A Multi-View Knowledge Graph Driven Industrial Chain Collaborative Recommendation Method

Abstract

The present invention relates to a multi-view knowledge graph-driven industrial chain collaborative recommendation method, including the following steps: constructing knowledge graph views of multiple business dimensions through multi-source heterogeneous data, performing feature decomposition and adaptive feature regulation; proposing a multi-head knowledge propagation mechanism integrating business context, and combining multi-view representations for hierarchical convolutional aggregation and dynamic path selection; constructing a dynamic representation of business entities through wavelet transform and multi-scale time series decomposition, combining an adaptive scale selection mechanism and an attention mechanism driven by industry characteristics; using a gated network to fuse static and dynamic features, constructing a multi-party matching scoring mechanism based on comprehensive representation, and supporting multi-party business collaboration and recommendation in the industrial chain. By integrating multi-view knowledge graph construction, dynamic business entity modeling, and contrast learning techniques, the present invention reveals the relevance between enterprise businesses, generates business recommendations for relevant enterprises, and lays a foundation for business resource allocation in the industrial chain.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and particularly relates to a multi-view knowledge graph-driven industrial chain collaborative recommendation method. Background Art

[0002] In the era of increasingly complex and dynamic industrial chains, enterprises need to collaborate to complete complex business requirements put forward by customers. However, due to the diverse business types of small and medium-sized enterprises, they face problems such as poor information circulation, low transaction efficiency, and insufficient manufacturing capacity matching when looking for collaborative partners, which hinder the efficient business matching and collaboration among industrial chain enterprises. In order to improve the overall operation efficiency of the regional industrial chain, constructing an accurate and dynamic enterprise business recommendation method has become the key to the optimal allocation of the industrial chain. By analyzing the business associations and resource characteristics among enterprises, the recommendation system can provide accurate business matching suggestions for enterprises, thereby improving the overall operation efficiency of the industrial chain.

[0003] In recent years, knowledge graphs have been widely applied in the field of recommendation systems due to their unique advantages in alleviating data sparsity and improving recommendation accuracy. By constructing a triple heterogeneous semantic structure of entities and their relationships, knowledge graphs can effectively mine the potential associations between enterprises and business entities.

[0004] However, existing methods tend to uniformly model the overall structure of enterprise data. Although the representation of relationships between entities is initially realized, it is vulnerable to interference from data noise, incorrect features, and weakly related features, resulting in insufficient accuracy of enterprise feature expression. There are a large number of heterogeneous nodes of multiple types and multiple relationships in the knowledge graph, and the multi-dimensional features exhibited by business entities from different perspectives are difficult to be fully captured by traditional methods, making the entity representation limited in both business perception ability and structural expression ability. In addition, the business behavior characteristics of enterprises often show dynamic changes over time, and short-term behavior trends have an important impact on current business matching decisions. Existing recommendation methods mostly focus on long-term stable static features, ignoring short-term dynamic behavior information and failing to effectively integrate static and dynamic features, resulting in the recommendation results being difficult to meet the actual needs of the rapidly changing industrial chain in terms of timeliness and adaptability. Summary of the Invention

[0005] The purpose of the present invention is to provide a multi-view knowledge graph-driven industrial chain collaborative recommendation method, aiming to solve the problems of insufficient accuracy, poor timeliness, and weak dynamic adaptability faced by existing technologies in the business recommendation of industrial chain enterprises. By integrating multi-view knowledge graph construction, dynamic business entity modeling, and contrast learning technologies, the present invention reveals the associations between enterprise businesses, generates business recommendations for relevant enterprises, and lays a foundation for the business resource allocation in the industrial chain.

[0006] To solve the above technical problems, the technical solution provided by the present invention is as follows:

[0007] A multi-view knowledge graph-driven industrial chain collaborative recommendation method includes the following steps:

[0008] S1. Construct knowledge graph views of multiple business dimensions from multi-source heterogeneous data, and perform feature decomposition and adaptive feature regulation;

[0009] S2. Propose a multi-head knowledge propagation mechanism that integrates business context, and perform hierarchical convolutional aggregation and dynamic path selection in combination with multi-view representations;

[0010] S3. Construct dynamic representations of business entities through wavelet transform and multi-scale time series decomposition, combined with an adaptive scale selection mechanism and an attention mechanism driven by industry characteristics;

[0011] S4. Use a gated network to fuse static and dynamic features, construct a multi-party matching scoring mechanism based on comprehensive representations, and support multi-party business collaboration and recommendation in the industrial chain.

[0012] Preferably, step S1 is specifically as follows:

[0013] S1.1. Based on multi-source heterogeneous data including enterprise information, business reports, and historical transaction data, construct an original knowledge graph G :

[0014] ;

[0015] Among them, represents a set of entities containing multiple business entities (business entities include enterprises, products, businesses, services, etc.); represents a set of relationships of multiple business relationships;

[0016] S1.2. Use domain expert knowledge and pre-trained models to perform multi-dimensional decomposition on the enterprise original feature matrix , extract key sub-features, where represents that this matrix belongs to a matrix space of dimension in the real number field; represents the size of the entity set , that is, the number of entities; represents the number of feature dimensions of each entity; for each view , calculate the correlation between the th feature of the th enterprise node and the business matching index , and obtain the feature retention probability through the normalization function : ;

[0017] Among them Indicates the values of all entities on the th feature dimension; " " indicates "all rows"; Indicates the th feature dimension, y is the supervision signal for business matching, ; Indicates the correlation score between the feature and the business goal; Indicates the base of the natural logarithm;

[0018] Generate a binary mask matrix through independent sampling based on the feature retention probability: ;

[0019] where Indicates that under the view , the th feature of the th enterprise node is retained, Indicates being suppressed;

[0020] By independent sampling with as the probability, selectively retain high-correlation features and suppress noise and low-correlation features, so as to obtain the entity feature matrix for different views : ;

[0021] S1.3. Introduce an adaptive edge weight calculation and sampling strategy when constructing relationships between various entities; for any two business entities in the original knowledge graph and , calculate the edge weight W ij : ;

