Track clustering method and device based on graph convolutional network

Abstract

The invention provides a trajectory clustering method based on a graph convolutional network, and the method comprises the steps: extracting feature vectors of all trajectories at different visual angles from an original event log, and obtaining multi-visual-angle trajectory features; using a K-nearest neighbor algorithm to construct a K-nearest neighbor graph for each view angle according to the multi-view-angle track features to obtain a multi-view-angle graph; performing convolution operation on the multi-view image by using a multi-view image encoder based on an image convolution network to obtain an initial public representation; inputting the multi-view image into a neural network for fusion, and optimizing the structure of a fusion image through rank constraint of a Laplacian matrix; using the initial common representation and the fusion image to learn the common representation, using a multi-view image decoder to reconstruct a multi-view image, and optimizing the common representation by minimizing the difference between the reconstructed multi-view image and the original multi-view image to obtain a final common representation; and clustering the tracks in the original event log based on the final common representation. By the adoption of the scheme, track clustering of the event logs is achieved.

Description

Technical Field

[0001] The present application relates to the field of process mining technology, and in particular to a trajectory clustering method and device based on a graph convolutional network. Background Art

[0002] The widespread use of information technology enables information systems to collect a large number of event logs, which can be deeply analyzed to improve service quality. Process mining is a set of technologies for process discovery, monitoring and optimization based on the logs recorded by enterprise information systems. These technologies help organizations break through the limitations of traditional process analysis methods, which often rely on manual observation or static modeling. Process mining uses actual execution data to reveal the true dynamics of the process. In essence, the core concept of process mining is to optimize the process by analyzing event logs to ensure that it runs as efficiently as possible.

[0003] Common research directions of process mining include the following:

[0004] Process discovery: aims to discover process models that optimize specific metrics (such as fitness and accuracy) to capture and represent the behavior of the underlying process.

[0005] Consistency check: Compare process models with event logs to identify and analyze deviations.

[0006] Process Optimization: Systematically improving processes by refining or optimizing existing models.

[0007] Predictive process monitoring: This important area aims to predict future process behavior using historical event logs [4].

[0008] Among them, process discovery has attracted much attention because it can automatically build process models from logs and help business personnel intuitively understand the overall process. At present, a large number of process discovery technologies have been proposed. However, in practical applications, a common problem is that process models generated from unprocessed logs are often very complex and difficult to understand, often referred to as "spaghetti models". This complexity stems from the inherent diversity of processes and the problems caused by trying to show all possible behaviors in a single model.

[0009] To solve this problem, the trajectories in the original log can be grouped through trajectory clustering, where each cluster corresponds to a group of related process executions. These clusters can then be used to generate more concise and interpretable process models. Trajectory clustering is a log preprocessing technology that aims to divide the original log into multiple sub-logs, each of which is then used to generate a process model through process discovery technology (such as heuristic mining algorithms). This method can effectively reduce the complexity of the process model and enhance the user's understanding of the business process. Trajectory clustering is a very effective partitioning method that can divide the original event log into multiple sub-logs to achieve the purpose of divide and conquer.

[0010] At present, most trajectory clustering algorithms are single-view clustering, which usually clusters based on a single view, such as control flow view or resource view. However, this method fails to reflect the characteristics of the process as a multi-view entity, ignores the information of other viewpoints, and easily leads to low accuracy of clustering results. In addition, multi-view clustering algorithms still lack the weight distribution of each viewpoint or ignore the concurrency, loop and selection structure in the process model. Summary of the invention

[0011] The present application aims to solve one of the technical problems in the related art at least to some extent.

[0012] To this end, the first purpose of this application is to propose a trajectory clustering method based on graph convolutional networks, which takes into account multi-dimensional information and effectively improves the accuracy of clustering results.

[0013] The second objective of this application is to propose a trajectory clustering device based on a graph convolutional network.

