A method and system for mining high-value collaborative clusters for industrial business processes

Abstract

A method and system for mining high-value collaborative clusters in industrial business processes is disclosed. The method includes: constructing a distance matrix of the original graph network to build a weighted complete graph; constructing a graph neural network encoder based on a multi-hop distance decay mechanism, performing multi-layer feature propagation and aggregation on the original graph network according to the weighted complete graph to obtain node embedding representations that integrate multi-hop global topological information and freezing them to form a global node embedding table; constructing collaborative clusters using an autoregressive reinforcement learning model; using the currently selected node sequence as the state and unselected nodes as candidate actions, the agent performs multi-round sequence decision-making based on the state transformed by the embedding layer and the global node embedding table, using an autoregressive cross-graph action selector, and adding nodes to the current subgraph in each round; calculating reward values ​​and training the autoregressive reinforcement learning model based on the changes in internal association conflicts of the current node subgraph before and after node addition and the stability of future subgraph expansion.

Description

Technical Field

[0001] This invention relates to the field of graph data mining technology, specifically to a method and system for mining high-value collaborative clusters for industrial business processes. Background Technology

[0002] The industrial chain is a complex ecological network composed of multiple links, including upstream raw material supply, midstream manufacturing, and downstream distribution, retail, and supporting services. The entire operation of the industrial chain revolves around the industry's business processes, with market players in each link forming multi-dimensional connections through supply and demand cooperation, technological collaboration, and resource sharing. Existing industrial chain collaboration models are still concentrated on adjacent links (such as upstream raw material suppliers and midstream manufacturers), failing to develop a holistic collaborative layout around the industry's business processes. This leads to severe homogenization competition, and the potential for long-term collaboration remains untapped. Furthermore, existing collaborations are mostly short-term, limited by technological compatibility, market capacity, or industrial chain support, preventing expansion to more complementary nodes or new cooperation scenarios. This hinders the formation of a sustainable ecosystem and fails to support long-term value enhancement of the industrial chain.

[0003] In the industrial chain network, the above problem can be abstracted into a graph structure. Nodes in the graph correspond to market players in each link of the industrial chain, and edges correspond to the actual relationships between these market players. Based on this, a set of nodes with sparse internal connections can be selected from the graph to achieve resource complementarity, technological integration, and market extension, avoiding homogeneous competition caused by collaboration between adjacent links. Simultaneously, without destroying the collaborative value, new nodes can be continuously absorbed, cooperation scenarios extended, and the industrial coverage expanded, forming a sustainable collaborative ecosystem. Existing technologies mainly employ rule-based methods, heuristic search methods, or traditional graph mining methods for processing. For example, by setting fixed association thresholds or constraints, a set of nodes that meet the conditions is selected from the industrial chain; this problem is modeled as a subgraph selection problem on a graph, and a set of nodes with sparse internal connections is found through traversal or greedy strategies.

[0004] In summary, existing technologies still have the following shortcomings: they only focus on the current subgraph structure and lack the ability to model the impact of future expansion; they lack effective expression of global structural information and are difficult to adapt to large-scale systems; and they lack a serialized, decision-oriented subgraph construction mechanism. There is an urgent need for a sparse relational subgraph construction method that can comprehensively consider the internal relationships of the current subgraph and the stability of future expansion in an industry chain collaborative network, and possesses good scalability. Summary of the Invention

[0005] This invention addresses the limitations of existing industry business process-driven supply chain collaboration models, which suffer from limited scope, strong structural constraints, and difficulty in continuous expansion. It proposes a method and system for mining high-value collaborative clusters within industry business processes. This method analyzes supply chain networks driven by industry business processes, aiming to identify a set of nodes with high collaboration potential from complex industry relationships, serving as the basic unit for industry collaboration and joint innovation. By maintaining the current collaborative value while considering the impact of introducing new relationships during subsequent expansion, this invention effectively prevents the collaborative structure from rapidly solidifying or failing during evolution, thereby enhancing the overall flexibility, stability, and long-term value creation capability of supply chain collaborative clusters.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] Firstly, a method for identifying high-value collaborative clusters in industrial business processes includes the following steps:

[0008] S1. Obtain and preprocess the original graph network of the industry chain; store the shortest path distance between any pair of nodes in the original graph network as a distance matrix; construct a weighted complete graph based on the distance matrix;

[0009] S2. Construct a graph neural network encoder based on a multi-hop distance decay mechanism. Perform multi-layer feature propagation and aggregation on the original graph network according to the weighted complete graph to obtain node embedding representations that integrate multi-hop global topology information and freeze them to form a global node embedding table.