[0022] where, conf( e i , e j ) is the edge confidence obtained from multi-source data, , U is the number of interaction times; is to measure the semantic similarity between entities by calculating the cosine similarity of the feature vectors obtained after encoding the entity descriptions using the BERT model (bidirectional transformer model); is an adjustable hyperparameter; the activation function normalizes the weight to the interval [0,1], ; x is the pre-activation value of the edge weight; indicates in ;

[0023] Sampling probability of the edge is defined as its weight value: ;

[0024] For each edge, based on Independent sampling is performed through the Bernoulli distribution to generate a subgraph; to capture structural diversity and reduce the impact of noise, the sampling process is repeated to generate multiple different subgraph views: ;

[0025] The adjacency matrix of each view is defined as:

[0026] ;

[0027] If the edge (e i , e j ) is retained in view v, the corresponding value of the adjacency matrix is 1; otherwise, the value is 0;

[0028] S1.4. Based on the entity feature matrix specific to each view and the corresponding adjacency matrix , construct a multi-layer graph convolutional network GCN, and introduce an enhanced residual connection strategy. Use the improved graph convolutional network to perform embedding representation learning on all business entities; first, define the initial node feature matrix :

[0029] ;

[0030] The propagation rule of the improved graph convolution in the th view and the th layer is: ;

[0031] Among them, , where is the identity matrix; is the initial node feature matrix; is the weight matrix of the th view and the th layer convolutional network; is the adjacency matrix after adding self-loops to view ; is the corresponding degree matrix; and are hyperparameters that control the retention ratios of the initial feature and the previous layer feature respectively; is the activation function to ensure the non-linear expression ability;

[0032] After After the layer-improved graph convolution propagation, the entity node embedding representation of the th view is obtained : ; denotes the node feature matrix of the v th layer of the graph convolutional network in the

[0033] As an option, in step S1.1, the multiple business entities include enterprises, products, services, and demands; the multiple business relationships include supply, cooperation, competition, and demand-supply; in step S1.2, the key sub-features include business demands, resource capabilities, and market positions; in step S1.3, the initial value ; in step S1.4, the initial value .

[0034] As an option, step S2 is specifically as follows:

[0035] S2.1. A multi-head knowledge propagation mechanism integrating business context is proposed for the diverse relationship features reflected among business entities from different business perspectives:

[0036] For any business entity in the view and its neighbor , the attention weight of the th attention head is defined as :

[0037] ;

[0038] where is the linear transformation matrix of the th attention head, which is used to map the entity embedding to a unified representation space; is the business relationship index between the entity and its neighbor ; the fusion function uses a multi-layer perceptron network (MLP) to fuse and model the entity pair and its business relationship, and the output semantic vector is weighted and scored by the query vector , and the normalized attention weight is constructed through softmax to reflect the semantic propagation contribution of the neighbor to the entity ;

[0039] S2.2. After the above multi-head propagation based on the attention weights, the updated representation of the entity is defined as:

[0040] ;

[0041] Among them, represents the set of neighbor nodes of the entity . is the transformation matrix in the th attention head. represents the feature representation of the neighbor entity under the th view. represents the concatenation operation.

[0042] S2.3. Entity node embedding representation based on the views generated above , through hierarchical convolutional aggregation and dynamic path selection mechanism, realizes deeper feature extraction and refinement:

[0043] For the initial node representation , the representations of each layer are obtained through multi-layer convolutional propagation , and a simplified residual structure is introduced during the final aggregation, and only the initial feature is weighted and fused with the propagation results of each level to obtain the aggregated representation of the node:

[0044] ;

[0045] Among them, is the number of propagation layers adaptively determined for the th node; is the trainable weight, which is optimized through backpropagation;

[0046] S2.4. Introduce a dynamic propagation depth control strategy. By calculating the change amount between the representations of adjacent convolutional layers of the node, combined with the adaptive threshold , where and are the statistical features of the change of the node in the previous layer, is a hyperparameter, to judge whether the current node feature has tended to be stable. When , it is considered that the feature of the th node has tended to be stable, and further propagation is stopped;

[0047] S2.5. Adopt the following fusion method:

[0048] For each business entity , the representations under each view are (that is, under each view v = 1, 2, …, V , represented by the symbol ), and the attention mechanism is used to assign dynamic weights to different perspectives:

[0049] ;

[0050] , where is the dimension of ; is the middle layer dimension is the traversal variable representing the view index, is the total number of views, represents the transpose operation of a vector or matrix, is the bias term in the attention mechanism, is the trainable query vector in the attention mechanism; obtaining the long-term static embedding representation of the business entity :

[0051] ;

[0052] S2.6. Introduce a consistency alignment mechanism based on contrastive learning. By maximizing the similarity between the representations of the same entity under different views and minimizing the distance between its representation and the representations of other entities, construct a consistency loss function , where is the cosine similarity, is the temperature parameter.

[0053] Preferably, step S3 is specifically as follows:

[0054] S3.1. In the dynamic behavior modeling of business entities, for each business entity collect its time series behavior data to form a behavior sequence , where is the behavior embedding vector of the entity at time , generated from the original data by a pre-trained model, represents the dimension size of the behavior embedding vector of the business entity at each time point ; Based on information including industry classification, enterprise scale, and geographical location, introduce industry feature embedding , represents the industry feature embedding ; is the dimension size of, used to encode the structural static information related to the business entity;

[0055] S3.2. Introduce discrete wavelet transform DWT to decompose the behavior sequence and automatically extract dynamic components of different frequencies:

[0056] Apply discrete wavelet transform to each dimension of the behavior sequence to generate a low-frequency component , reflecting the long-term trend of the entity; generate a high-frequency component , representing the decomposition level, capturing short-term fluctuations;

[0057] S3.3. Dynamically fuse features at each scale through the attention mechanism:

[0058] The set of hidden states at each scale is input into the multi-head attention network MHA, with industry characteristics embedded as the query vector to guide the model to focus on the time series components most relevant to industry characteristics; calculate the weights through normalization by the softmax function and obtain the comprehensive time series representation through weighted fusion ;

