Data processing method and system
By segmenting the bipartite graph of the material kitting problem and processing its feature information, the problem of low solution efficiency in large-scale material kitting problems is solved, and a more efficient objective solution is obtained.
Patent Information
- Application Number
- CN202511570639.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-30
- Publication Date
- 2026-02-03
AI Technical Summary
With the accelerated intelligent transformation of the manufacturing industry, the scale of decision variables and constraints in the material kitting problem is growing exponentially, resulting in a significant decrease in the efficiency of finding the optimal solution.
The variable nodes and constraint nodes in the bipartite graph corresponding to the material kitting problem are divided into multiple subgraphs. The nodes of each subgraph are processed to obtain global feature information. Through self-attention and cross-attention processing, the global feature information of the bipartite graph is generated, and finally the target solution that satisfies the production objective is determined.
While ensuring the accuracy of the target solution, the amount of data processing is reduced and the efficiency of solving the target solution is improved.
Smart Images
Figure CN121457698A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of artificial intelligence, and in particular to a data processing method and system. BACKGROUND
[0002] At present, for the material matching problem, a bipartite graph constructed by decision variables and constraint conditions is converted into a high-dimensional hidden vector, and similarity comparison is performed with the bipartite graph high-dimensional vector of the sample in the historical set, so as to recommend the historical parameter configuration for the solver, so as to obtain the optimal solution output by the solver, so as to maximize the enterprise benefit.
[0003] However, with the acceleration of intelligent transformation of manufacturing industry, the scale of decision variables and constraint conditions in the material matching problem presents exponential growth, and may even reach millions or higher, which leads to a significant decrease in the solving efficiency of the optimal solution. SUMMARY
[0004] Therefore, the present application provides a data processing method and system, as follows:
[0005] A data processing method, comprising:
[0006] segmenting the variable nodes and constraint nodes in the bipartite graph corresponding to the material matching problem to obtain a plurality of subgraphs, each of the subgraphs comprising a plurality of variable nodes and constraint nodes, the variable nodes corresponding to production variables in the material matching problem, and the constraint nodes corresponding to matching constraint conditions in the material matching problem;
[0007] processing the nodes in each of the subgraphs to obtain global feature information corresponding to the subgraph; the global feature information at least representing the relevance between the nodes in the subgraph and the semantics of the nodes in the subgraph;
[0008] processing a sequence composed of the global feature information corresponding to the plurality of subgraphs to obtain global feature information corresponding to the bipartite graph;
[0009] determining a target solution satisfying a production target of the material matching problem according to the global feature information corresponding to the bipartite graph, the target solution comprising a variable value corresponding to each of the production variables.
[0010] The above method, preferably, processing the nodes in each of the subgraphs to obtain global feature information corresponding to the subgraph comprises:
[0011] obtaining semantic feature information of the variable nodes and semantic feature information of the constraint nodes in each of the subgraphs;
[0012] According to the semantic feature information, the association between the variable nodes and the constraint nodes in each of the subgraphs is analyzed to obtain global feature information corresponding to each of the subgraphs.
[0013] Preferably, the method further comprises:
[0014] The variable nodes in the subgraph are subjected to self-attention processing to obtain semantic feature information of the variable nodes in the subgraph.
[0015] The constraint nodes in the subgraph are subjected to self-attention processing to obtain semantic feature information of the constraint nodes in the subgraph.
[0016] Preferably, the method further comprises:
[0017] According to the semantic feature information, the variable nodes and the constraint nodes in each of the subgraphs are subjected to cross-attention processing to obtain global feature information corresponding to each of the subgraphs.
[0018] Preferably, the method further comprises:
[0019] The sequence of the global feature information corresponding to the plurality of subgraphs is subjected to self-attention processing to obtain global feature information corresponding to the bipartite graph.
[0020] Preferably, the method further comprises:
[0021] The variable nodes in the bipartite graph corresponding to the material matching problem are divided to obtain a plurality of initial graphs, each of which includes a plurality of variable nodes.
[0022] The variable nodes in different subgraphs are different.
[0023] A plurality of constraint nodes are sequentially added to each of the initial graphs to obtain a plurality of subgraphs, and each constraint node in the subgraph is connected to at least one variable node in the subgraph.
[0024] Preferably, the method further comprises:
[0025] obtaining variable semantics of production variables corresponding to each variable node in a bipartite graph corresponding to a material lot-sizing problem;
[0026] According to the variable semantics, the corresponding variable nodes in the bipartite graph are divided into pre-constructed initial graphs.
[0027] The above method preferably adds a plurality of constraint nodes to each of the initial graphs in sequence to obtain a plurality of subgraphs, comprising:
[0028] Determining the remaining constraint nodes in the bipartite graph that are not added to any initial graph;
[0029] In the remaining constraint nodes, all initial constraint nodes connected to any variable node in the current initial graph are screened out;
[0030] At least part of the initial constraint nodes are added to the current initial graph;
[0031] Wherein, the total number of nodes in the current initial graph is less than or equal to the maximum number of nodes.
[0032] A data processing system, comprising:
[0033] A graph representation network module is used to segment variable nodes and constraint nodes in a bipartite graph corresponding to a material lot-sizing problem to obtain a plurality of subgraphs, each of which includes a plurality of variable nodes and constraint nodes. The variable nodes correspond to production variables in the material lot-sizing problem, and the constraint nodes correspond to lot-sizing constraint conditions in the material lot-sizing problem. Each node in each of the subgraphs is processed to obtain global feature information corresponding to the subgraph. The global feature information represents the relevance between the nodes in the subgraph and the semantics of the nodes in the subgraph. A sequence composed of global feature information corresponding to a plurality of subgraphs is processed to obtain global feature information corresponding to the bipartite graph.
[0034] A target solution solving module is used to provide at least the global feature information corresponding to the bipartite graph to a solver;
[0035] The solver is used to output a target solution that satisfies the production target corresponding to the material lot-sizing problem according to at least the global feature information corresponding to the bipartite graph, and the target solution includes a variable value corresponding to each production variable.
[0036] The above system, preferably, the obtaining process of the graph representation network module comprises:
[0037] A graph representation network module is used to process a first bipartite graph according to a plurality of segmentation methods to obtain a plurality of first sample global feature information;
[0038] According to the plurality of first sample global feature information, the model parameters of the graph representation network module are adjusted, so that the vector distance between the global feature information output by the graph representation network module for the same bipartite graph is less than or equal to a first threshold, to obtain the graph representation network module;
[0039] and / or,
[0040] The second bipartite graph and the third bipartite graph are processed using the graph representation network module to obtain a plurality of second sample global feature information.