[0010] S3. Construct a collaborative cluster using an autoregressive reinforcement learning model; with the currently selected node sequence as the state and the unselected nodes as candidate actions, the agent performs multi-round sequence decision-making based on the state transformed by the embedding layer and the global node embedding table, and adds the selected node to the current subgraph in each round.

[0011] S4. Based on the changes in internal association conflicts of the current node subgraph before and after the node is added, as well as the stability of the subgraph's future expansion, calculate the reward value and train an autoregressive reinforcement learning model.

[0012] S5. Using the trained autoregressive reinforcement learning model, the final collaborative cluster is gradually constructed through greedy selection; and the association score TS is used to quantitatively evaluate the final collaborative cluster.

[0013] Preferably, S2 includes:

[0014] S21, in the During the layer encoding process, based on the distance matrix, the neighboring nodes of each node are divided into neighborhood sets with different hop counts according to the shortest path distance; the node features in the neighborhood sets with different hop counts are independently aggregated to obtain the hop count feature representation corresponding to each hop count;

[0015] S22. Using a decay function that varies with the number of hops, distance decay gating is applied to the hop count feature representation to obtain decay features for different hop counts; the decay features for different hop counts are concatenated in hop count order, and a linear transformation is used to obtain the node at the hop count position. The output representation of the layer; after multiple layers of encoding, the final embedded representation of each node is obtained and frozen to form a global node embedding table.

[0016] Preferably, S3 includes:

[0017] S31. Map the currently selected node to a node embedding sequence according to the global node embedding table;

[0018] S32. Use the attention mechanism to calculate the correlation between the currently selected node and the global node;

[0019] S33. Based on the correlation between the currently selected nodes and the global nodes, generate the action value of candidate nodes; select the candidate node with the highest action value and add it to the current node subgraph;

[0020] As a preferred embodiment, in S31, the initial state is an empty node sequence, and a virtual starting node is introduced as the starting point of the sequence; the currently selected nodes are the nodes generated by each round of decision-making; the node sequence is formed according to the generation order, and mapped to a node embedding sequence through a global node embedding table.

[0021] As a preferred embodiment, in S32, the autoregressive cross-graph action selector obtains the query vector matrix by linearly mapping the node embedding sequence of the current state; obtains the key vector matrix and value vector matrix by linearly mapping the global node embedding table; calculates the correlation between the currently selected node and the global node through multi-head attention; and integrates the correlation into the attention output matrix through the feedforward layer.

[0022] Preferably, S33 includes:

[0023] Take the representation vector H of the last position in the attention output matrix last As a global representation of the current state; for any unselected candidate node v i By calculating the candidate node embeddings and the global representation H of the current state last The inner product between them yields the corresponding action value:

[0024]

[0025] Among them, S t This is the current state. Embed candidate nodes;

[0026] Select the node with the highest action value and add it to the current node's subgraph a. t At the same time, the next state S is obtained. t+1 .

[0027] Preferably, S4 includes:

[0028] S41. Calculate the instant reward based on the change in the number of node pairs in the subgraph before and after adding a node, where the shortest distance is less than or equal to the degree of association.

[0029] S42. A Q-learning-based reinforcement learning algorithm is adopted to store the experience samples composed of each complete interaction in the experience replay pool; the parameters of the autoregressive reinforcement learning model are iteratively updated by sampling a batch of experience samples and minimizing the temporal difference error.

[0030] S43. Use the padding operation to set the length of each node sequence to a fixed value and generate an aligned padding sequence; generate a mask matrix based on the padding sequence; calculate attention using the padding sequence and the corresponding mask matrix.

[0031] As a preferred option, the instant reward calculation in S41 is as follows:

[0032]

[0033] in, The purpose is to maintain the hop count relationship of subgraph nodes in the original network topology G. k-lines represent the number of node pairs in the subgraph that satisfy the shortest path distance being less than or equal to the association degree k. When the subgraph is constructed and the number of selected nodes reaches the preset scale p, the termination reward is further calculated.

[0034] The current subgraph is combined with any unselected node, and the number of short-distance connections introduced by adding that node is calculated. The expected value of all candidate nodes is then calculated to obtain the subgraph's scalability metric, MKL. A termination reward or penalty is given based on the size of the MKL. g is an adaptive positive reward, obtained by rounding up the optimal MKL (F(p)) in the model's interaction history. When the model gets stuck in a local optimum, the reward becomes a penalty, and the scaling factor is applied. This penalty will be amplified; when the subgraph construction does not reach the preset size, the negative of the k-lines increment will be given as a penalty.

[0035] As a preferred embodiment, S5 includes: fixing the model parameters after training; inputting the original graph network into the autoregressive reinforcement learning model, selecting the candidate node with the largest action value function in each round of decision-making and adding it to the current subset; and outputting the final collaborative cluster TSG when the number of selected nodes reaches the preset target size.