[0059] S3.4. Adopt the industry-characteristic-driven multi-head attention mechanism:

[0060] First, combine the comprehensive time series representation with the industry characteristics embedding to calculate the attention weights . Subsequently, introduce the value mapping matrix to linearly project the weighted multi-scale representation and aggregate the outputs of all attention heads at all time steps, thereby obtaining the short-term dynamic comprehensive representation of the business entity at the current moment , and the calculation method is as follows:

[0061] ;

[0062] is the number of attention heads, is the dimension of the dynamic representation;

[0063] S3.5. Adopt a contrastive learning framework that combines cross-time and cross-entity; this framework is based on the multi-scale hidden states specific to time steps to establish a contrastive mechanism:

[0064] In the time dimension, introduce a cross-time contrastive learning mechanism; regard the representations of adjacent time steps and as a positive sample pair to construct a positive guidance for behavioral continuity; at the same time, form a negative sample pair with the representation of the current time step and the competing entities to enhance the ability to distinguish dynamic behaviors between entities, that is , where represents the set of entities competing with entity ; its time loss function is:

[0065] ;

[0066] where is the temperature hyperparameter, used to adjust the sparsity of the similarity distribution;

[0067] S3.6. Further introduce a cross-entity contrastive learning strategy:

[0068] Use industry-similar entities as positive samples , where represents the set of entities belonging to the same industry as entity , to reflect the potential business synergy among entities in the same industry; at the same time, select entities with large industry differences as negative samples, so as to prompt the model to effectively widen the distribution of entities in different industries in the representation space, that is , where represents the set of entities belonging to different industries from entity ;

[0069] The loss function of cross-entity contrastive learning is defined as follows:

[0070] .

[0071] Preferably, in step S3.2, the low-frequency component is regarded as the 0th order scale, denoted as ; design an independent temporal convolutional network TCN model for each component to extract features: ;

[0072] where is the hidden representation at each scale;

[0073] In step S3.4, in the initial implementation, set the number of attention heads to 4 to balance the computational complexity and the ability to capture different temporal features; at the same time, the dimension of the dynamic representation is set to match the dimension of the static embedding , that is , to ensure the dimensional compatibility when fusing the subsequent static and dynamic representations.

[0074] Preferably, step S4 is specifically:

[0075] S4.1. Combine the long-term and stable static representation with the dynamic representation reflecting recent behaviors, to improve the model's comprehensive modeling ability for business entities; among them, is derived from the multi-view knowledge graph and convolutional propagation; then captures the dynamic behavior response of the enterprise through temporal data and the attention mechanism;

[0076] Adopt a learnable gating network to adaptively adjust the fusion ratio of static features and dynamic features:

[0077] ​ ;

[0078] Among them and are projection parameter matrices used to project static features and is used to project dynamic features ; is a bias term; is an activation function that normalizes the output value to the interval [0, 1];

[0079] Each element represents the importance of the stable business advantages of the enterprise in the corresponding dimension; the larger the value, the more the dimension depends on static features, and vice versa, it pays more attention to recent dynamics;

[0080] Subsequently, the two types of features are weighted and fused through element-wise multiplication to obtain the final comprehensive representation of the industrial chain business entity :

[0081] ;

[0082] Among them represents element-wise multiplication;

[0083] S4.2. Construct a multi-party matching scoring mechanism based on the comprehensive representation; for any two enterprises and with potential business synergy effects, calculate their respective fused comprehensive representations and , and construct a matching score :

[0084] ;

[0085] Among them, is the Sigmoid activation function used to normalize the output to the interval [0, 1]; represents the multi-layer perceptron function, and its form is as follows:

[0086] ;

[0087] Among them, represents the vector concatenation operation, , is the weight matrix, , is the bias term;

[0088] S4.3. Adopt a multi-party collaborative matching mechanism to comprehensively consider the business complementarity and overall synergy among multiple enterprises:

[0089] For the target enterprise , select a group of enterprises from the candidate enterprise set to optimize the overall business matching effect; the specific implementation method is as follows:

[0090] Based on pairwise matching scores , use the attention mechanism to quantify the interaction importance between the target enterprise and the candidate enterprises , and calculate the attention weights through exponential normalization operation: ;

[0091] Among them, represents the relative importance of enterprise to the target enterprise in business collaboration;

[0092] After obtaining the attention weights , further use these weights to perform weighted aggregation on the comprehensive representations of each enterprise in the candidate enterprise set to generate the context representation of the target enterprise: ;

[0093] Among them, reflects the overall interaction relationship between the target enterprise and the candidate enterprise set;

[0094] Combine the comprehensive representation and the context representation of the target enterprise to calculate its final business recommendation score: ;

[0095] Among them, represents the vector concatenation operation, is a learnable weight matrix, is a bias term.

[0096] Adopting the technical solution provided by the present invention, compared with the prior art, it has the following beneficial effects:

[0097] 1. By constructing multiple knowledge graph views for different business dimensions and combining domain feature decomposition, adaptive feature regulation, and edge weight sampling strategies, the present invention effectively suppresses the interference of data noise and weakly related features; compared with traditional knowledge graph construction methods, the present invention can more accurately depict the multi-dimensional attributes of enterprises, improve the expression ability of the graph, and thus significantly improve the accuracy and robustness of business matching.

[0098] 2. The present invention introduces a multi-head knowledge propagation mechanism integrating business context, captures rich business context information during the propagation between entities, and further optimizes the depth and accuracy of entity representation by combining hierarchical convolutional aggregation and dynamic path selection mechanism. Compared with the relatively single graph embedding method in the prior art, the present invention can generate entity representations more suitable for actual business scenarios, improving the structural expression ability and business perception ability of business entities.

[0099] 3. In view of the characteristics of the evolution of business entity behavior over time, the present invention proposes a dynamic business entity modeling framework, extracts and fuses the long-term trend and short-term fluctuation characteristics of business entities through wavelet transform-driven multi-scale time series decomposition and adaptive scale selection mechanism. In addition, combined with the attention mechanism driven by industry characteristics, it highlights the key dynamic behaviors related to the industry background. Compared with the existing methods, the present invention significantly improves the flexibility and accuracy of dynamic modeling, providing a decision-making basis with strong timeliness and wide adaptability for business recommendation.