[0014] To achieve the above objectives, the first embodiment of the present application proposes a trajectory clustering method based on a graph convolutional network, comprising:

[0015] Extract feature vectors of all trajectories from different perspectives from the original event log to obtain multi-perspective trajectory features;

[0016] Use the K nearest neighbor algorithm to construct a K nearest neighbor graph for each view according to the multi-view trajectory features to obtain a multi-view graph;

[0017] A multi-view graph encoder based on a graph convolutional network is used to perform convolution operations on the multi-view graph to obtain an initial common representation;

[0018] The multi-view graphs are input into the neural network for fusion, and the structure of the fusion graph is optimized through the rank constraint of the Laplacian matrix to obtain the fusion graph;

[0019] The initial common representation and the fusion graph are used to learn the common representation, and the multi-view graph decoder is used to reconstruct the multi-view graph. The common representation is optimized by minimizing the difference between the reconstructed multi-view graph and the original multi-view graph to obtain the final common representation.

[0020] Based on the final common representation, the trajectories in the original event log are clustered.

[0021] Optionally, in one embodiment of the present application, different perspectives include an activity perspective, a resource perspective, and a transition perspective, and the activity perspective, the resource perspective, and the transition perspective are used to describe each track in the original event log, wherein:

[0022] The activity perspective describes the activities contained in the trajectory;

[0023] The resource perspective describes the execution resources of the activities contained in the trace;

[0024] The transition perspective reflects the transition relationship between the activities contained in the trajectory.

[0025] Optionally, in one embodiment of the present application, the nodes in the k-nearest neighbor graph represent trajectories, the edges between the nodes represent the similarities between the trajectories, and the k-nearest neighbor algorithm is used to construct a k-nearest neighbor graph for each perspective according to the multi-perspective trajectory features to obtain a multi-perspective graph, including:

[0026] For each view, the similarity between trajectories is calculated based on the corresponding trajectory features, and the k nearest neighbor trajectories are selected for each trajectory, and the k nearest neighbor graph of the view is constructed using the K nearest neighbor algorithm.

[0027] Optionally, in one embodiment of the present application, a multi-view graph encoder based on a graph convolutional network is used to perform a convolution operation on the multi-view graph to obtain an initial public representation, including:

[0028] Input the multi-view trajectory features and multi-view graphs into the multi-layer GCN and output the features of each view;

[0029] The attention mechanism is introduced to learn the weight of each perspective, fuse the features of each perspective, and obtain the initial public representation.

[0030] Optionally, in one embodiment of the present application, a multi-layer GCN performs a convolution operation, and feature information is propagated between the multi-layer GCNs. Assume that the representation of the lth layer is H (l) , and its propagation rules are:

[0031]

[0032] The result of the first layer of GCN convolution is:

[0033]

[0034] in, is the multi-view trajectory feature, is the adjacency matrix of the multi-view graph, v is the view angle;

[0035] Assuming that the multi-layer GCN is layer l, the result of the convolution of the l-th layer of GCN is:

[0036]

[0037] The initial public representation is:

[0038]

[0039] in, is the weight matrix for each view.

[0040] Optionally, in one embodiment of the present application, fusing the multi-view images through a neural network includes:

[0041] Use neural network to fuse multi-view images. During fusion, learn the attention coefficient matrix of each image, obtain the weighted accumulation results of each image, and pass the weighted accumulation results through the linear layer to obtain the fusion image, where:

[0042] The weighted cumulative results of each graph are expressed as:

[0043]

[0044] Among them, G (v) Represents a multi-view graph, is the attention coefficient matrix for each perspective, v is the perspective,

[0045] The fusion graph is represented as:

[0046] G=φ(W g G (*) +b g )

[0047] Among them, φ is the activation function of the fully connected layer, W g and b g They are the weight matrix and bias matrix in the neural network respectively;

[0048] The structure of the fusion graph is optimized through the rank constraint of the Laplacian matrix, including:

[0049] The fusion graph loss is constructed based on the rank constraint of the Laplace matrix, and the loss function of the fusion graph is constructed based on the fusion graph reconstruction loss and the fusion graph loss, where:

[0050] The loss function of the fusion graph is expressed as:

[0051] Loss G =Lossgre +λ1Loss gtr

[0052] Among them, λ1 is a hyperparameter that balances the fusion image reconstruction loss and the fusion image trace loss, Loss gre The reconstruction loss for the fusion graph is expressed as:

[0053]

[0054] A is the adjacency matrix of the fusion graph, A (v) is the adjacency matrix of the multi-view graph,

[0055] Loss gtr For the fusion trace loss, it is expressed as:

[0056] Loss gtr =tr(Q T L A Q),stQ T Q=I

[0057] Among them, L A is the Laplacian matrix of A, Q represents the eigenvector matrix of A, the trace operator tr() calculates the trace of the matrix, and I is the identity matrix.