[0041] According to the plurality of second sample global feature information, the model parameters of the graph representation network module are adjusted, so that the vector distance between the global feature information output by the graph representation network module for different bipartite graphs is greater than a second threshold, to obtain the graph representation network module. BRIEF DESCRIPTION OF DRAWINGS
[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.
[0043] Figure 1 A flowchart of a data processing method provided by the embodiment of the present application;
[0044] Figure 2 An example graph for dividing a bipartite graph into a plurality of subgraphs in the embodiment of the present application;
[0045] Figure 3 A schematic diagram for obtaining a target solution in the embodiment of the present application;
[0046] Figure 4 A partial flowchart of a data processing method provided by the embodiment of the present application;
[0047] Figure 5 A schematic diagram for processing a subgraph to obtain global feature information in the embodiment of the present application;
[0048] Figure 6 Another partial flowchart of a data processing method provided by the embodiment of the present application;
[0049] Figure 7 Still another partial flowchart of a data processing method provided by the embodiment of the present application;
[0050] Figure 8 A structural schematic diagram of a data processing system provided by the embodiment of the present application;
[0051] Figure 9 This is a flowchart of the graph representation network module obtained in an embodiment of this application;
[0052] Figure 10 This is another flowchart of the graph representation network module in this application embodiment;
[0053] Figure 11 This is a code example diagram illustrating the process of obtaining a subgraph by segmenting a bipartite graph in an embodiment of this application.
[0054] Figure 12 This is a schematic diagram of the graph-level learning framework based on adaptive graph segmentation and subgraph sequence comparison learning in the embodiments of this application;
[0055] Figure 13 This is a schematic diagram of a framework for solving large-scale combinatorial optimization problems based on graph representation networks and KNN (K-Nearest Neighbor) parameter recommendation. Detailed Implementation
[0056] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0057] refer to Figure 1 The diagram shown is a flowchart illustrating the implementation of a data processing method provided in this application. This method can be applied to electronic devices capable of data processing, such as computers or servers. The technical solution in this embodiment is mainly used to improve the efficiency of solving the objective solution for the material kitting problem.
[0058] Specifically, the method in this embodiment may include the following steps:
[0059] Step 101: Divide the variable nodes and constraint nodes in the bipartite diagram corresponding to the material kitting problem to obtain multiple sub-graphs.
[0060] Each subgraph includes multiple variable nodes and constraint nodes. The variable nodes in each subgraph come from the bipartite graph and correspond to production variables in the material kitting problem, such as product quantity, raw material demand, order deadline, and production loss. The constraint nodes in each subgraph come from the bipartite graph and correspond to kitting constraints in the material kitting problem, such as timeline constraints, raw material constraints, capacity constraints, downtime constraints, correlation constraints, and target penalty constraints.
[0061] It should be noted that the bipartite graph corresponding to the material matching problem is constructed based on the production variables and matching constraints corresponding to the material matching problem. For example, as shown in FIG. 3, the bipartite graph includes a plurality of variable nodes and a plurality of constraint nodes, each variable node is connected to at least one constraint node, and each constraint node is connected to at least one variable node, to represent the association between the production variables and the matching constraints. Figure 2
[0062] Based on this, the bipartite graph is segmented in this embodiment, and each subgraph obtained from the bipartite graph includes variable nodes and constraint nodes. For example, as shown in FIG. 4, the bipartite graph is segmented into a plurality of subgraphs, and the variable nodes included in all subgraphs are the variable nodes in the bipartite graph, and the constraint nodes included in all subgraphs are the constraint nodes in the bipartite graph. Figure 2
[0063] Step 102: Process the nodes in each subgraph respectively to obtain global feature information corresponding to the subgraph.
[0064] The global feature information at least represents the association between the nodes in the subgraph and the semantics of the nodes in the subgraph. The association between the nodes can be understood as the association between the semantics of the variable nodes and the semantics of the constraint nodes in the subgraph.
[0065] For example, in this embodiment, the semantics of each node in each subgraph is obtained by performing semantic analysis on the description information corresponding to each variable node and constraint node in each subgraph, such as production variables and matching constraints. Then, for each subgraph, the semantics of the variable nodes in the subgraph and the semantics of the constraint nodes in the subgraph are associated and analyzed to obtain the global feature information corresponding to the subgraph.
[0066] Step 103: Process the sequence composed of the global feature information corresponding to the plurality of subgraphs to obtain the global feature information corresponding to the bipartite graph.
[0067] The global feature information corresponding to the bipartite graph represents the semantics of all nodes in the bipartite graph and the association between the nodes.
[0068] It should be noted that the global feature information corresponding to the bipartite graph is a high-dimensional vector.
[0069] Step 104: Determine a target solution satisfying the production target corresponding to the material matching problem according to the global feature information corresponding to the bipartite graph.
[0070] The target solution includes a variable value corresponding to each production variable. In the production process, the variable value in the target solution is set to the production variable, which can achieve the production target, such as the highest yield or the lowest energy consumption.
[0071] Specifically, in the embodiment, a historical set can be pre-deployed, and the historical set stores a plurality of sample feature information, each sample feature information corresponds to an optimal solution solved for a historical nesting problem, and the optimal solution includes a variable value corresponding to each production variable in the material nesting problem. Each sample feature is global feature information corresponding to a historical bipartite graph constructed for the corresponding historical nesting problem. Based on this, as shown in Figure 3 step 104, the global feature information corresponding to the bipartite graph can be provided to the solver, and the sample feature information in the historical set can be provided to the solver. The solver compares the global feature information corresponding to the bipartite graph of the material nesting problem with the sample feature information in the historical set in terms of similarity, filters out sample feature information with the highest similarity, and finally, the solver adjusts the optimal solution corresponding to the sample feature information with the highest similarity to obtain a target solution that meets the production target corresponding to the material nesting problem.