[0100] 4. The present invention designs a learnable gating network to adaptively adjust the fusion ratio of static features and dynamic features, and combines a cross-time and cross-entity contrast learning framework to enhance the model's understanding ability of the evolution of business entity behavior and industry characteristics. Compared with the simple feature splicing or fixed-weight fusion methods in the prior art, the comprehensive representation generated by the present invention has higher discriminability and robustness, and can more comprehensively reflect the business status of enterprises. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] Figure 1 It is a flowchart of an industrial chain collaborative recommendation method driven by a multi-view knowledge graph;

[0102] Figure 2 It is a box plot of the performance of this model. DETAILED DESCRIPTION OF THE INVENTION

[0103] To further understand the content of the present invention, the present invention will be described in detail in combination with embodiments. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0104] Refer to Figure 1 , a multi-view knowledge graph-driven industrial chain collaborative recommendation method, includes the following steps:

[0105] S1. Construct knowledge graph views of multiple business dimensions through multi-source heterogeneous data, perform feature decomposition and adaptive feature regulation to realize entity embedding representation learning (multi-view knowledge graph construction). The specific steps are as follows:

[0106] S1.1. First, construct an original knowledge graph based on multi-source heterogeneous data such as enterprise information and business reports :

[0107] ;

[0108] Among them, the entity set contains various business entities such as enterprises, products, services, and demands; the relationship set contains various business relationships such as supply, cooperation, competition, and demand-supply.

[0109] S1.2. Use domain expert knowledge and pre-trained models to perform multi-dimensional decomposition (domain feature decomposition) on the original enterprise feature matrix , extract key sub-features such as business requirements, resource capabilities, and market positions, so as to make the feature expression of enterprise nodes clearer, where represents that this matrix belongs to a matrix space with a dimension of in the real number field; represents the size of the entity set , that is, the number of entities; represents the number of feature dimensions of each entity. For each view , calculate the correlation between the th feature of the th enterprise node and the business matching index, and obtain the feature retention probability through the normalization function : ;

[0110] Among them , represents the value of all entities on the th feature dimension; " " represents "all rows"; represents the th feature dimension, is the supervision signal for business matching, . represents the correlation score between the feature and the business goal; represents the base of the natural logarithm;

[0111] Based on this probability, generate a binary mask matrix through independent sampling: ;

[0112] Among them represents that under the view , the th feature of the th enterprise node is retained, represents being suppressed;

[0113] By using For independent sampling of probabilities, selectively retain highly correlated features and suppress noise and low-correlated features, thus obtaining an entity feature matrix for different views. : .

[0114] S1.3. To enhance the robustness of the graph structure, an adaptive edge weight calculation and sampling strategy (edge weight calculation and sampling) is introduced when constructing relationships between various entities. For the original knowledge graph any two entities and in it, calculate the edge weight for the relationship edge between them:

[0115] where conf( e i , e j ) is the edge confidence obtained from multi-source data, , U is the number of interaction times; represents that after encoding the entity description using BERT, the cosine similarity of the calculated feature vectors is used to measure the semantic similarity between entities; is an adjustable hyperparameter, with an initial value ; the activation function normalizes the weight to the interval [0,1], ;

[0116] The sampling probability of the edge is defined as its weight value: ;

[0117] For each edge, based on generate a subgraph through independent sampling according to the Bernoulli distribution. To capture structural diversity and reduce the influence of noise, repeat the sampling process to generate multiple different subgraph views: ;

[0118] The adjacency matrix of each view is defined as:

[0119] ;

[0120] If the edge (e i ,e j ) is retained in view v, then the value of the corresponding adjacency matrix is 1; otherwise, has a value of 0.

[0121] S1.4. Based on the entity feature matrix specific to each view and the corresponding adjacency matrix , the present invention constructs a multi - layer graph convolutional network (GCN), introduces an enhanced residual connection strategy, and uses the improved graph convolutional network to perform embedded representation learning on all business entities (graph convolutional embedding learning). First, define the initial node feature matrix:

[0122] ;

[0123] The -th view's -th layer's improved graph convolutional propagation rule is: ;

[0124] Among them, , where is the identity matrix; is the initial node feature matrix; is the -th view's -th layer's convolutional network weight matrix; is the -th view's adjacency matrix after adding self - loops; is the corresponding degree matrix; and are hyperparameters that control the retention ratios of the initial feature and the previous layer's feature, with the initial value ; is the activation function to ensure non - linear expression ability;

[0125] After layers of improved graph convolutional propagation, the entity node embedding representation of the -th view is obtained: .

[0126] S2. Propose a multi - head knowledge propagation mechanism that fuses business context, combines multi - view representations for hierarchical convolutional aggregation and dynamic path selection to achieve more refined entity representation and business perception (business entity modeling). The specific steps are as follows:

[0127] S2.1. A multi - head knowledge propagation mechanism that fuses business context is proposed for the diverse relationship characteristics shown by business entities from different business perspectives. This mechanism can effectively capture business context information during the entity propagation process, thereby enhancing the business perception ability and structural expression ability of entity representations;

[0128] For any business entity in view and its neighbor , the attention weight of the -th attention head is defined as:

[0129] ;

[0130] Among them, is the linear transformation matrix of the -th attention head, which is used to map the entity embedding to a unified representation space; is the business relationship index between the entity and its neighbor ; The fusion function adopts a multi-layer perceptron network (MLP) to fuse and model the entity pair and its business relationship. The output semantic vector is weighted and scored by the query vector , and the normalized attention weight is constructed through softmax to reflect the semantic propagation contribution of the neighbor to the entity ;

[0131] mainly serves the feature mapping before attention calculation, highlighting the business context and the relationship information between nodes.