[0036] Use correlation scores (TS) for evaluation:

[0037]

[0038] Wherein, MKL(TSG) represents the mathematical expectation of the k-lines formed by adding any remaining node in the original graph network to the final cooperative cluster TSG to obtain a new subgraph. This represents the theoretical upper limit of MKL(TSG); the larger the TS value, the fewer global structural associations there are in the final collaborative cluster TSG.

[0039] Secondly, a high-value collaborative cluster mining system for industrial business processes includes:

[0040] The graph processing and distance calculation module is used to acquire and preprocess the original graph network of the industry chain; store the shortest path distance between any pair of nodes in the original graph network as a distance matrix; and construct a weighted complete graph based on the distance matrix.

[0041] The multi-hop distance decay coding module is used to construct a graph neural network encoder based on the multi-hop distance decay mechanism. It performs multi-layer feature propagation and aggregation on the original graph network according to the weighted complete graph to obtain node embedding representations that fuse multi-hop global topology information and freeze them to form a global node embedding table.

[0042] The autoregressive reinforcement learning decision module is used to build a collaborative cluster using an autoregressive reinforcement learning model. The agent takes the currently selected node sequence as the state and the unselected nodes as candidate actions. Based on the state transformed by the embedding layer and the global node embedding table, the agent uses an autoregressive cross-graph action selector to make multi-round sequence decisions and adds the selected node to the current subgraph in each round.

[0043] The reward calculation and model training module is used to calculate the reward value and train the autoregressive reinforcement learning model based on the changes in the internal association conflicts of the current node subgraph before and after the node is added, as well as the stability of the subgraph's future expansion.

[0044] The collaborative cluster reasoning and evaluation module is used to construct the final collaborative cluster step by step through greedy selection using a pre-trained autoregressive reinforcement learning model; and uses the association score TS to quantitatively evaluate the final collaborative cluster.

[0045] The high-value collaborative cluster mining system for industry business processes is used to implement the high-value collaborative cluster mining method and its steps for industry business processes described in the first aspect.

[0046] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0047] 1. Unlike traditional technologies that focus only on the current structure and use static subgraph construction, this invention adopts a technical solution that models the construction of collaborative clusters as an autoregressive reinforcement learning sequence decision-making process with sparse rewards. In node selection, the mathematical expectation of the internal association conflict of the current subgraph and the association conflict of future expansion are evaluated simultaneously. A two-layer reward mechanism of immediate punishment and termination reward is designed to guide the model decision-making, so that the discovered collaborative clusters can effectively avoid the problem of rapid structural solidification, maintain high stability and robustness in the dynamic evolution of the industry chain, and significantly improve the long-term value creation capability of collaborative clusters.

[0048] 2. Unlike traditional technologies that only utilize local structural information and lack global topology representation capabilities, this invention adopts a graph neural network encoding solution that integrates global modeling of shortest path distance with a multi-hop distance decay mechanism. Through a unique encoding logic of hop feature aggregation and Sigmoid gating decay, it expands the feature receptive field while suppressing noise from distant nodes, integrates multi-hop global topology information, and generates node embeddings, enabling these embeddings to efficiently represent large-scale complex industrial chain networks.

[0049] 3. Unlike traditional node selection techniques that use manual rule design and heuristic search, this invention adopts a cross-graph attention-driven sequential adaptive node selection technique. It maps the selected node sequence to an embedding sequence and constructs a query matrix. It then performs multi-head attention interaction with the global node embedding. At the same time, it designs a sequence alignment scheme for padding and masking to improve training efficiency, enabling the model to adaptively learn the optimal subgraph construction strategy under complex conflict constraints. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0051] Figure 1 This is a schematic diagram of the overall principle of the method in Embodiment 1 of the present invention;

[0052] Figure 2 This is a schematic diagram illustrating the principle of the graph neural network encoding method based on multi-hop distance decay in Embodiment 1 of the present invention;

[0053] Figure 3 This is a flowchart of the method steps in Embodiment 1 of the present invention. Detailed Implementation

[0054] To make the technical means, inventive features, objectives, and effects of the invention readily understandable, the invention is further described below with reference to specific illustrations. However, the invention is not limited to the embodiments described below.