[0058] Optionally, in one embodiment of the present application, the multi-view image decoder is represented as:

[0059]

[0060] Among them, A ′ (v) represents the adjacency matrix of each view reconstructed by the final common representation Z2, W (v) is the self-learning parameter matrix in the decoder, and v is the viewing angle;

[0061] The initial common representation and the fusion graph are used to learn the common representation, including:

[0062] Input the initial common representation and fusion graph into the multi-layer GCN to obtain the final common representation;

[0063] By minimizing the difference between the reconstructed multi-view image and the original multi-view image, the public representation is optimized, including:

[0064] Through back-propagation training, the difference between the reconstructed multi-view image and the original multi-view image is minimized and the common representation is optimized. The loss function during optimization is:

[0065]

[0066] Among them, Loss rec is the reconstruction loss for all views, A(v) is the adjacency matrix of the multi-view graph.

[0067] To achieve the above-mentioned purpose, the second embodiment of the present invention proposes a trajectory clustering device based on a graph convolutional network, comprising:

[0068] A feature extraction module is used to extract feature vectors of all trajectories from different perspectives from the original event log to obtain multi-perspective trajectory features;

[0069] A graph construction module, used to construct a K nearest neighbor graph for each view according to the multi-view trajectory features using a K nearest neighbor algorithm to obtain a multi-view graph;

[0070] A graph convolution module, used to perform convolution operations on multi-view graphs using a multi-view graph encoder based on a graph convolution network to obtain an initial common representation;

[0071] The graph fusion module is used to input the multi-view graph into the neural network for fusion, and optimize the structure of the fusion graph through the rank constraint of the Laplace matrix to obtain the fusion graph;

[0072] An optimization module is used to learn a common representation using the initial common representation and the fusion graph, and reconstruct the multi-view graph using a multi-view graph decoder, and optimize the common representation by minimizing the difference between the reconstructed multi-view graph and the original multi-view graph to obtain a final common representation;

[0073] The clustering module is used to cluster the trajectories in the raw event log based on the final common representation.

[0074] Optionally, in one embodiment of the present application, different perspectives include an activity perspective, a resource perspective, and a transition perspective, and the activity perspective, the resource perspective, and the transition perspective are used to describe each track in the original event log, wherein:

[0075] The activity perspective describes the activities contained in the trajectory;

[0076] The resource perspective describes the execution resources of the activities contained in the trace;

[0077] The transition perspective reflects the transition relationship between the activities contained in the trajectory.

[0078] Optionally, in one embodiment of the present application, the nodes in the k-nearest neighbor graph represent trajectories, the edges between the nodes represent the similarities between the trajectories, and the k-nearest neighbor algorithm is used to construct a k-nearest neighbor graph for each perspective according to the multi-perspective trajectory features to obtain a multi-perspective graph, including:

[0079] For each view, the similarity between trajectories is calculated based on the corresponding trajectory features, and the k nearest neighbor trajectories are selected for each trajectory, and the k nearest neighbor graph of the view is constructed using the K nearest neighbor algorithm.

[0080] The trajectory clustering method and device based on graph convolutional network of the embodiment of the present application fully utilize the advantages of the encoder-decoder framework and significantly improve the clustering effect by adapting the encoder-decoder strategy to the information fusion in trajectory clustering; utilize multiple trajectory feature configuration files under the multi-view learning framework to effectively fuse feature information and improve the accuracy and robustness of trajectory clustering; introduce the attention mechanism to automatically learn the weight of each view; the features or attributes provided for each view are different, and by integrating multi-view information, the intrinsic structural characteristics of the data are more comprehensively captured.

[0081] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0083] Figure 1 This is an example diagram of an event log.

[0084] Figure 2 A schematic diagram of a flow chart of a trajectory clustering method based on a graph convolutional network provided in Example 1 of the present application;

[0085] Figure 3 A schematic diagram of the process of extracting a trajectory profile and converting it into a directed graph in an embodiment of the present application;

[0086] Figure 4 A schematic diagram of the structure of a trajectory clustering device based on a graph convolutional network provided in an embodiment of the present application. DETAILED DESCRIPTION

[0087] Embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0088] Figure 1 This is an example of an event log. Figure 1 It can be seen that each trace in the log contains not only the activity, but also the timestamp and related resources of the activity. Therefore, when dividing the traces into different clusters, multiple perspectives must be considered simultaneously to improve the homogeneity of the clusters.