[0132] S2.2. Further, after the multi-head propagation of the above attention weights, the updated representation of the entity is defined as:

[0133] ;

[0134] Among them, represents the set of neighbor nodes of the entity , is the transformation matrix in the -th attention head, represents the feature representation of the neighbor entity under the -th view, represents the concatenation operation;

[0135] mainly serves the transformation and aggregation of neighbor information, ensuring that the final updated representation can synthesize neighbor features from multiple perspectives;

[0136] This representation is obtained by concatenating the neighbor representations weighted and aggregated by attention heads respectively, comprehensively integrating the refined neighbor information obtained from multiple business perspectives, and effectively improving the multi-view expression ability and business interpretability of entity representations.

[0137] S2.3. Based on the view-specific embedding representations generated above, it aims to further optimize these initial representations to better adapt to specific business scenarios and entity characteristics. Through hierarchical convolutional aggregation and dynamic path selection mechanisms, deeper feature extraction and refinement are achieved;

[0138] For the initial representation of nodes each layer representation is obtained through multi-layer convolutional propagation and a simplified residual structure is introduced during the final aggregation, only the initial features are weighted and fused with the propagation results at each level to obtain the aggregated representation of the node:

[0139] ;

[0140] wherein, is the number of propagation layers adaptively determined by the -th node; is the trainable weight, which is optimized through backpropagation;

[0141] This design effectively avoids the problem of introducing redundant information in each layer of propagation in the traditional residual structure, simplifies the model structure while retaining multi-layer semantic information.

[0142] S2.4. To further improve the efficiency and expression accuracy of information propagation, the present invention introduces a dynamic propagation depth control strategy; by calculating the change amount between the representations of a node in adjacent convolutional layers , combined with an adaptive threshold (wherein and are the statistical features of the change of the previous layer node, is a hyperparameter), it is judged whether the current node feature has tended to be stable. When , it is considered that the feature of the -th node has tended to be stable, thereby stopping further propagation; this mechanism can effectively control the noise interference during the propagation process while retaining the business difference features between nodes.

[0143] S2.5. Considering the multi-dimensionality of enterprise business attributes, each view provides a local but unique semantic entity representation. To integrate the information of each view, the present invention proposes the following innovative fusion method (cross-view fusion);

[0144] For each business entity , the representations under each view are , and the attention mechanism is used to assign dynamic weights to different perspectives:

[0145] ;

[0146] , wherein is the dimension of , is the intermediate layer dimension; is the traversal variable representing the view index, represents the total number of views, Denotes the transpose operation of a vector or matrix, which is the bias term in the attention mechanism, and is the trainable query vector in the attention mechanism; the final comprehensive representation is:

[0147] .

[0148] S2.6. Further, to enhance the consistency between the representations of the same entity under different views and at the same time distinguish the semantic differences between different entities, the present invention introduces a consistency alignment mechanism based on contrastive learning. By maximizing the similarity between the representations of the same entity under different views and minimizing the distance between it and the representations of other entities, a consistency loss function is constructed , where is the cosine similarity, and is the temperature parameter. This loss is based on the cosine similarity and combines the temperature parameter to adjust the optimization amplitude, thereby enhancing the discriminative ability and robustness of the overall representation.

[0149] S3. Through wavelet transform and multi-scale time series decomposition, combined with an adaptive scale selection mechanism and an attention mechanism driven by industry characteristics, a dynamic representation of business entities (dynamic business entity modeling) is constructed. The specific steps are as follows:

[0150] The present invention proposes a dynamic business entity modeling framework aimed at solving the problem of the dynamic evolution of the behavioral characteristics of business entities over time. On the basis of the foregoing, long-term static embedding representations of business entities (such as enterprises, products, services) have been obtained . However, the short-term behavioral trends of business entities have an important impact on the current business matching decision. Therefore, this framework depicts the short-term dynamic patterns of business entities and fuses them with the static representations to generate a comprehensive representation , thereby providing strong timeliness and high accuracy technical support for business recommendations between enterprises.

[0151] S3.1. First, in the dynamic behavior modeling of business entities, the present invention collects the time series behavior data of each business entity to form a behavior sequence , where is the behavior embedding vector of the entity at time , which is generated from the original data by a pre-trained model, represents the behavior embedding vector of the business entity at each time point of dimension size. To incorporate industry background characteristics, the present invention introduces industry characteristic embeddings based on information such as industry classification, enterprise scale, geographical location, etc., laying a foundation for subsequent dynamic analysis, Indicates the dimensionality of the industry feature embedding , which is used to encode the structural static information related to business entities.

[0152] S3.2. Traditional time series modeling usually adopts fixed time granularity (such as days, weeks, months), which is difficult to flexibly adapt to the diverse dynamic patterns of business entities. The present invention introduces the Discrete Wavelet Transform (DWT) to decompose the behavior sequence and automatically extract the dynamic components of different frequencies:

[0153] Apply DWT to each dimension of the behavior sequence to generate the low-frequency component , which reflects the long-term trend of the entity (such as stable trading patterns or operating strategies); generate the high-frequency component ( represents the decomposition level), which captures short-term fluctuations (such as sudden demands or product updates);

[0154] In this solution, for the convenience of unified processing, it is agreed that the low-frequency component is regarded as the "0th-order scale", denoted as ; design an independent Temporal Convolutional Network (TCN) model for each component to extract features: ;

[0155] where is the hidden representation at each scale; TCN can efficiently capture temporal dependencies through convolutional operations, and has stronger parallel computing capabilities and anti-gradient vanishing capabilities.

[0156] S3.3. Further, to overcome the subjectivity that may be introduced by manually specifying the time scale, the present invention proposes an adaptive scale selection mechanism to dynamically fuse the features of each scale through the attention mechanism;

[0157] Input the set of hidden states of each scale into the Multi-Head Attention (MHA) network, using as the query vector to guide the model to focus on the temporal components most relevant to the industry features. Calculate the weights through softmax normalization, and weighted fusion to obtain the comprehensive temporal representation ;

[0158] This method can not only adapt to the behavior characteristics of different business entities and industries, but also highlight the relevance between short-term trends and industry features, thus significantly improving the flexibility and accuracy of dynamic modeling.