[0072] As can be seen from the above technical solutions, in the data processing method provided by the embodiment, after the variable nodes and the constraint nodes in the bipartite graph corresponding to the material nesting problem are segmented, the variable nodes and the constraint nodes in each subgraph obtained are processed respectively to obtain global feature information corresponding to the subgraph. The global feature information at least represents the association between the nodes in the subgraph and the semantics of the nodes in the subgraph. Then, the sequence composed of the global feature information corresponding to the plurality of subgraphs is processed to obtain the global feature information corresponding to the bipartite graph. Finally, the target solution that meets the production target corresponding to the material nesting problem can be determined according to the global feature information corresponding to the bipartite graph, and the target solution includes variable values corresponding to the production variables corresponding to the variable nodes. As can be seen, in the embodiment, after the plurality of subgraphs of the bipartite graph are segmented, the feature information representing the association between the nodes and the semantics of the nodes is obtained first, and then the feature information is processed in sequence. The target node is determined according to the global feature information of the bipartite graph. In this way, the accuracy of the target solution can be ensured, the data processing amount for obtaining the target solution can be reduced, the efficiency of obtaining the global feature information of the bipartite graph can be improved, and the efficiency of solving the target solution can be improved.
[0073] In an implementation manner, step 102 can obtain the global feature information corresponding to each subgraph in the following manner, as shown in Figure 4
[0074] Step 401: Obtain the semantic feature information of the variable nodes and the semantic feature information of the constraint nodes in each subgraph.
[0075] For example, as shown in Figure 5 As shown in the figure, in this embodiment, the semantics of the variable nodes in each subgraph are parsed to obtain the semantic feature information of the variable nodes, and the semantics of the constraint nodes in each subgraph are parsed to obtain the semantic feature information of the constraint nodes.
[0076] It should be noted that in this embodiment, the semantics of each variable node and each constraint node can be parsed by multiple parallel processes or threads to obtain the semantic feature information of each node.
[0077] In a specific implementation, step 401 can perform self-attention processing on each variable node in the subgraph to obtain the semantic feature information of the variable node in the subgraph. At the same time, self-attention processing can be performed on each constraint node in the subgraph to obtain the semantic feature information of the constraint node in the subgraph.
[0078] For example, in this embodiment, multiple parallel self-attention processing modules can be used to perform semantic parsing on the corresponding variable nodes and constraint nodes respectively, so as to obtain the semantic feature information of each variable node and the semantic feature information of each constraint node.
[0079] Each subgraph corresponds to two self-attention processing modules. One module performs self-attention processing on the variable nodes in the subgraph to obtain their semantic features, while the other module performs self-attention processing on the constraint nodes to obtain their semantic features. Each self-attention processing module runs in parallel, thereby improving the efficiency of obtaining the semantic features of each node.
[0080] Step 402: Based on the semantic feature information, analyze the relationship between variable nodes and constraint nodes in each subgraph to obtain the global feature information corresponding to each subgraph.
[0081] For example, such as Figure 5 As shown in the figure, in this embodiment, for each subgraph, the correlation between the variable nodes and constraint nodes in the subgraph can be analyzed based on the semantic feature information of the variable nodes and the semantic feature information of the constraint nodes in the subgraph, so as to obtain the global feature information corresponding to the subgraph.
[0082] It should be noted that in this embodiment, the correlation between variable nodes and constraint nodes in each subgraph can be parsed through multiple parallel processes or threads to obtain the global feature information corresponding to each subgraph.
[0083] In the specific implementation, step 402 can perform cross-attention processing on the variable nodes and constraint nodes in each subgraph based on semantic feature information to obtain the global feature information corresponding to each subgraph.
[0084] For example, in this embodiment, the association between the variable nodes and the constraint nodes in each subgraph can be analyzed by multiple cross-attention processing modules running in parallel to obtain the global feature information corresponding to each subgraph.
[0085] Each subgraph corresponds to a cross-attention processing module, and the cross-attention processing module performs cross-attention processing on the variable nodes and the constraint nodes in the corresponding subgraph to obtain the global feature information corresponding to the subgraph. Each cross-attention processing module runs in parallel, so that the efficiency of obtaining the global feature information corresponding to each subgraph can be improved by the cross-attention processing modules running in parallel.
[0086] In an implementation manner, when obtaining the global feature information corresponding to the bipartite graph, the sequence composed of the global feature information corresponding to the multiple subgraphs can be self-attention processed to obtain the global feature information corresponding to the bipartite graph.
[0087] For example, in this embodiment, the sequence composed of the global feature information corresponding to the multiple subgraphs can be self-attention processed by a self-attention processing module to obtain the global feature information corresponding to the bipartite graph.
[0088] It should be noted that in this embodiment, multiple subgraphs are obtained by bipartite graph segmentation, and after obtaining the global feature information corresponding to each subgraph, the global feature information corresponding to the bipartite graph can be obtained by sequential self-attention processing, and the global feature information corresponding to the bipartite graph represents the global information in the bipartite graph.
[0089] Based on this, in this embodiment, after the subgraph segmentation, the global feature information of the subgraph is obtained respectively, and then the global feature information of the subgraph is processed sequentially, which can reduce the data processing amount of obtaining the global feature information of the bipartite graph, thereby improving the efficiency of obtaining the global feature information of the bipartite graph, and further improving the solving efficiency of the target solution of the material matching problem.
[0090] In an implementation manner, in step 101, when the bipartite graph is segmented, the segmentation can be implemented in the following manner, as shown in Figure 6
[0091] Step 601: dividing the variable nodes in the bipartite graph corresponding to the material matching problem to obtain multiple initial graphs.
[0092] Each initial graph includes multiple variable nodes. The variable nodes in different subgraphs are different, so the variable nodes in different subgraphs are not repeated, which avoids inaccurate situations when obtaining the global feature information subsequently.
[0093] Specifically, in step 601, the variable semantics of the production variables corresponding to each variable node in the bipartite graph corresponding to the material matching problem can be obtained first, for example, the semantic recognition model can be used to analyze the variable semantics of each production variable, such as the semantic information of the decision variables such as product quantity, raw material demand, order deadline and production loss, and then the corresponding variable nodes in the bipartite graph are divided into the pre-constructed initial graph according to the variable semantics.
[0094] For example, in the embodiment, a corresponding number of variable nodes can be randomly selected from the bipartite graph according to the variable semantics and divided into the corresponding initial graph, thereby ensuring the randomness of the variable nodes in the subsequent obtained subgraph.