[0055] It should be noted that the structures, proportions, sizes, etc., illustrated in the accompanying drawings of this specification are only used to complement the content disclosed in the specification for those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0056] In the technical solution of this invention, the collaborative relationships within the industrial chain are first modeled in a network manner. Market entities within the industrial chain are abstracted as nodes in a graph, and the supply and demand cooperation, technological linkages, and other relationships between nodes are abstracted as edges in the graph. Based on this, the structural relationships between nodes are characterized from a global perspective, constructing node representations that reflect multi-hop association characteristics to avoid relying solely on local associations for collaborative decision-making. Furthermore, the construction process of the collaborative cluster is viewed as a progressively evolving decision-making process. During node selection, not only are the current collaborative relationships within the cluster evaluated, but the potential impact of a node's addition on subsequent collaborative expansion is also assessed simultaneously, thereby guiding the formation of a high-value collaborative cluster that balances current collaborative value with future expansion.

[0057] Example 1:

[0058] like Figure 1 , Figure 3 The method for mining high-value collaborative clusters in industry business processes, as shown, includes the following steps:

[0059] S1. Obtain and preprocess the original graph network of the industry chain; construct a weighted complete graph based on the shortest path distance between any pair of nodes in the original graph network;

[0060] The graph structure is preprocessed by first obtaining the correlation degree k, subset size p, and undirected, unweighted connected graph G=(V,E). To efficiently obtain the hop count relationship between any pair of nodes during subsequent subgraph construction, this example performs unified preprocessing on the graph structure in the initial stage.

[0061] Specifically, for any two nodes u and v in graph G, calculate their shortest path length in graph G, denoted as Dis(u,v), and store the shortest path distance between node pairs in the distance matrix D. Based on this, construct a weighted complete graph G', where the set of nodes in G' is consistent with G, and there exists an edge between any two nodes, the weight of which is determined by the corresponding shortest path distance Dis(u,v). This method transforms the path lookup in the original graph into a constant-time table lookup, significantly improving the computational efficiency of subsequent subgraph construction. The shortest path distance represents the minimum number of hops between nodes; when selecting subsequent nodes, candidate nodes with a hop count less than or equal to a preset association level are considered as neighbors.

[0062] S2. Construct a graph neural network encoder based on a multi-hop distance decay mechanism. Perform multi-layer feature propagation and aggregation on the original graph network according to the weighted complete graph to obtain node embedding representations that integrate multi-hop global topology information and freeze them to form a global node embedding table.

[0063] To construct a global node embedding, the distance matrix D of all nodes in the undirected and unweighted connected graph G and the node pairs in S1 is input into the graph neural network to generate a global embedding table.

[0064] like Figure 2 As shown, a graph neural network (KA-GNN) based on a multi-hop distance decay mechanism is used to encode the undirected, unweighted connected graph G to obtain node embeddings that incorporate multi-hop global topological information. The initial features of each node in the graph are obtained through a normal random distribution. To ensure that the initial state of subsequent autoregressive reinforcement learning is not empty, a virtual node is encoded as the starting token during graph encoding, participating in the encoding process along with all other nodes in the graph. This node encoding serves as the starting token for subsequent autoregressive action prediction and is not included in reward calculations or sequence length statistics.

[0065] The encoding process of an L-layer graph neural network is as follows:

[0066] S21. Feature aggregation based on shortest path distance (SPD), in the... During the layer encoding process, for any node v, firstly, based on the distance matrix D in step 1, the neighboring nodes u of node v are divided into neighborhood sets with different hop counts:

[0067]

[0068] Where m represents different hop counts and k represents the degree of association;

[0069] Subsequently, the node features in the neighborhood sets of different hop counts are independently aggregated to obtain the corresponding hop count feature representations:

[0070]

[0071] Where Aggregate represents mean aggregation. This represents the node characteristics of the previous layer;

[0072] S22. Distance-based gating mechanism: To suppress the noise introduced by distant nodes during multi-layer propagation, a hop-dependent attenuation function, Sigmoid-gated attenuation (SGA), is introduced during feature fusion.

[0073]

[0074] Where Sigmoid is a mathematical function, s is the decay coefficient, b is the adaptive decay window, L is the number of layers in the graph neural network, and c is the offset used to control the position of the decay interval.

[0075] Subsequently, the decay characteristics of different hop counts are concatenated in hop count order, and the result of a linear transformation is obtained for node v at the hop count. The output of the layer is represented as follows:

[0076]

[0077] Here, Concat represents the concatenation operation. Indicates the first The L-layer is a linear layer. The node features of layer 0 are initialized using a normal random distribution. After the above encoding process in layer L, the final embedding representation of each node is obtained and frozen, forming the global node embedding table Z. G .

[0078]

[0079] In order to achieve a balance between expanding the receptive field and maintaining a high signal-to-noise ratio during node feature propagation, multi-hop feature aggregation is used to enhance the feature information of nodes, and distance-aware feature attenuation is used to reduce noise from distant nodes.