[0089] The following describes a trajectory clustering method and device based on a graph convolutional network according to an embodiment of the present application with reference to the accompanying drawings.

[0090] Figure 2 A schematic diagram of the flow of a trajectory clustering method based on a graph convolutional network provided in Example 1 of the present application.

[0091] like Figure 2 As shown in FIG, the trajectory clustering method based on graph convolutional network includes the following steps:

[0092] Step 201, extracting feature vectors of all trajectories from different perspectives from the original event log to obtain multi-perspective trajectory features;

[0093] Step 202, using a K-nearest neighbor algorithm to construct a K-nearest neighbor graph for each view according to the multi-view trajectory features, to obtain a multi-view graph;

[0094] Step 203, using a multi-view graph encoder based on a graph convolutional network to perform a convolution operation on the multi-view graph to obtain an initial public representation;

[0095] Step 204, input the multi-view graphs into a neural network for fusion, and optimize the structure of the fused graph by using the rank constraint of the Laplacian matrix to obtain a fused graph;

[0096] Step 205, learning a common representation using the initial common representation and the fusion graph, and reconstructing the multi-view graph using a multi-view graph decoder, optimizing the common representation by minimizing the difference between the reconstructed multi-view graph and the original multi-view graph, and obtaining a final common representation;

[0097] Step 206 , clustering the trajectories in the original event log based on the final common representation.

[0098] The trajectory clustering method based on graph convolutional network in the embodiment of the present application adapts the encoder-decoder strategy to the information fusion in trajectory clustering, fully utilizes the advantages of the encoder-decoder framework, and significantly improves the clustering effect; utilizes multiple trajectory feature configuration files under the multi-view learning framework to effectively fuse feature information and improve the accuracy and robustness of trajectory clustering; introduces the attention mechanism to automatically learn the weights of each view; the features or attributes provided for each view are different, and by integrating multi-view information, the intrinsic structural characteristics of the data are more comprehensively captured.

[0099] The embodiment of the present application also proposes a multi-view trajectory clustering framework, which mainly consists of three parts: a multi-view graph encoder, a fusion graph generation module, and a multi-view graph decoder. This embodiment implements multi-view learning through an encoder and a decoder, and can integrate a common representation of each trajectory in the log based on the features of multiple perspectives of the trajectory, including activity perspective, resource perspective, and transition perspective.

[0100] The overall process is as follows: First, the feature vector of each view is extracted from the original log. Then, the KNN algorithm is used to construct the k-nearest neighbor graph of the trajectory, and the graph convolution network (GCN) is used to perform convolution operations on these graphs. Secondly, the graph of each view is input into the neural network to learn the fusion graph, and the structure of the fusion graph is optimized through the rank constraint of the Laplacian matrix to make it more sparse and conducive to clustering. Then, the graph convolution results of each view and the fusion graph are used to learn the common representation. By combining multi-view features, a more complete and accurate representation can be obtained. Finally, the common representation is continuously improved through the internal decoder module to ensure that the common representation can accurately reproduce the original directed graph.

[0101] Finding the similarity between two trajectories is the first step to effective clustering. Trajectories are described as profiles, and profiles are a collection of elements that describe the trajectory from a specific perspective. In other words, a profile containing n elements can be viewed as a transformation of the trajectory into a vector <i1,i2,...,i n > operation. Then, the similarity between two trajectories can be calculated based on these vectors. This embodiment uses three profiles (ie, activity profile, resource profile and transition profile) to describe each trajectory.

[0102] Activity profile: Activities are considered as the main and key elements of a trace. Therefore, activity profile plays an important role in defining the similarity of two traces. Activity profile is obtained by aggregating and statistically analyzing the raw logs. It contains all the activities in the process, and each activity has a corresponding value in the activity profile, which indicates the number of times the activity appears in the trace. Figure 3 As shown, seven unique activities (i.e., SI, RD, VI, PF, CA, AD, and IS) are extracted from the log, so each trajectory has a feature vector of length seven. The value in the vector is the number of times the activity appears in the trajectory. For example, activity RD appears once in trajectory 1 and 0 times in trajectory 2, so the activity profile feature vectors for trajectory 1 and trajectory 2 are <1,1,1,1,1,0,0> and <1,0,1,1,0,1,1>, respectively.