[0159] S3.4. To dynamically evaluate the impact of the historical behavior of business entities on the current business needs and resource matching and highlight the key time points related to industry characteristics, the present invention proposes an industry characteristic-driven multi-head attention mechanism;

[0160] First, combine the comprehensive time series representation with the industry characteristic embedding to calculate the attention weights . Subsequently, the present invention introduces a value mapping matrix to perform a linear projection on the weighted multi-scale representation and aggregate the outputs of all attention heads at all time steps, thereby obtaining the short-term dynamic comprehensive representation of the business entity at the current moment . The calculation method is as follows:

[0161] ;

[0162] In the initial implementation, set the number of attention heads to 4 to balance the computational complexity and the ability to capture different time features. At the same time, the dimension of the dynamic representation is set to match the dimension of the static embedding , that is , to ensure the dimensional compatibility when fusing the subsequent static and dynamic representations.

[0163] S3.5. To improve the model's ability to understand and distinguish the dynamic behavior of business entities, the present invention proposes a contrastive learning framework that combines cross-time and cross-entity. This framework establishes a contrastive mechanism based on the time-step specific multi-scale hidden states to characterize the time evolution features and the fine-grained differences between entities;

[0164] In the time dimension, to model the behavioral continuity and evolution features of a single business entity at different times, the present invention introduces a cross-time contrastive learning mechanism. Consider the representations of adjacent time steps ( and ) as positive sample pairs to construct a positive guidance for behavioral continuity. At the same time, form negative sample pairs with the representations of the current time step and competing entities to enhance the ability to distinguish the dynamic behavior between entities, that is , where represents the set of entities competing with entity . Its loss function is:

[0165] ;

[0166] where is a temperature hyperparameter that adjusts the sparsity of the similarity distribution. This loss encourages the model to bring the representations of the same entity at adjacent moments closer together in the representation space, while distinguishing them from the representations of competing entities.

[0167] S3.6. In the entity dimension, in order to enhance the model's ability to distinguish entity relationships in different industries, the present invention further introduces a cross-entity comparative learning strategy. This strategy uses industry similarity information to construct positive and negative sample pairs, thereby improving the model's ability to distinguish behavior patterns in different industries;

[0168] Specifically, industry-similar entities are used as positive samples ,in Representation and Entity A collection of entities belonging to the same industry to reflect the potential business synergy between the same industry. At the same time, entities with large industry differences are selected as negative samples, so that the model can effectively expand the distribution of entities from different industries in the representation space, that is, ,in Representation and Entity A collection of entities belonging to different industries.

[0169] The loss function for cross-entity contrastive learning is defined as follows:

[0170] .

[0171] S4. Use the gated network to fuse static and dynamic features, build a multi-party matching scoring mechanism based on comprehensive representation, and support multi-party business collaboration and recommendation (feature fusion and modeling prediction) in the industrial chain. The specific steps are:

[0172] S4.1. The present invention combines long-term stable static representation and dynamic representations that reflect recent behavior , in order to enhance the model's comprehensive modeling capabilities for business entities; among them, It comes from multi-view knowledge graph and convolutional propagation, reflecting its role and resource advantages in the industrial chain; The dynamic behavior response of enterprises is captured through time series data and attention mechanism;

[0173] In order to fuse the two types of features, a learnable gating network is designed to adaptively adjust the fusion ratio of static features and dynamic features:

[0174] ;

[0175] in and is the projection parameter matrix, Mainly used for static features To project, is used to project the dynamic features for projection; is the bias term; is the activation function, which normalizes the output value to the interval [0, 1];

[0176] Each element represents the importance of the stable business advantages of the enterprise in the corresponding dimension; the larger the value, the more the dimension depends on static features, and vice versa, it pays more attention to recent dynamics;

[0177] Subsequently, the two features are weighted and fused through element-wise multiplication to obtain the final comprehensive representation of the industrial chain business entity:

[0178] ;

[0179] where represents element-wise multiplication.

[0180] S4.2. For the complex multi-party interaction relationships among different business entities in the supply chain, the present invention constructs a multi-party matching scoring mechanism based on the comprehensive representation. For any two enterprises and with potential business synergy effects, their fused comprehensive representations and are calculated respectively, and a matching score is constructed:

[0181] ;

[0182] where is the Sigmoid activation function, which is used to normalize the output to the interval [0, 1]; represents the multi-layer perceptron function, and its form is as follows:

[0183] ;

[0184] where represents the vector concatenation operation, , is the weight matrix, , is the bias term.

[0185] S4.3. In practical applications, business matching in the industrial chain usually involves multiple links and multiple parties. Therefore, a multi-party collaborative matching mechanism is designed to comprehensively consider the business complementarity and overall synergy effects among multiple enterprises;

[0186] For the target enterprise , a group of enterprises is selected from the candidate enterprise set to make the overall business matching effect optimal; the specific implementation method is as follows:

[0187] Based on pairwise matching scores and using the attention mechanism to quantify the target enterprise and the candidate enterprises The interaction importance between them is calculated by exponential normalization operation to obtain the attention weights: ;

[0188] Among them, represents the relative importance of enterprise to the target enterprise in business collaboration;

[0189] After obtaining the attention weights these weights are further used to weighted aggregate the comprehensive representations of each enterprise in the candidate enterprise set to generate the context representation of the target enterprise: ;

[0190] Among them, reflects the overall interaction relationship between the target enterprise and the candidate enterprise set;

[0191] To further quantify the business matching potential of the target enterprise in the multi-party collaboration scenario, the model combines the comprehensive representation and the context representation of the target enterprise to calculate its final business recommendation score: ;

[0192] Among them, represents the vector concatenation operation, is the learnable weight matrix, is the bias term.