[0080] S3. Construct a collaborative cluster using an autoregressive reinforcement learning model; with the currently selected node sequence as the state and the unselected nodes as candidate actions, the agent performs multi-round sequence decision-making based on the state transformed by the embedding layer and the global node embedding table, and adds the selected node to the current subgraph in each round.

[0081] The autoregressive construction of collaborative clusters is modeled as an autoregressive reinforcement learning process, see [link to documentation]. Figure 1In reinforcement learning modeling, the already selected sequence of nodes is used as the state, and all remaining nodes represent feasible actions. The agent uses an autoregressive cross-graph action selector to decide on an action based on the state transformed by the embedding layer, and the environment calculates the corresponding reward feedback model based on that action.

[0082] S31. Map the currently selected node to a node embedding sequence according to the global node embedding table;

[0083] The initial state is an empty sequence of nodes. A virtual starting node is introduced as the starting point of the sequence to unify the state representation format. This virtual node does not participate in the sequential point selection process and does not introduce new associations. Each subsequent point selection is based on the previous point selection. In the t-th round of decision-making, the currently selected nodes (i.e., the nodes selected in rounds 0 to t-1) are arranged into a node sequence S according to their selection order. t =(v1,v2,…,v t And embed table Z through global nodes in S2. G Mapped to a sequence of node embeddings:

[0084] S32. Use the attention mechanism to calculate the correlation between the currently selected node and the global node;

[0085] During the action selection phase, a cross-graph attention mechanism is employed to model the interaction between the current state and global nodes. Specifically, in the cross-graph action selector, the node of the current state is embedded into sequence Z. S The query vector matrix Q is obtained through linear mapping. S Embed global nodes into table Z G The bond vector matrix K is obtained through linear mapping. G Sum value vector matrix V G Multi-head attention is used to calculate the correlation between the currently selected node and the global nodes:

[0086]

[0087]

[0088] Among them, W O represents a linear transformation, and d represents the dimension of the embedded vector;

[0089] Subsequently, the attention output matrix, denoted as H, is obtained by further integrating the correlation relationships through a feedforward layer and is calculated as follows:

[0090]

[0091] Where Linear represents a linear transformation, and ReLU is a mathematical function.

[0092] S33. Based on the correlation between the currently selected nodes and the global nodes, generate the action value of candidate nodes; select the candidate node with the highest action value and add it to the current node subgraph;

[0093] Take the representation vector H of the last position of H last This serves as a global representation of the current state. For any unselected candidate node v... i By calculating the candidate node embedding and state representation H last The inner product between them yields the corresponding action value:

[0094]

[0095] Then select the node with the highest action value and add it to the current node's subgraph, denoted as a. t At the same time, the next state S is obtained. t+1 .

[0096] S4. Based on the changes in internal association conflicts of the current node subgraph before and after the node is added, as well as the stability of the subgraph's future expansion, calculate the reward value and update the reinforcement learning model.

[0097] Reward calculation and model training: Reward calculation and model training are performed for each interaction. The process terminates after a specified number of training rounds and updates the reinforcement learning model parameters based on the reward value.

[0098] S41. During the subgraph construction process, when a new node is selected to be added to the current subgraph, the immediate reward is first calculated based on the change in the number of node pairs in the subgraph before and after the addition of the node that satisfy Dis(u,v)≤k (shortest distance is less than or equal to the degree of association k):

[0099]

[0100] Here, F(t) serves to maintain the hop count relationship between subgraph nodes in the original network topology G, where any subgraph S... t Both can be represented as induced subgraphs of a weighted complete graph G'. k-lines represent the number of node pairs in the subgraph whose shortest path distance is less than or equal to k. When the subgraph is constructed and the number of selected nodes reaches a preset size p, the termination reward is further calculated.

[0101] Specifically, the current subgraph is combined with any unselected node, the number of short-distance connections introduced by adding that node (i.e., the new k-lines) is calculated, and the expected value of all candidate nodes is obtained to obtain the subgraph's scalability metric, MKL. A termination reward or penalty is given based on the size of the MKL to guide the model to choose node combinations with better future scalability stability. g is an adaptive positive reward, obtained by rounding up the optimal MKL(F(p)) in the model's interaction history. When the model gets stuck in a local optimum, the reward becomes a penalty, a scaling factor. This penalty will be amplified. When the subgraph construction does not reach the preset size, the negative of the k-lines increment will be given as a penalty.

[0102] S42. During model training, a reinforcement learning algorithm based on Q-learning is used to process each complete interaction (state S). t Action a t Rewards r t Next state S t+1 The empirical samples, consisting of a set of samples, are stored in the empirical replay pool. The model parameters are iteratively updated by sampling a batch of samples and minimizing the temporal difference error.