[0095] Wherein, the set of variable nodes in the bipartite graph can be represented by V, V = {vi|i = 1, 2, 3,..., n}, n is the number of variable nodes in the bipartite graph, and vi is the i-th variable node. The set of constraint nodes in the bipartite graph can be represented by C, C = {cj|j = 1, 2, 3,..., m}, m is the number of constraint nodes in the bipartite graph, and cj is the j-th constraint node. The set of edges between the variable nodes and the constraint nodes in the bipartite graph can be represented by E, (vi, cj) represents the connection relationship between the i-th variable node and the j-th constraint node.
[0096] It should be noted that the maximum node number can be pre-set in the embodiment, which represents the node size of the subgraph, and can be represented by gsize. The total number of variable nodes and constraint nodes in the subgraph is less than or equal to the maximum node number. Based on this, the total subgraph number is determined based on the maximum node number in the embodiment, and the total subgraph number can be represented by S, S = ⌈(|V| + |C|) / gsize⌉, |V| + |C| is the total number of all nodes in the bipartite graph.
[0097] Based on this, in step 601, the variable nodes V in the bipartite graph can be randomly divided into S non-intersecting subsets according to the variable semantics (such as material category, order category, etc.) according to the total subgraph number S, and each subset is an initial graph, represented by V1, V2,..., V s .
[0098] Step 602: adding a plurality of constraint nodes to each initial graph in turn to obtain a plurality of subgraphs.
[0099] Wherein, each constraint node in the subgraph is connected to at least one variable node in the subgraph, thereby based on the connection state of the variable nodes and the constraint nodes in the bipartite graph, the variable nodes and the constraint nodes in the divided subgraph exist connection, to ensure the connectivity between the nodes in the subgraph.
[0100] It should be noted that in the embodiment, the constraint nodes are sequentially added to each initial graph in order to obtain the subgraph corresponding to each initial graph.
[0101] In an implementation manner, the plurality of constraint nodes can be added to each initial graph in step 602 to obtain the plurality of subgraphs, as shown in the following formula: Figure 7
[0102] Step 701: determining the remaining constraint nodes in the bipartite graph that are not added to any initial graph.
[0103] In the initial stage, the remaining constraint nodes include all the constraint nodes in the bipartite graph.
[0104] Step 702: screening out all the initial constraint nodes connected to any variable node in the current initial graph from the remaining constraint nodes.
[0105] Specifically, in the embodiment, each initial graph is taken as the current initial graph, and all the initial constraint nodes connected to any variable node in the current initial graph are screened out for each current initial graph.
[0106] The set composed of the initial constraint nodes can be represented by C s . For example, for the i-th initial graph (the subset Vᵢ), the constraint nodes connected to any variable node in Vᵢ are determined to form the set C s .
[0107] Step 703: adding at least part of the initial constraint nodes to the current initial graph.
[0108] The total number of nodes in the current initial graph is less than or equal to the maximum number of nodes.
[0109] It can be seen that in the embodiment, the variable nodes in the bipartite graph are divided in advance, and then the constraint nodes connected to the variable nodes are preferentially added to the initial graph based on the variable nodes in the initial graph. Therefore, the constraint nodes and the variable nodes in the divided subgraph are connected by edges, which can ensure the connectivity of the nodes in the divided subgraph, thereby improving the accuracy of obtaining the global feature information of the bipartite graph and the solving accuracy of the target solution.
[0110] Specifically, at least part of the initial constraint nodes can be added to the current initial graph in step 703 in the following manner:
[0111] First, the size relationship between the number of initial constraint nodes and the maximum number of constraint nodes corresponding to the current initial graph is determined, and based on the size relationship, there are two adding manners:
[0112] In a case where the number of initial constraint nodes is greater than or equal to the maximum constraint node number corresponding to the current initial graph, constraint nodes of the maximum constraint node number are selected from the initial constraint nodes and added to the current initial graph.
[0113] In a case where the number of initial constraint nodes is less than the maximum constraint node number corresponding to the current initial graph, candidate constraint nodes of the difference number are selected from the constraint nodes contained in the bipartite graph and added to the current initial graph, and the initial constraint nodes are added to the current initial graph.
[0114] The maximum constraint node number is the difference between the maximum node number and the number of variable nodes in the current initial graph. The difference number is the difference between the maximum constraint node number and the number of initial constraint nodes.
[0115] It should be noted that the candidate constraint nodes are different from the initial constraint nodes. The candidate constraint nodes can be constraint nodes in the bipartite graph that are connected to any variable node in the initial graph but are different from the initial constraint nodes (because as constraint nodes are added to each initial graph, the remaining constraint nodes can become fewer and fewer, so if the initial constraint nodes are too few, the constraint nodes are reselected from the bipartite graph to be added).
[0116] It can be seen that in the embodiment, the maximum node number is taken as the upper limit, the maximum number of constraint nodes that can be added to the initial graph is determined based on the number of variable nodes that have been divided in the initial graph, that is, the maximum constraint node number, and thus after the initial constraint nodes that are connected to any variable node in the initial graph are selected, the initial constraint nodes are preferentially added to the initial graph. If the initial constraint nodes are too many (that is, greater than or equal to the maximum constraint node number corresponding to the current initial graph), constraint nodes of the maximum constraint node number are selected from the initial constraint nodes and added to the current initial graph. If the initial constraint nodes are too few (that is, less than the maximum constraint node number corresponding to the current initial graph), in addition to adding all the initial constraint nodes to the current initial graph, candidate constraint nodes of the difference number are also selected from the bipartite graph and added to the current initial graph, and thus the same constraint nodes can exist between different subgraphs.
[0117] Reference Figure 8 A structure schematic diagram of a data processing system provided in the embodiment of the present application, which can be deployed in an electronic device capable of data processing, such as a computer or a server. The technical solution in the embodiment is mainly used to improve the solving efficiency of the target solution for the material matching problem.
[0118] Specifically, the data processing system in the embodiment can include the following structure:
[0119] The graph representation network module 801 is configured to split variable nodes and constraint nodes in a bipartite graph corresponding to a material matching problem to obtain a plurality of subgraphs, each of the subgraphs including a plurality of variable nodes and constraint nodes, the variable nodes corresponding to production variables in the material matching problem, and the constraint nodes corresponding to matching constraint conditions in the material matching problem; process the nodes in each of the subgraphs to obtain global feature information corresponding to the subgraph; and process a sequence of global feature information corresponding to the plurality of subgraphs to obtain global feature information corresponding to the bipartite graph.
[0120] The target solution solving module 802 is configured to provide the global feature information corresponding to the bipartite graph to a solver 803.