[0193] To verify the effectiveness of the above method, the following experiments were conducted:

[0194] Benchmark dataset

[0195] To verify the effectiveness of the proposed method, the present invention uses the industrial chain business dataset from 2018 to 2024 provided by a regional manufacturing industry alliance. This dataset is constructed based on Internet of Things and cloud technologies, covering six R & D centers, ten manufacturing bases, and eight service centers, and systematically collects high-frequency interaction data and resource management information in the operation of the industrial chain. The dataset reflects the characteristics of a digital and intelligent regional industrial chain, covering key entities such as suppliers, R & D enterprises, manufacturers, and service enterprises, and comprehensively reflects the industrial chain process from upstream resource supply to downstream business demand. In addition, the dataset includes business category information and a large number of enterprise interaction records, demonstrating the diversity and dynamic characteristics of the industrial chain. To ensure the temporal consistency and fairness of the experiment, 80% of the interaction history records of each enterprise are extracted in chronological order as the training set, and the remaining 20% as the test set. The statistical data of the dataset is shown in Table 1:

[0196] Table 1

[0197]

[0198] Experimental parameter settings

[0199] The experiment was completed on a computer equipped with the following hardware: the operating system is Ubuntu 22.04, equipped with Intel(R) Xeon(R) Platinum 8488C*2 and NVIDIA RTX A6000 *4 graphics processors, and the memory capacity is 256 GB. The entire experiment uses the Adam optimizer, with a learning rate of 0.001, a batch size of 128, and 100 epochs of training (using early stopping strategy). In the multi-view knowledge graph construction module, the original feature dimension is set to 256; the edge weight parameters and are set to 0.6 and 0.4 respectively; the GCN module adopts a 3-layer structure, , . The business entity modeling module uses 4-head multi-head attention (the dimension of the mapping matrix is 256), the dimension of the intermediate mapping layer in cross-view fusion is 128, and the contrast learning temperature parameter . In the dynamic business entity modeling module, the behavior embedding dimension is 128, and the industry embedding dimension is 64; decomposition is performed using discrete wavelet transform (high-frequency level S = 2), the TCN uses 2 layers of convolution (the convolution kernel size is 3), and the output dimension of the multi-head attention module is 256. In the feature fusion and model prediction module, both static and dynamic features are 256-dimensional, and adaptive fusion is performed through a gated network (matrix size 256×256).

[0200] Evaluation metrics

[0201] To comprehensively analyze the recommendation performance of the model of the present invention, recall (Recall@K) and normalized discounted cumulative gain (NDCG@K) are used as evaluation metrics. Recall@K represents the proportion of the top K recommended services that are relevant to the actual needs of the enterprise, reflecting the recall ability of the model. NDCG@K takes into account the ranking of relevant services in the recommendation list and evaluates the ranking quality of the recommendations. To evaluate the performance of the model under different recommendation list lengths, K is set to 10, 20, and 50 in the experiment. The values of these metrics range between 0 and 1, and the higher the value, the better the recommendation performance.

[0202] Comparative experiment

[0203] To prove the effectiveness of the method (model) of the present invention, several advanced methods (models) are selected in the experiment for comparison with this method. The existing methods include: Bayesian personalized ranking matrix factorization (BPR-MF): By optimizing with the Bayesian personalized ranking (BPR) loss, it enhances the pairwise matrix factorization of implicit feedback, thus effectively learning the implicit interaction between enterprises and services. Lightweight graph convolutional network (LightGCN): A lightweight graph neural network that simplifies GCN operations and only focuses on neighbor aggregation, making it highly scalable in large-scale recommendation tasks. Neural graph collaborative filtering model (NGCF): A graph-based framework that incorporates collaborative filtering signals into message passing to capture the high-order connectivity between users and items. Knowledge graph attention network (KGAT): Introducing an attention mechanism to learn high-order dependencies in the knowledge graph and enhancing the representation of users and items. Knowledge graph-based intention learning framework (KGIN): A GNN-based model that can capture the relational structure in the knowledge graph and use auxiliary semantic signals to model user preferences. Knowledge-enhanced graph contrastive learning (KGCL): Using contrastive learning to improve the representation of the knowledge graph, reducing noise and solving the long-tail item problem. Item-specific graph attention network (IGAT): Combining interaction and knowledge graph information to capture high-order neighbor information through knowledge-aware and item-specific attention mechanisms, improving recommendation performance. Multi-view knowledge graph convolutional network (MKGCN): By constructing multi-views, introducing initial residual connections, and graph self-attention mechanisms, it alleviates noise interference and over-smoothing problems and improves recommendation performance. Knowledge-aware fine-grained attention network (KFGAN): Proposing a knowledge-aware fine-grained attention network framework to improve personalized recommendation performance by capturing high-order collaborative signals and refining knowledge graph embeddings.

[0204] Experimental results

[0205] Table 2 shows the experimental results of the model of the present invention (hereinafter referred to as this model) and each baseline model on the industrial chain multi-dimensional dataset:

[0206] Table 2

[0207] Model Recall@10 NDCG@10 Recall@20 NDCG@20 Recall@50 NDCG@50 BPR-MF 0.0448 0.4003 0.1024 0.4024 0.2563 0.3643 LightGCN 0.0597 0.4164 0.1171 0.4559 0.2713 0.3832 NGCF 0.0747 0.5284 0.1317 0.4928 0.2864 0.4258 KGAT 0.0597 0.4719 0.1463 0.5194 0.3165 0.4339 KGIN 0.0747 0.5288 0.1610 0.5562 0.3216 0.4447 KGCL 0.0896 0.5475 0.1756 0.5767 0.3466 0.4824 IGAT 0.1014 0.5884 0.2196 0.6667 0.3617 0.4957 MKGCN 0.1044 0.6065 0.2049 0.6460 0.3768 0.5094 KFGAN 0.0896 0.5536 0.1903 0.6114 0.3919 0.5263 This model 0.1194 0.6357 0.2478 0.6904 0.4205 0.5714

[0208] As can be seen from Table 2, the proposed model outperforms the comparison methods in all evaluation metrics, fully verifying the innovative designs of the present invention in aspects such as multi-view knowledge graph construction, business entity multi-head fusion, dynamic time-series modeling, and feature adaptive fusion.