[0103]

[0104]

[0105] Here, E represents the mean operation, and γ is the discount factor, which measures the importance of future rewards relative to immediate rewards.

[0106] S43. Use the padding operation to set the length of each node sequence to a fixed value and generate an aligned padding sequence; generate a mask matrix based on the padding sequence; calculate attention using the padding sequence and the corresponding mask matrix.

[0107] During model training, the node sequences stored in the experience replay pool are of misaligned lengths, and the feature sequences obtained after embedding transformation also have different lengths. This misalignment prevents the direct application of efficient batch matrix operations, resulting in very high training costs.

[0108] To align these lengths, a padding operation P is first used to set the length of each node sequence to the maximum value p and fill the remaining parts with virtual nodes at index number -1. When performing the embedding transformation, a 0-vector mapping to the -1 index is used.

[0109] Subsequently, a corresponding mask M is generated to prevent the attention module from seeing future tokens during prediction, i.e., to mask the influence of the 0 vector. Therefore, the S originally stored in the experience replay pool... t and St+1 Convert to P(S) t ), M (S t ) and P(S t+1 ), M (S t+1 The formula for calculating the attention head after introducing the mask is as follows:

[0110]

[0111] Through the above operations, all sequences are normalized to a uniform dimension, and attention computation can be applied to the entire batch. The overhead of processing a large matrix at a time is far lower than the loop overhead of processing small matrices multiple times, greatly improving training efficiency.

[0112] S5. Using the trained reinforcement learning model, the final collaborative cluster is gradually constructed through greedy selection; and the association score TS is used to quantitatively evaluate the final collaborative cluster.

[0113] The node subset output stage begins after the reinforcement learning model has been trained or after a specified number of interaction rounds, entering the model testing and inference phase. In this phase, model parameters are no longer updated; the trained model is used to perform collaborative cluster inference on the target network. The trained model parameters are fixed, and the graph network G is used as input. In each round of decision-making, the candidate node with the highest action value function is directly selected and added to the current subset. As the node selection process progresses, when the number of selected nodes reaches a preset target size p, the node selection process terminates, and the constructed node subset, denoted as TSG, is output. When the number of selected nodes reaches a preset size, the final constructed collaborative cluster is output.

[0114] For the node subset TSG, the association score TS is used for evaluation: ,

[0115] Here, MKL(TSG) represents the mathematical expectation of the k-lines formed by adding any remaining node in G to the TSG to obtain a new subgraph. This represents the theoretical upper limit of MKL(TSG). The larger the TS value, the less global structural correlation there is in the TSG.

[0116] The results will be presented below. This invention uses ten open-source graph datasets, the data distribution is shown in Table 1, and the important parameter settings are shown in Table 2.

[0117] Table 1: Data Distribution of the Dataset

[0118]

[0119] Table 2: Important Parameter Settings for Dataset

[0120]

[0121] Table 3: Comparison of TS scores for selected methods

[0122]

[0123] The implementation results are shown in Table 3. Table 3 presents a comparison of the results of different methods on ten real-world datasets. Compared with traditional algorithms based on static structure pruning (TERA, KTG), the present invention achieves an average improvement of over 19%. Compared with graph representation learning-based models (GNNM, R_TGF), the present invention achieves an average improvement of over 20%. Compared with mainstream graph neural network encoders (GCN, GraphSAGE, k-hop GNN), the present invention achieves an average improvement of over 6%. Furthermore, the data shows that the present invention maintains stable performance on both sparse and dense graphs, demonstrating strong generalization ability and scalability.

[0124] Example 2:

[0125] A high-value collaborative cluster mining system for industry business processes includes:

[0126] The graph processing and distance calculation module is used to acquire and preprocess the original graph network of the industry chain; store the shortest path distance between any pair of nodes in the original graph network as a distance matrix; and construct a weighted complete graph based on the distance matrix.

[0127] The multi-hop distance decay coding module is used to construct a graph neural network encoder based on the multi-hop distance decay mechanism. It performs multi-layer feature propagation and aggregation on the original graph network according to the weighted complete graph to obtain node embedding representations that fuse multi-hop global topology information and freeze them to form a global node embedding table.

[0128] The autoregressive reinforcement learning decision module is used to build a collaborative cluster using an autoregressive reinforcement learning model. The agent takes the currently selected node sequence as the state and the unselected nodes as candidate actions. Based on the state transformed by the embedding layer and the global node embedding table, the agent uses an autoregressive cross-graph action selector to make multi-round sequence decisions and adds the selected node to the current subgraph in each round.