[0121] The solver 803 is configured to output a target solution satisfying a production target corresponding to the material matching problem according to the global feature information corresponding to the bipartite graph, the target solution including variable values corresponding to each of the production variables.
[0122] As can be seen from the above technical solutions, in the data processing system provided by the embodiments of the present application, after the variable nodes and constraint nodes in the bipartite graph corresponding to the material matching problem are split, the variable nodes and constraint nodes in each of the obtained subgraphs are processed to obtain global feature information corresponding to the subgraph, the global feature information at least representing the relevance between the nodes in the subgraph and the semantics of the nodes in the subgraph, then the sequence of global feature information corresponding to the plurality of subgraphs is processed to obtain global feature information corresponding to the bipartite graph, and finally the target solution satisfying the production target corresponding to the material matching problem can be determined according to the global feature information corresponding to the bipartite graph, the target solution including variable values corresponding to the production variables corresponding to the variable nodes. It can be seen that, in the embodiments, after the plurality of subgraphs of the bipartite graph are split, the feature information representing the relevance between the nodes and the semantics of the nodes is obtained first, and then the feature information is processed in sequence, the target node is determined according to the global feature information of the bipartite graph, so that the accuracy of the target solution is ensured, the data processing amount for obtaining the target solution is reduced, the efficiency of obtaining the global feature information of the bipartite graph is improved, and the efficiency of solving the target solution is further improved.
[0123] In an implementation manner, the obtaining process of the graph representation network module can include the following flow as shown in Figure 9
[0124] Step 901: using the graph representation network module to process the first bipartite graph according to a plurality of splitting manners to obtain a plurality of first sample global feature information.
[0125] The segmentation manner refers to a subgraph segmentation manner of the bipartite graph by the graph representation network module. Therefore, in the process of segmenting the subgraph and dividing the initial graph and adding the constraint node, randomness exists. Therefore, each subgraph segmentation of the graph representation network module is a different segmentation manner. Based on this, the global feature information obtained by using the graph representation network module to process the same bipartite graph may be different in the embodiment.
[0126] Step 902: adjusting the model parameters of the graph representation network module according to the plurality of first sample global feature information, so that the vector distance between the global feature information output by the graph representation network module for the same bipartite graph is less than or equal to the first threshold, to obtain the graph representation network module.
[0127] Specifically, in the embodiment, a loss function can be constructed, and a loss value is obtained according to the plurality of first sample global feature information. The loss value represents the vector distance between the plurality of first sample global feature information. In the embodiment, the model parameters of the graph representation network module are adjusted based on the loss value, so that the vector distance between the global feature information output by the graph representation network module for the same bipartite graph is less than or equal to the first threshold. Therefore, the obtained graph representation network module can output global feature information with a smaller vector distance for the same bipartite graph. Even if the subgraph segmentation has randomness, the accuracy of the global feature information output by the graph representation network module will not be affected.
[0128] In an implementation manner, the obtaining process of the graph representation network module can further include the following flow, as shown in Figure 10
[0129] Step 1001: using the graph representation network module to process the second bipartite graph and the third bipartite graph respectively to obtain a plurality of second sample global feature information.
[0130] The second bipartite graph and the third bipartite graph are different.
[0131] Step 1002: adjusting the model parameters of the graph representation network according to the plurality of second sample global feature information, so that the vector distance between the global feature information output by the graph representation network module for different bipartite graphs is greater than a second threshold, to obtain the graph representation network module.
[0132] The second threshold is greater than the first threshold.
[0133] Specifically, in the embodiment, the model parameters of the graph representation network module are adjusted based on the loss value, so that the vector distance between the global feature information output by the graph representation network module for different bipartite graphs is greater than the second threshold. Thus, the graph representation network module obtained can output global feature information with a greater vector distance for different bipartite graphs, which can improve the accuracy of the global feature information output by the graph representation network module.
[0134] In an implementation manner, when the graph representation network module 801 processes nodes in each of the subgraphs respectively to obtain global feature information corresponding to the subgraph, the graph representation network module 801 is specifically configured to: obtain semantic feature information of variable nodes and semantic feature information of constraint nodes in each of the subgraphs; and analyze the association between the variable nodes and the constraint nodes in each of the subgraphs according to the semantic feature information to obtain global feature information corresponding to each of the subgraphs.
[0135] In an implementation manner, when the graph representation network module 801 obtains the semantic feature information of the variable nodes and the semantic feature information of the constraint nodes in each of the subgraphs, the graph representation network module 801 is specifically configured to: perform self-attention processing on the variable nodes in the subgraph to obtain the semantic feature information of the variable nodes in the subgraph; and perform self-attention processing on the constraint nodes in the subgraph to obtain the semantic feature information of the constraint nodes in the subgraph.
[0136] In an implementation manner, when the graph representation network module 801 analyzes the association between the variable nodes and the constraint nodes in each of the subgraphs according to the semantic feature information to obtain global feature information corresponding to each of the subgraphs, the graph representation network module 801 is specifically configured to: perform cross-attention processing on the variable nodes and the constraint nodes in each of the subgraphs according to the semantic feature information to obtain global feature information corresponding to each of the subgraphs.
[0137] In an implementation manner, when the graph representation network module 801 processes a sequence composed of the global feature information corresponding to the plurality of subgraphs to obtain global feature information corresponding to the bipartite graph, the graph representation network module 801 is specifically configured to: perform self-attention processing on the sequence composed of the global feature information corresponding to the plurality of subgraphs to obtain global feature information corresponding to the bipartite graph.
[0138] In an implementation manner, when the graph representation network module 801 segments the variable nodes and the constraint nodes in the bipartite graph corresponding to the material matching problem to obtain a plurality of subgraphs, the graph representation network module 801 is specifically configured to: divide the variable nodes in the bipartite graph corresponding to the material matching problem to obtain a plurality of initial graphs, and each of the initial graphs includes a plurality of variable nodes; different variable nodes are included in different subgraphs; add a plurality of constraint nodes to each of the initial graphs in sequence to obtain a plurality of subgraphs; and each constraint node in the subgraph is connected to at least one variable node in the subgraph.
[0139] In the method, when the network module 801 divides the variable nodes in the bipartite graph corresponding to the material matching problem to obtain a plurality of initial graphs, the following is specifically used: obtaining variable semantics of a production variable corresponding to each variable node in the bipartite graph corresponding to the material matching problem; and dividing the corresponding variable nodes in the bipartite graph into the pre-constructed initial graphs according to the variable semantics.