[0103] Resource profiles: In some cases, even if two traces have the same number and frequency of activities, they may still differ because the roles (i.e., resources) performing the activities are different, leading to differences in execution time or quality.

[0104] Therefore, it is equally important to consider the resource perspective and the activity perspective. The resource profile contains all the resources in the process and is constructed by counting the number of events triggered by a specified resource in each trace. Figure 3As shown in Figure 1, five unique resources are extracted from the log, and the feature vector length of each trajectory is five. The values ​​in the vector represent the frequency of the resource in the trajectory. For example, the resource Alex appears twice in trajectory 1 and once in trajectory 2, so the resource profile feature vectors of trajectory 1 and trajectory 2 are <2,1,1,1,0> and <1,0,2,1,1> respectively.

[0105] Transition profile: The transition profile focuses on the order of events in the trajectory. By considering the transition relationship between activities, we can understand the order of events. The transition profile reflects the direct relationship between activities, and each transition feature is characterized by the order in which the activities occur. For example, if activity SI is followed by activity RD, the transition feature is expressed as (SI→RD). Figure 3 As shown, nine unique transitions are extracted from the log, resulting in a feature vector of length nine for each trajectory. The values ​​in the vector represent the frequency of each transition. For example, the transition SI→RD appears once in trajectory 1, so the feature vector of trajectory 1 in the transition profile is <1,1,1,...,0>.

[0106] After obtaining the trajectory profiles of different perspectives, the k-nearest neighbor graph is constructed for each perspective using the k-nearest neighbor algorithm. For each perspective, the similarity between trajectories is calculated, and the k nearest neighbor trajectories are selected for each trajectory. Figure 3 As shown in the activity perspective of , the feature vectors are obtained and the corresponding graph is constructed using the k-nearest neighbor algorithm. It is worth noting that different k values ​​will lead to different graph structures. If the k value is too small, the similarity between trajectories may not be accurately captured. On the contrary, if the k value is too large, the edges of the graph may become redundant. Therefore, the default k=3. In the k-nearest neighbor graph, each node represents a trajectory, and the edges between nodes represent the similarity between trajectories.

[0107] After constructing the k-nearest neighbor graphs at different views, the goal of this embodiment is to fuse the features of each view into a unified representation. To this end, a multi-view encoder based on a graph convolutional network (GCN) is introduced. In the encoder, the feature vector of each view and its corresponding k-nearest neighbor graph are first input. Then, the GCN layer performs convolution operations to propagate feature information between multiple layers. Assume that the representation of the lth layer is H (l) , and its propagation rules are as follows:

[0108]

[0109] In the convolution process, two layers of GCN are used. The input is the multi-view trajectory features of each view v and its corresponding graph adjacency matrix Each GCN module learns a representation for each view Its structure is as follows:

[0110]

[0111] in, Represents the result of the l-th layer of convolution, Z (v) Represents the result of the second layer of convolution. After obtaining the convolution results of each view, considering that the importance of different views in the trajectory features may be different, this embodiment introduces an attention mechanism to learn the weight of each view. This mechanism can adaptively assign importance to different views, thereby integrating the features of each view. The specific attention process is as follows:

[0112]

[0113] in, is the weight matrix for each view, and Z1 represents the initial common representation.

[0114] In order to improve the representation effect of the initial common representation, this embodiment inputs Z1 and the fusion graph into a two-layer GCN to obtain the final common representation Z2.

[0115] The initial public representation simply merges the vectors from different perspectives, without considering the overall structural information of the log. Since the k-nearest neighbor graph of each perspective only represents the similarity between trajectories from a specific perspective, a single-perspective graph cannot fully capture the overall structure of the trajectory data. To solve this problem, this embodiment designs a fusion graph module to merge graphs from different perspectives into a unified fusion graph.