[0129] The reward calculation and model training module is used to calculate the reward value and train the autoregressive reinforcement learning model based on the changes in the internal association conflicts of the current node subgraph before and after the node is added, as well as the stability of the subgraph's future expansion.

[0130] The collaborative cluster reasoning and evaluation module is used to construct the final collaborative cluster step by step through greedy selection using a trained autoregressive reinforcement learning model; and to quantitatively evaluate the final collaborative cluster using the association score TS.

[0131] The embodiments and implementation process of the present invention have been described in detail above with reference to the accompanying drawings and tables, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments, including components, without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A method for mining high-value collaborative clusters in industrial business processes, characterized in that, Includes the following steps: S1. Obtain and preprocess the original graph network of the industrial chain; firstly, perform network modeling of the collaborative relationship of the industrial chain, abstract the market entities in the industrial chain as nodes in the graph, and abstract the supply and demand cooperation relationship and technology association relationship between nodes as edges in the graph; store the shortest path distance between any pair of nodes in the original graph network as a distance matrix; construct a weighted complete graph based on the distance matrix; S2. Construct a graph neural network encoder based on a multi-hop distance decay mechanism. Perform multi-layer feature propagation and aggregation on the original graph network according to the weighted complete graph to obtain node embedding representations that fuse multi-hop global topological information and freeze them, forming a global node embedding table; including: S21, in the During the layer encoding process, based on the distance matrix, the neighboring nodes of each node are divided into neighborhood sets with different hop counts according to the shortest path distance; the node features in the neighborhood sets with different hop counts are independently aggregated to obtain the hop count feature representation corresponding to each hop count; S22. Using a decay function that varies with the number of hops, distance decay gating is applied to the hop count feature representation to obtain decay features for different hop counts; the decay features for different hop counts are concatenated in hop count order, and a linear transformation is used to obtain the node at the hop count position. The output representation of the layer; after multi-layer encoding, the final embedding representation of each node is obtained and frozen to form a global node embedding table; S3, construct a collaborative cluster using an autoregressive reinforcement learning model; with the currently selected node sequence as the state and the unselected nodes as candidate actions, the agent performs multi-round sequence decision-making based on the state transformed by the embedding layer and the global node embedding table, adding the selected node to the current subgraph in each round; including: S31. Map the currently selected node to a node embedding sequence according to the global node embedding table; S32. Use the attention mechanism to calculate the correlation between the currently selected node and the global node; S33. Based on the correlation between the currently selected nodes and the global nodes, generate the action value of candidate nodes; select the candidate node with the highest action value and add it to the current node subgraph; S4. Based on the changes in internal association conflicts of the current node subgraph before and after the node is added, as well as the stability of the subgraph's future expansion, calculate the reward value and train an autoregressive reinforcement learning model. S5. Using the trained autoregressive reinforcement learning model, the final collaborative cluster is gradually constructed through greedy selection; and the association score TS is used to quantitatively evaluate the final collaborative cluster.

2. The method for mining high-value collaborative clusters oriented towards industrial business processes according to claim 1, characterized in that, In S31, the initial state is an empty node sequence, and a virtual starting node is introduced as the starting point of the sequence; the currently selected nodes are the nodes generated by each round of decision-making; the node sequence is formed according to the generation order, and mapped to a node embedding sequence through a global node embedding table.

3. The method for mining high-value collaborative clusters oriented towards industrial business processes according to claim 1, characterized in that, In S32, the autoregressive cross-graph action selector obtains the query vector matrix by linearly mapping the node embedding sequence of the current state. The global node embedding table is linearly mapped to obtain the key vector matrix and the value vector matrix, respectively. The correlation between the currently selected node and the global node is calculated using multi-head attention; The correlation relationships are integrated into an attention output matrix through a feedforward layer.

4. The method for mining high-value collaborative clusters oriented towards industrial business processes according to claim 1, characterized in that, S33 includes: Take the representation vector H of the last position in the attention output matrix last As a global representation of the current state; for any unselected candidate node v i By calculating the candidate node embeddings and the global representation H of the current state last The inner product between them yields the corresponding action value: in, This is the current state. Embed candidate nodes; Select the node with the highest action value and add it to the current node's subgraph. At the same time, the next state is obtained. .

5. The method for mining high-value collaborative clusters oriented towards industrial business processes according to claim 1, characterized in that, S4 include: S41. Calculate the instant reward based on the change in the number of node pairs in the subgraph before and after adding a node, where the shortest distance is less than or equal to the degree of association. S42. A Q-learning-based reinforcement learning algorithm is adopted to store the experience samples composed of each complete interaction in the experience replay pool; the parameters of the autoregressive reinforcement learning model are iteratively updated by sampling a batch of experience samples and minimizing the temporal difference error. S43. Use the padding operation to set the length of each node sequence to a fixed value and generate an aligned padding sequence; generate a mask matrix based on the padding sequence; calculate attention using the padding sequence and the corresponding mask matrix.