[0140] In the method, when the network module 801 adds a plurality of constraint nodes to each of the initial graphs in sequence to obtain a plurality of sub-graphs, the following is specifically used: determining remaining constraint nodes in the bipartite graph that are not added to any initial graph; screening out, from the remaining constraint nodes, initial constraint nodes that are connected to any variable node in a current initial graph; and adding at least part of the initial constraint nodes to the current initial graph, wherein a total number of nodes in the current initial graph is less than or equal to a maximum number of nodes.
[0141] It should be noted that the specific implementation of the graph representation network module can refer to the corresponding content in the foregoing, which will not be described in detail here.
[0142] Taking the material matching problem as an example, the data processing in the present application is described as follows:
[0143] The present application proposes an adaptive graph partitioning and hierarchical contrastive learning framework to solve the general representation learning problem of super-large-scale combinatorial optimization problems, and the core innovation of the present application includes two key technical points:
[0144] I. Adaptive graph partitioning algorithm:
[0145] The present application is based on a conditional selection mechanism of connected constraints: a unilateral node random subset A (i.e., a variable node, the scale is adjustable) is selected, and only a candidate node set B (i.e., a constraint node) that is directly connected to A is allowed to be selected, which can be represented as , where u is a variable node, v is a constraint node, and V2 is a set of constraint nodes of the bipartite graph, that is, C in the foregoing. B is selected as much as possible without repetition, but a small amount of repetition is allowed (which can be removed when calculating the aggregated features in the forward direction). Experiments show that this will not greatly interfere with the learning of the whole graph representation.
[0146] In this way, it can be ensured that each sub-graph is at least a connected graph, and sufficient randomness is ensured. At the same time, the present application also strictly limits the proportion of variable and constraint nodes in each sub-graph, ensures that the node distribution is as detailed as possible as the original bipartite graph, and thus ensures that the representation learning of different sub-graphs has a certain comparability; details are shown in Figure 11 .
[0147] It can be seen that the application can realize the balance of "randomness" and "connectivity", and proposes a bipartite graph partitioning method based on the material matching problem, which lays a foundation for super large scale graph learning.
[0148] The adaptive graph partitioning algorithm includes the following steps (for reference Figure 11 ):
[0149] (1) Input preparation stage:
[0150] Input variable node set: V = {vi | i = 1, 2, 3,..., n};
[0151] Input constraint node set: C = {cj | j = 1, 2, 3,..., m};
[0152] Input edge set E: Indicates the connection relationship between variable nodes and constraint nodes.
[0153] In addition, input the maximum subgraph size parameter: gsize (used to control the maximum number of nodes in each partition).
[0154] (2) Calculate the total number of subgraphs S:
[0155] According to the formula S = ⌈(|V| + |C|) / gsize⌉, calculate S. This value represents the minimum number of partitions (i.e. the number of subgraphs) required to divide the entire bipartite graph node set into several subgraphs, ensuring that the number of nodes in each subgraph does not exceed gsize.
[0156] (3) Randomly partition the variable node set V:
[0157] Randomly partition the variable node set V into S mutually disjoint subsets according to the decision variable semantics of the material matching problem (such as material category, order category, etc.): V1, V2,..., VS. s .
[0158] Among them, this step is the specific implementation of the "select single-sided node random subset A" mentioned in the previous text, where each Vi is a "single-sided node random subset A", and its size can be adjusted (determined by gsize and S).
[0159] (4) Construct subgraph Pi (i from 1 to S loop execution):
[0160] For the current subgraph i:
[0161] A, determine the constraint node set C i directly connected to V s .
[0162] For example, filter out all the nodes connected to any node in V i by edges from the constraint node set C s .
[0163] B. Determine whether the lower limit of the subgraph size is met:
[0164] If |C s |≥gsize-|V i |, it means that only the constraint nodes directly connected to V i are enough to fill the target size gsize. At this time, randomly select (gsize-|V s |) constraint nodes from C i , denoted as C t , then add C t to the current subgraph P i (P i .add(C t )) and remove these constraint nodes from the global constraint node set C (C.remove(C t )).
[0165] Otherwise (i.e., |C s |<gsize-|V i |), it means that only the constraint nodes directly connected to V i are not enough to fill the subgraph size. At this time, additionally select (gsize-|V i |-|C s |) nodes from the remaining constraint node set C-C s , which ensure connectivity as much as possible, denoted as C t . Then, add C s and C t to the current partition P i (P i .add(C s +C t )) and remove C s and C t from the global constraint node set C (C.remove(C s +C t )).
[0166] (5) Output the final subgraph result:
[0167] After the loop ends, S subgraphs are obtained: P1, P2,..., P s , each containing a set of variable nodes V i and a set of constraint nodes (C t or C s +C t), which jointly constitute a subgraph. Based on this, the output result is a partition node set P = {P1, P2, …, P s}.
[0168] II. Hierarchical contrastive learning architecture
[0169] Traditional graph representation learning methods face two major challenges: one is that directly processing the whole graph leads to memory explosion and poor efficiency; and simple aggregation of node features will lose global structure information. This application proposes a "divide and conquer" hierarchical learning paradigm to build a learning path for super large-scale graph representation (such as Figure 12 ) :
[0170] (1) Multi-view encoding of subgraph semantic alignment: Each segmented subgraph is first processed by a graph structure feature extraction module (self-attention processing module and cross-attention processing module): Graph Transformer architecture can be used to model the relevance between nodes from different perspectives but with semantic alignment through self-attention on single nodes and cross-attention between two nodes, thereby obtaining the global representation of each subgraph (aggregating different nodes), i.e. the global feature information corresponding to the subgraph in the previous section;
[0171] (2) Sequential graph information construction: The global representation of all subgraphs is regarded as a special "word" and input into the Transformer sequence model for semantic splicing. It can be understood that: the whole graph is regarded as a sentence, and each subgraph is regarded as a word. Different graph segmentation schemes are different expressions of the same semantics. Self-attention can model and infer the representation of the whole graph through its powerful long sequence modeling capability.
[0172] (3) Multi-view contrastive learning: By automatically generating several different segmentation schemes for the same problem (equivalent to multiple expressions of a sentence), the model learns to identify essential features. For example, for the alignment problem, whether the graph is segmented by material type or by order priority, the model can extract the invariant global representation. The essence of contrastive learning is that the same graph with different segmentation views is regarded as a positive sample, and its representation is as close as possible, while the global representation of other problem instances is regarded as a negative sample; through this adaptive contrastive learning, effective global hierarchical representation can be learned under the premise of balancing efficiency and performance.