[0116] This module uses a neural network to learn the fusion graph G. First, the k-nearest neighbor graph of each view is input into the module. Then, the module automatically learns the attention coefficient matrix of each graph to characterize the importance of each graph in the fusion process. The operation is described as follows:

[0117]

[0118] Among them, G (v) represents the k-nearest neighbor graph for each view, is the attention coefficient matrix for each view, G (*) is the weighted cumulative result of each graph. Next, the cumulative result G (*) The final fusion graph G is obtained through the linear layer.

[0119] G=φ(W g G (*) +b g )

[0120] Where φ is the activation function of the fully connected layer, W g and b gare the weight matrix and bias matrix in the neural network, respectively. Through this process, the module is able to fuse information from each perspective and generate a comprehensive graph containing multi-perspective features.

[0121] In order to maximize the retention of the k-nearest neighbor graph information of each perspective and optimize the fusion graph, this embodiment introduces the loss function Loss gre (Fusion graph reconstruction loss function) is used to measure the difference between the fusion graph and each view graph. In addition, in order to make the fusion graph more suitable for clustering tasks, this embodiment uses the rank constraint of the Laplacian matrix to eliminate redundant edges in the fusion graph. Therefore, this embodiment introduces another loss function Loss gtr (fusion graph loss function) to further improve the performance of the fusion graph in the clustering task. The total loss of the fusion graph is defined as follows:

[0122] Loss G =Loss gre +λ1Loss gtr

[0123] in,

[0124]

[0125] Loss gtr =tr(Q T L A Q),stQ T Q=I

[0126] A is the adjacency matrix of the fusion graph, L A is its Laplacian matrix. Q represents the eigenvector matrix of A, and Q T is the transposed matrix of Q. The trace operator tr() is used to calculate the trace of the matrix, and λ1 is a hyperparameter that balances the two loss functions.

[0127] After inputting the fusion image and the initial common representation z1, the present embodiment can obtain the final common representation Z2. In order to ensure that Z2 accurately contains all features from different perspectives, the present embodiment designs an internal decoder to reconstruct the multi-view image from Z2. This internal decoder is specifically designed to:

[0128]

[0129] Among them, A ′ (v) W represents the adjacency matrix of each view reconstructed by the common representation Z2. (v) is the self-learning parameter matrix in the decoder, one for each view. is the transposed matrix of Z2.

[0130] Each parameter matrix W(v) It can be regarded as the weight in the decoder, controlling the reconstruction process of the common representation between different views. Specifically, the parameter matrix W (v) It is able to identify and enhance the values ​​in the common representation Z2 that are similar to a specific view v. This process enables the common representation to accurately reconstruct the graph corresponding to view v.

[0131] In obtaining the reconstruction of Figure A ′ (v) After that, this embodiment minimizes A through back propagation training. ′ (v) Compared with the original figure A (v) This process continuously optimizes the model to improve the accuracy of the joint representation.

[0132]

[0133] Among them, Loss rec is the reconstruction loss of all views. As mentioned above, the core goal of this embodiment is to obtain a common representation containing multi-view data, which is suitable for trajectory clustering. Therefore, based on this common representation, the k-means algorithm is used to cluster the trajectories in the log.

[0134] In order to implement the above embodiments, the present application also proposes a trajectory clustering device based on a graph convolutional network.

[0135] Figure 4 A schematic diagram of the structure of a trajectory clustering device based on a graph convolutional network provided in an embodiment of the present application.

[0136] like Figure 4 As shown, the trajectory clustering device based on graph convolutional network includes:

[0137] A feature extraction module is used to extract feature vectors of all trajectories from different perspectives from the original event log to obtain multi-perspective trajectory features;

[0138] A graph construction module, used to construct a K nearest neighbor graph for each view according to the multi-view trajectory features using a K nearest neighbor algorithm to obtain a multi-view graph;

[0139] A graph convolution module, used to perform convolution operations on multi-view graphs using a multi-view graph encoder based on a graph convolution network to obtain an initial common representation;

[0140] The graph fusion module is used to input the multi-view graph into the neural network for fusion, and optimize the structure of the fusion graph through the rank constraint of the Laplace matrix to obtain the fusion graph;

[0141] An optimization module is used to learn a common representation using the initial common representation and the fusion graph, and reconstruct the multi-view graph using a multi-view graph decoder, and optimize the common representation by minimizing the difference between the reconstructed multi-view graph and the original multi-view graph to obtain a final common representation;

[0142] The clustering module is used to cluster the trajectories in the raw event log based on the final common representation.

