph isomorphic networks and obtains graph structure information features. The second branch structure is the time-series structure processing branch. Its input is the node feature matrix, i.e., the sequence features, which is used to process time information using a bidirectional gated cyclic unit to obtain time-series information features. After feature extraction is completed in the two branch structures, the output features of the two branch structures are first connected by a Concat, and then the connected features are input into the gated fusion module. This gating fusion module is used to dynamically adjust the flow ratio of graph structure information and timing information to highlight key information; The flow ratios of the obtained graph structure information and time series information are multiplied by the corresponding branch structure outputs to obtain the first multiplication result and the second multiplication result. Then, the two multiplication results are added together and the prediction result is output through the dense connection layer.

8. The business process prediction method based on multi-scale feature fusion and graph temporal modeling according to claim 1, characterized in that, The graph structure processing branch consists of multiple graph convolutional modules; Each graph convolutional module contains a graph neural network, a hybrid weighted graph normalization method, and an activation function; During graph neural network updates, edge features are fused with the features of adjacent nodes. The fused node features and edge features are aggregated by averaging, and then nonlinear transformation is performed by a multilayer perceptron to generate new node features. The hybrid weighted graph normalization method employs four normalization methods: node, adjacency, graph level, and batch. It also uses learned shared weights for weighted fusion to ensure stability while enhancing the ability to capture local and global structural information. Graph structure processing branch outputs effectively extract graph structure and its node feature information; The timing structure processing branch uses a bidirectional gated cyclic unit (Bi-GRU) to model the timing information of the event log. At the same time, a linear transformation layer is added to this branch to restore the output dimension to its original length for subsequent processing. The input of the temporal structure processing branch only utilizes node features (B,C,L). The input first passes through a Bi-GRU, which doubles the number of channels, and then passes through a linear layer to restore the number of channels to C. Among them, the number of stacked channels for graph structure processing branches and time sequence structure processing branches becomes 2×C.

9. A business process prediction system based on multi-scale feature fusion and graph temporal modeling, used to implement the business process prediction method based on multi-scale feature fusion and graph temporal modeling as described in claim 1, characterized in that, The business process prediction system based on multi-scale feature fusion and graph temporal modeling includes the following modules: The preprocessing module is used to obtain object-centric raw event logs from historical loan business processes and transform the event logs into a process execution graph by extracting process execution. The process execution graph consists of nodes and edges between nodes. The attribute information of nodes and the attribute information of edges are stored in the node feature matrix and the edge feature matrix respectively, and are initially empty. The feature extraction module is used to process event logs using edge and node feature extraction functions respectively, and encodes the obtained features into node features and edge features respectively, and then adds them to an empty process execution graph; We enhanced edge and node features using a multi-scale feature fusion model based on CRA-Unet and a k-means clustering algorithm, respectively; and used the process execution graph containing node and edge features, as well as the node feature matrix, as the dataset. The system also includes a prediction module, which is used to build a time graph prediction model (GTPM) that can simultaneously process graph structure information and log time series information. Its input is the process execution graph and the node feature matrix, i.e., sequence features. The model output is the next activity in the process execution. The model is trained based on the dataset and the trained model is used to predict activities.

Images