[0173] As can be seen, the application can realize hierarchical knowledge extraction and representation learning of "local features-global semantics", and adaptively learn the global hierarchical representation of super large-scale graphs without external labels.
[0174] Specifically, the "divide and conquer" hierarchical learning method proposed in this scheme is as follows:
[0175] The present application constructs a complete path suitable for ultra-large scale graph representation learning, as shown in Figure 12 By dividing the full graph into multiple subgraphs, modeling local features and global semantics layer by layer, and finally realizing efficient and expressive full graph representation learning. The specific detailed execution steps are as follows:
[0176] (1) Multi-view encoding of subgraph semantic alignment
[0177] Input object: Each subgraph generated by the adaptive graph segmentation algorithm (such as subgraph Figure 1 , subgraph Figure 2 , …, subgraph M).
[0178] Processing module: Graph structure feature extraction module, using Graph Transformer architecture.
[0179] Specific execution steps:
[0180] a. Perform self-attention operation on single-sided nodes (variable nodes V or constraint nodes C) in the subgraph, to model the relevance and importance distribution within the same type of node.
[0181] b. Perform cross-attention operation between variable nodes and constraint nodes, including V→C and C→V attention interaction, to model the dependency relationship between the two types of nodes.
[0182] c. All attention mechanisms share parameters to ensure semantic consistency when encoding different subgraphs.
[0183] d. Through the above multi-view attention mechanism (self-attention + cross-attention), the relevance between nodes is modeled to form a semantically aligned local representation.
[0184] e. Finally, output the global representation vector of each subgraph.
[0185] This step realizes the modeling goal of "different perspectives but semantic alignment", that is, each subgraph is cut from different angles, but its encoding process maintains consistent semantic space, which is convenient for subsequent splicing and comparison.
[0186] (2) Sequential full graph information construction, as shown in Figure 12
[0187] Input object: The global representation vector set obtained after encoding all subgraphs (such as subgraph Figure 1 representation, subgraph Figure 2 representation, …, subgraph M representation).
[0188] Processing approach: The global representation of each subgraph is regarded as a "word", and the whole graph is regarded as a "sentence" composed of these "words".
[0189] Model architecture: Input into the Transformer sequence model for semantic splicing.
[0190] Specific implementation steps:
[0191] a. Arrange the M subgraph representations in any order to form a sequence of length M.
[0192] b. Input into the Transformer encoder to model the long-distance dependency between "words" (i.e. subgraph representations, which are the global feature information corresponding to the subgraphs mentioned earlier) in the sequence using its self-attention (sub-graphs) mechanism.
[0193] c. Through the context awareness ability of the Transformer, the overall semantics of the whole graph is modeled and inferred.
[0194] d. After pooling, a unified whole graph level representation vector is finally output (such as Figure 12 "whole graph representation" in Chinese, which is the global feature information corresponding to the bipartite graph mentioned earlier).
[0195] Note: Different graph partitioning schemes correspond to different "word arrangement orders", i.e. different expressions of the same semantics.
[0196] (3) Multi-view contrastive learning
[0197] Learning goal: Let the graph representation network module learn to identify the essential structural features of the graph, rather than relying on a specific partitioning method.
[0198] Specific implementation steps:
[0199] a. Generate several different graph partitioning schemes for the same problem instance (equivalent to "multiple expressions of a sentence").
[0200] b. Perform representation learning for each partitioning scheme to obtain the corresponding subgraph representation.
[0201] c. In the graph-level contrastive learning framework: the full graph representations obtained from different partition views of the same original bipartite graph (i.e., the first bipartite graph in the foregoing) are regarded as positive sample pairs, and the loss function is used to make the representations as close as possible (i.e., the vector distance is less than or equal to the first threshold). The full graph representations obtained from other different problem instances (such as different matching problems, i.e., the second bipartite graph and the third bipartite graph in the foregoing) are regarded as negative samples, and the loss function is used to make the representations as far apart as possible (i.e., the vector distance is greater than the second threshold).
[0202] d. Through this adaptive contrastive learning mechanism, the model learns full graph-level representations with generalization ability without external labels.
[0203] e. The entire process is efficiently executed under the support of the "parallel acceleration" module, ensuring training efficiency.
[0204] The essence of contrastive learning is to "pull the same graph and different views closer, and push the different graphs farther apart", thereby forcing the model to focus on the internal structure of the graph rather than the surface partition method.
[0205] As can be seen, the technical solution of the present application has the following advantages:
[0206] 1. Super large-scale processing capability: theoretically, it can support representation learning of combinatorial optimization problems of any node size, and dynamically adjust the subgraph size according to resource constraints, breaking through the constraints of graphics processors and video memory.
[0207] 2. Structure-preserving partitioning: the connectivity constraint mechanism ensures subgraph connectivity, solving the problem of topological structure destruction caused by random partitioning.
[0208] 3. Cross-scene generalization performance: the learned universal graph representation extraction network can be used for different downstream optimization problems such as material matching.
[0209] 4. Meet the industrial-level response demand: single inference time is in the order of seconds (million node level), and optimal parameter recommendation can improve the solver's solving efficiency by several times, meeting the production system response demand;
[0210] 5. Parallel acceleration: in the subgraph partitioning and forward processing stage, multiple graphics processors can be used to accelerate inference, further improving response speed.
[0211] For example, the present application proposes a graph representation learning framework for super large-scale material matching problems in real-world scenarios to achieve optimal configuration of solving parameters and significantly improve the solving efficiency of complex matching problems. The overall architecture is as shown in Figure 13
[0212] 1) System input combinatorial optimization problem: an industrial-level material matching problem instance containing millions of nodes (materials / orders);
[0213] 2) Graph partitioning stage: the original problem is randomly decomposed into k connected subgraphs by a topology-aware partitioning algorithm, each subgraph tries to maintain the complete material-order supply chain relationship as much as possible, and tries to minimize the damage to the graph structure of the bipartite graph;
[0214] 3) Feature extraction stage: in the graph representation network module, the Graph Transformer architecture is adopted: the subgraph global dependencies between variable nodes, between constraint nodes, and between variable nodes-constraint nodes in each subgraph are constructed from different perspectives. After obtaining the representation of each subgraph, it is transmitted into the upper subgraph-level Transformer network to construct the dependency relationship on a larger view, and finally the full graph representation corresponding to the input problem instance is obtained;
[0215] 4) In the parameter recommendation stage: compare the full graph representation with the set of hidden vectors in the reference set (sample1 to n), and use KNN algorithm to select the closest reference set sample, and recommend its optimal parameter configuration to the current sample.