[0143] Optionally, in one embodiment of the present application, different perspectives include an activity perspective, a resource perspective, and a transition perspective, and the activity perspective, the resource perspective, and the transition perspective are used to describe each track in the original event log, wherein:

[0144] The activity perspective describes the activities contained in the trajectory;

[0145] The resource perspective describes the execution resources of the activities contained in the trace;

[0146] The transition perspective reflects the transition relationship between the activities contained in the trajectory.

[0147] Optionally, in one embodiment of the present application, the nodes in the k-nearest neighbor graph represent trajectories, the edges between the nodes represent the similarities between the trajectories, and the k-nearest neighbor algorithm is used to construct a k-nearest neighbor graph for each perspective according to the multi-perspective trajectory features to obtain a multi-perspective graph, including:

[0148] For each view, the similarity between trajectories is calculated based on the corresponding trajectory features, and the k nearest neighbor trajectories are selected for each trajectory, and the k nearest neighbor graph of the view is constructed using the K nearest neighbor algorithm.

[0149] It should be noted that the aforementioned explanation of the embodiment of the trajectory clustering method based on the graph convolutional network is also applicable to the trajectory clustering device based on the graph convolutional network of this embodiment, which will not be repeated here.

[0150] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0151] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0152] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0153] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute the instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0154] It should be understood that the various parts of the present application can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0155] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0156] In addition, each functional unit in each embodiment of the present application may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0157] The storage medium mentioned above may be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limiting the present application. A person of ordinary skill in the art may change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A trajectory clustering method based on graph convolutional network, characterized in that: include: Extract feature vectors of all trajectories from different perspectives from the original event log to obtain multi-perspective trajectory features; Using a K-nearest neighbor algorithm to construct a K-nearest neighbor graph for each perspective according to the multi-perspective trajectory features, to obtain a multi-perspective graph; Using a multi-view graph encoder based on a graph convolutional network to perform a convolution operation on the multi-view graph to obtain an initial common representation; Inputting the multi-view graph into a neural network for fusion, and optimizing the structure of the fused graph through the rank constraint of the Laplacian matrix to obtain a fused graph; The initial common representation and the fusion graph are used to learn the common representation, and the multi-view graph decoder is used to reconstruct the multi-view graph. The common representation is optimized by minimizing the difference between the reconstructed multi-view graph and the original multi-view graph to obtain the final common representation. Based on the final common representation, the trajectories in the original event log are clustered.

2. The method according to claim 1, characterized in that The different perspectives include an activity perspective, a resource perspective, and a transition perspective, wherein the activity perspective, the resource perspective, and the transition perspective are used to describe each track in the original event log, wherein: The activity perspective describes the activities contained in the trajectory; The resource perspective describes the execution resources of the activities contained in the trace; The transition perspective reflects the transition relationship between activities included in the trajectory.

3. The method according to claim 1, characterized in that The nodes in the k-nearest neighbor graph represent trajectories, and the edges between the nodes represent the similarities between the trajectories. The k-nearest neighbor algorithm is used to construct a k-nearest neighbor graph for each perspective according to the multi-perspective trajectory features to obtain a multi-perspective graph, including: For each view, the similarity between trajectories is calculated based on the corresponding trajectory features, and the k nearest neighbor trajectories are selected for each trajectory, and the k nearest neighbor graph of the view is constructed using the K nearest neighbor algorithm.

4. The method according to claim 1, characterized in that The method of using a multi-view graph encoder based on a graph convolutional network to perform a convolution operation on the multi-view graph to obtain an initial public representation includes: Input the multi-view trajectory features and multi-view graphs into the multi-layer GCN and output the features of each view; The attention mechanism is introduced to learn the weight of each perspective, fuse the features of each perspective, and obtain the initial public representation.

5. The method according to claim 4, characterized in that Multi-layer GCN performs convolution operations and propagates feature information between multiple layers of GCN. Assume that the representation of the lth layer is H (l) , and its propagation rules are: The result of the first layer of GCN convolution is: in, is the multi-view trajectory feature, is the adjacency matrix of the multi-view graph, v is the view angle; Assuming that the multi-layer GCN is layer l, the result of the convolution of the l-th layer of GCN is: The initial public representation is: in, is the weight matrix for each view.