6. The method for mining high-value collaborative clusters oriented towards industrial business processes according to claim 5, characterized in that, The instant reward calculation in S41 is as follows: in, The purpose is to maintain the hop count relationship of subgraph nodes in the original network topology G. k-lines represent the number of node pairs in the subgraph that satisfy the shortest path distance being less than or equal to the association degree k. When the subgraph is constructed and the number of selected nodes reaches the preset scale p, the termination reward is further calculated. The current subgraph is combined with any unselected node, and the number of short-distance connections introduced by adding that node is calculated. The expected value of all candidate nodes is then calculated to obtain the subgraph's scalability metric, MKL. A termination reward or penalty is given based on the size of the MKL. g is an adaptive positive reward, obtained by rounding up the optimal MKL (F(p)) in the model's interaction history. When the model gets stuck in a local optimum, the reward becomes a penalty, and the scaling factor is applied. This penalty will be amplified; when the subgraph construction does not reach the preset size, the negative of the k-lines increment will be given as a penalty.

7. The method for mining high-value collaborative clusters oriented towards industrial business processes according to claim 1, characterized in that, S5 includes: fixing the model parameters after training; inputting the original graph network into the autoregressive reinforcement learning model, selecting the candidate node with the largest action value function in each round of decision-making and adding it to the current subset; when the number of selected nodes reaches the preset target size, outputting the final collaborative cluster TSG; Use correlation scores (TS) for evaluation: Wherein, MKL(TSG) represents the mathematical expectation of the k-lines formed by adding any remaining node in the original graph network to the final cooperative cluster TSG to obtain a new subgraph. This represents the theoretical upper limit of MKL(TSG); the larger the TS value, the fewer global structural associations there are in the final collaborative cluster TSG.

8. A high-value collaborative cluster mining system for industrial business processes, characterized in that, include: The graph processing and distance calculation module is used to acquire and preprocess the original graph network of the industrial chain. First, the collaborative relationship of the industrial chain is modeled in a network, and the market entities in the industrial chain are abstracted as nodes in the graph. The supply and demand cooperation relationship and the technical relationship between the nodes are abstracted as edges in the graph. The shortest path distance between any pair of nodes in the original graph network is stored as a distance matrix. A weighted complete graph is constructed based on the distance matrix. The multi-hop distance decay encoding module is used to construct a graph neural network encoder based on a multi-hop distance decay mechanism. It performs multi-layer feature propagation and aggregation on the original graph network according to the weighted complete graph, obtaining and freezing node embedding representations that fuse multi-hop global topological information, forming a global node embedding table; including: In the During the layer encoding process, based on the distance matrix, the neighboring nodes of each node are divided into neighborhood sets with different hop counts according to the shortest path distance; the node features in the neighborhood sets with different hop counts are independently aggregated to obtain the hop count feature representation corresponding to each hop count; Using a decay function that varies with the number of hops, distance decay gating is applied to the hop count feature representation to obtain decay features for different hop counts. These decay features are then concatenated in hop count order, and a linear transformation is applied to obtain the node at the [hop count] position. The output representation of the layer; after multiple layers of encoding, the final embedded representation of each node is obtained and frozen to form a global node embedding table; The autoregressive reinforcement learning decision module is used to build collaborative clusters using an autoregressive reinforcement learning model. It uses the currently selected node sequence as the state and unselected nodes as candidate actions. The agent performs multi-round sequence decisions based on the state transformed by the embedding layer and the global node embedding table, using an autoregressive cross-graph action selector. In each round, the selected node is added to the current subgraph. This includes: Map the currently selected node to a node embedding sequence based on the global node embedding table; An attention mechanism is used to calculate the correlation between the currently selected node and the global node; Based on the correlation between the currently selected nodes and the global nodes, generate the action value of candidate nodes; select the candidate node with the highest action value and add it to the current node subgraph; The reward calculation and model training module is used to calculate the reward value and train the autoregressive reinforcement learning model based on the changes in the internal association conflicts of the current node subgraph before and after the node is added, as well as the stability of the subgraph's future expansion. The collaborative cluster reasoning and evaluation module is used to construct the final collaborative cluster step by step through greedy selection using a pre-trained autoregressive reinforcement learning model; and uses the association score TS to quantitatively evaluate the final collaborative cluster. The high-value collaborative cluster mining system for industry business processes is used to implement the high-value collaborative cluster mining method and steps for industry business processes as described in claim 1.

Images