[0216] It should be noted that the forward reasoning and parameter selection can be completed in seconds (million-level nodes).
[0217] 5) Solver solving stage: input the original problem and the selected optimal parameter combination into the solver to solve the original problem.
[0218] Experiments show that the solving efficiency can be improved by more than 6-8 times compared with the default parameter configuration, so as to ensure the performance without affecting the advancement of the production process.
[0219] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same and similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.
[0220] The skilled person can further realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in the present text can be realized in electronic hardware, computer software or a combination of both. In order to clearly show the interchangeability of hardware and software, the components and steps of each example have been described in the above description. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.
[0221] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.
[0222] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A data processing method, comprising: The variable nodes and constraint nodes in the bipartite graph corresponding to the material kitting problem are divided to obtain multiple subgraphs. Each subgraph includes multiple variable nodes and constraint nodes. The variable nodes correspond to the production variables in the material kitting problem, and the constraint nodes correspond to the kitting constraints in the material kitting problem. Each node in the subgraph is processed separately to obtain the global feature information corresponding to the subgraph; The global feature information at least characterizes the relationships between nodes in the subgraph and the semantics of the nodes in the subgraph; The sequence of global feature information corresponding to multiple subgraphs is processed to obtain the global feature information corresponding to the bipartite graph; Based on the global feature information corresponding to the bipartite graph, a target solution that satisfies the production objective corresponding to the material kitting problem is determined. The target solution includes the variable value corresponding to each of the production variables.
2. The method according to claim 1, wherein each node in the subgraph is processed to obtain global feature information corresponding to the subgraph, including: Obtain the semantic feature information of the variable nodes and the semantic feature information of the constraint nodes in each of the subgraphs; Based on the semantic feature information, the correlation between variable nodes and constraint nodes in each subgraph is analyzed to obtain the global feature information corresponding to each subgraph.
3. The method according to claim 2, obtaining the semantic feature information of the variable nodes and the semantic feature information of the constraint nodes in each of the subgraphs, includes: Self-attention processing is performed on the variable nodes in the subgraph to obtain the semantic feature information of the variable nodes in the subgraph; Self-attention processing is performed on the constraint nodes in the subgraph to obtain the semantic feature information of the constraint nodes in the subgraph.
4. The method according to claim 2, wherein the association between variable nodes and constraint nodes in each subgraph is parsed based on the semantic feature information to obtain global feature information corresponding to each subgraph, including: Based on the semantic feature information, cross-attention processing is performed on the variable nodes and constraint nodes in each subgraph to obtain the global feature information corresponding to each subgraph.
5. The method according to claim 1, processing a sequence of global feature information corresponding to multiple subgraphs to obtain global feature information corresponding to the bipartite graph, comprising: Self-attention processing is performed on the sequence of global feature information corresponding to multiple subgraphs to obtain the global feature information corresponding to the bipartite graph.
6. The method according to claim 1, wherein the variable nodes and constraint nodes in the bipartite graph corresponding to the material kitting problem are segmented to obtain multiple subgraphs, including: The variable nodes in the bipartite diagram corresponding to the material kitting problem are divided to obtain multiple initial diagrams, each of which includes multiple variable nodes; The different subgraphs contain different variable nodes; Multiple constraint nodes are added sequentially to each initial graph to obtain multiple subgraphs; each constraint node in the subgraph is connected to at least one variable node in its respective subgraph.
7. The method according to claim 6, wherein the variable nodes in the bipartite diagram corresponding to the material kitting problem are divided to obtain multiple initial diagrams, including: Obtain the variable semantics of the production variables corresponding to each variable node in the bipartite diagram corresponding to the material kitting problem; According to the variable semantics, the corresponding variable nodes in the bipartite graph are assigned to the pre-constructed initial graph.
8. The method according to claim 6, wherein multiple constraint nodes are added sequentially to each of the initial graphs to obtain multiple subgraphs, comprising: Determine that there are no remaining constraint nodes in the bipartite graph that have been added to any initial graph; Among the remaining constraint nodes, select all initial constraint nodes that have an edge connection with any variable node in the current initial graph; Add at least some of the constraint nodes from the initial constraint nodes to the current initial graph; Wherein, the total number of nodes in the current initial graph is less than or equal to the maximum number of nodes.
9. A data processing system, comprising: The graph representation network module is used to segment the variable nodes and constraint nodes in the bipartite graph corresponding to the material kitting problem to obtain multiple subgraphs. Each subgraph includes multiple variable nodes and constraint nodes. The variable nodes correspond to the production variables in the material kitting problem, and the constraint nodes correspond to the kitting constraints in the material kitting problem. The nodes in each subgraph are processed to obtain the global feature information corresponding to the subgraph. Global feature information characterizes the relationships between nodes in the subgraph and the semantics of the nodes in the subgraph; the sequence of global feature information corresponding to multiple subgraphs is processed to obtain the global feature information corresponding to the bipartite graph; The objective solution solving module is used to provide at least the global feature information corresponding to the bipartite graph to the solver; The solver is configured to output a target solution that satisfies the production objective corresponding to the material kitting problem, based at least on the global feature information corresponding to the bipartite graph. The target solution includes the variable value corresponding to each of the production variables.
10. The system according to claim 9, wherein the process of obtaining the graph representation network module includes: The graph representation network module is used to process the first binary graph according to various segmentation methods to obtain multiple global feature information of the first sample; Based on the global feature information of the multiple first samples, the model parameters of the graph representation network module are adjusted so that the vector distance between the global feature information output by the graph representation network module for the same bipartite graph is less than or equal to a first threshold, so as to obtain the graph representation network module. And / or, The graph representation network module is used to process the second and third bipartite graphs respectively to obtain global feature information of multiple second samples; Based on the global feature information of the multiple second samples, the model parameters of the graph representation network module are adjusted so that the vector distance between the global feature information output by the graph representation network module for different bipartite graphs is greater than a second threshold, thereby obtaining the graph representation network module.