6. The method according to claim 1, characterized in that The multi-view images are fused through neural networks, including: Use neural network to fuse multi-view images. During fusion, learn the attention coefficient matrix of each image, obtain the weighted accumulation results of each image, and pass the weighted accumulation results through the linear layer to obtain the fusion image, where: The weighted cumulative results of each graph are expressed as: Among them, G (v) Represents a multi-view graph, is the attention coefficient matrix for each perspective, v is the perspective, The fusion graph is represented as: G=φ(W g G (*) +b g ) Among them, φ is the activation function of the fully connected layer, W g and b g They are the weight matrix and bias matrix in the neural network respectively; The structure of the fusion graph is optimized through the rank constraint of the Laplacian matrix, including: The fusion graph loss is constructed based on the rank constraint of the Laplace matrix, and the loss function of the fusion graph is constructed based on the fusion graph reconstruction loss and the fusion graph loss, where: The loss function of the fusion graph is expressed as: Loss G =Loss gre +λ1Loss gtr Among them, λ1 is a hyperparameter that balances the fusion image reconstruction loss and the fusion image trace loss, Loss gre The reconstruction loss for the fusion graph is expressed as: A is the adjacency matrix of the fusion graph, A (v) is the adjacency matrix of the multi-view graph, Loss gtr For the fusion trace loss, it is expressed as: Loss gtr =tr(Q T L A Q),s.t.Q T Q=I Among them, L A is the Laplacian matrix of A, Q represents the eigenvector matrix of A, the trace operator tr() calculates the trace of the matrix, and I is the identity matrix.

7. The method according to claim 1, characterized in that The multi-view image decoder is expressed as: Among them, A ′ (v) represents the adjacency matrix of each view reconstructed by the final common representation Z2, W (v) is the self-learning parameter matrix in the decoder, and v is the viewing angle; The method of learning a common representation by using an initial common representation and a fusion graph includes: Input the initial common representation and fusion graph into the multi-layer GCN to obtain the final common representation; By minimizing the difference between the reconstructed multi-view image and the original multi-view image, the public representation is optimized, including: Through back-propagation training, the difference between the reconstructed multi-view image and the original multi-view image is minimized and the common representation is optimized. The loss function during optimization is: Among them, Loss rec is the reconstruction loss for all views, A (v) is the adjacency matrix of the multi-view graph.

8. A trajectory clustering device based on graph convolutional network, characterized in that: include: A feature extraction module is used to extract feature vectors of all trajectories from different perspectives from the original event log to obtain multi-perspective trajectory features; A graph construction module, configured to construct a K nearest neighbor graph for each view according to the multi-view trajectory features using a K nearest neighbor algorithm to obtain a multi-view graph; A graph convolution module, configured to perform a convolution operation on the multi-view graph using a multi-view graph encoder based on a graph convolution network to obtain an initial common representation; A graph fusion module, used for inputting the multi-view graph into a neural network for fusion, and optimizing the structure of the fused graph through the rank constraint of the Laplace matrix to obtain a fused graph; An optimization module is used to learn a common representation using the initial common representation and the fusion graph, and reconstruct the multi-view graph using a multi-view graph decoder, and optimize the common representation by minimizing the difference between the reconstructed multi-view graph and the original multi-view graph to obtain a final common representation; A clustering module is used to cluster the trajectories in the original event log based on the final common representation.

9. The device according to claim 8, characterized in that The different perspectives include an activity perspective, a resource perspective, and a transition perspective, wherein the activity perspective, the resource perspective, and the transition perspective are used to describe each track in the original event log, wherein: The activity perspective describes the activities contained in the trajectory; The resource perspective describes the execution resources of the activities contained in the trace; The transition perspective reflects the transition relationship between activities included in the trajectory.

10. The device according to claim 8, characterized in that The nodes in the k-nearest neighbor graph represent trajectories, and the edges between the nodes represent the similarities between the trajectories. The k-nearest neighbor algorithm is used to construct a k-nearest neighbor graph for each perspective according to the multi-perspective trajectory features to obtain a multi-perspective graph, including: For each view, the similarity between trajectories is calculated based on the corresponding trajectory features, and the k nearest neighbor trajectories are selected for each trajectory, and the k nearest neighbor graph of the view is constructed using the K nearest neighbor algorithm.

Images