[0209] Compared with other methods, models based on traditional graph collaborative filtering such as LightGCN and NGCF, although performing well in capturing local structures, fail to effectively integrate enterprise multi-dimensional features and time-series information. In the present invention, by constructing a multi-view knowledge graph, static features of enterprises can be accurately extracted; the lack of traditional models in this regard results in relatively weak performance in Recall@10 and NDCG@10, and they are unable to fully capture the long-term needs of enterprises in the long list recommendation (Recall@50 and NDCG@50) task.

[0210] Models such as KGAT and KGIN improve the expression ability of multi-dimensional behaviors through graph attention mechanisms and relation-aware information propagation, but they mainly focus on static knowledge graph information and fail to conduct in-depth modeling by combining business context and dynamic time-series features. In contrast, the proposed model effectively integrates enterprise background and business relationships, making the model more advantageous in capturing complex enterprise behaviors, especially more obvious in Recall@20, Recall@50, and the corresponding NDCG metrics.

[0211] Although models such as IGAT and MKGCN improve the recommendation accuracy when modeling various enterprise business types, they have deficiencies in time-series feature extraction and noise suppression, unable to distinguish long-term trends and short-term fluctuations, thus performing slightly worse when dealing with complex industrial chain data. Similarly, KGCL and KFGAN enhance the model robustness through contrastive learning but fail to achieve in-depth fusion of business context and dynamic information. The proposed dynamic business entity modeling and feature fusion and prediction mechanism, through modules such as discrete wavelet transform, time-series convolutional network, and learnable gating network, significantly improves the ability to capture the dynamics of enterprise behaviors and the matching effect, obtaining more robust performance in long-tail recommendation metrics.

[0212] To further demonstrate the performance fluctuations and stability of the proposed model in different evaluation metrics, box plots were used to visualize the performance of the model. Figure 2Shows the performance distribution of this model on multiple metrics such as Recall@10, Recall@20, Recall@50, NDCG@10, NDCG@20, and NDCG@50. Each box plot represents the performance distribution under different metrics, providing the minimum, maximum, quartiles, and median of the model's performance in multiple experiments. These results can intuitively demonstrate the consistency and reliability of the model on multiple evaluation metrics, further confirming the advantages of this model in different recommendation tasks.

[0213] Computational efficiency and memory usage

[0214] The comparison of computational efficiency and memory usage is shown in Table 3. It can be seen that this model achieves the best balance between computational complexity and efficiency:

[0215] Table 3

[0216] Model Training time (minutes) Inference time (seconds) Training memory (GB) Inference memory (GB) BPR-MF 23 0.29 2.38 0.26 LightGCN 31 0.35 4.18 0.43 NGCF 51 0.45 5.42 0.53 KGAT 68 0.42 7.52 0.81 KGIN 43 0.56 6.89 0.91 KGCL 74 0.69 7.84 1.15 IGAT 84 0.62 8.23 1.03 MKGCN 62 0.72 8.51 1.25 KFGAN 47 0.51 6.05 0.85 This model 66 0.67 7.35 0.97

[0217] Compared with simple models such as BPR-MF and LightGCN, this model uses slightly more memory during the training and inference processes. This is a trade-off for its ability to capture complex enterprise behaviors and provide context-aware recommendations, which is particularly suitable for dynamic industrial chain environments. In terms of training time, this model provides a balanced approach that can meet the requirements of large-scale datasets while maintaining competitive running efficiency. Although the training time of this model is slightly longer compared to some benchmark models, it can process richer data and generate accurate, context-driven recommendations, which justifies the additional computational cost. During the inference process, this model performs efficiently and can provide real-time recommendations with reasonable latency even when dealing with complex temporal and relational data.

[0218] The present invention has been described in detail above in conjunction with the embodiments, but the above content is only the preferred embodiments of the present invention and cannot be considered as limiting the scope of implementation of the present invention. All equivalent changes and improvements made in accordance with the scope of the present invention application shall still fall within the scope covered by the patent of the present invention.

Claims

1. A multi-view knowledge graph-driven industrial chain collaborative recommendation method, characterized in that Including the following steps: S1. Construct knowledge graph views of multiple business dimensions from multi-source heterogeneous data, and perform feature decomposition and adaptive feature regulation; S2. Propose a multi-head knowledge propagation mechanism that integrates business context, and perform hierarchical convolutional aggregation and dynamic path selection in combination with multi-view representations; S3. Through wavelet transform and multi-scale time series decomposition, combine the adaptive scale selection mechanism and the attention mechanism driven by industry characteristics to construct a dynamic representation of business entities; S4. Use a gated network to fuse static and dynamic features, construct a multi-party matching scoring mechanism based on comprehensive representation, and support multi-party business collaboration and recommendation in the industrial chain; Specifically, step S2 is as follows: S2.

1. For the view below, for any business entity and its neighbors , the attention weight of the -th attention head is defined as : ； Among them, is the linear transformation matrix of the th attention head, is the entity embedding; is an entity and its neighbor the business relationship indicator between them; is a fusion function, is a query vector; S2.

2. After the above-mentioned multi-head propagation of attention weights, the updated representation of the entity is defined as: ； Among them, represents the set of neighbor nodes of the entity is the transformation matrix in the th attention head represents the feature representation of the neighbor entity under the th view represents the concatenation operation; S2.

3. For the initial representation of nodes , obtain the representations of each layer through multi-layer convolutional propagation , and introduce a simplified residual structure during the final aggregation. Only the initial features are weighted and fused with the propagation results of each level to obtain the aggregated representation of the node: ； Among them, is the number of propagation layers adaptively determined by the th node; is the trainable weight; S2.

4. Calculate the change amount between the representations of adjacent convolutional layers by the computing nodes , combined with the adaptive threshold , where and are the statistical features of the node changes in the previous layer, is a hyperparameter. When , it is considered that the features of the th node have tended to be stable, and further propagation is stopped; S2.

5. Adopt the following fusion method: For each business entity , the representation under each view is , and the attention mechanism is used to assign dynamic weights to different perspectives: ； , where is the dimension of ; the middle layer dimension is the traversal variable for the view index, is the total number of views, represents the transpose operation of a vector or matrix, is the bias term in the attention mechanism, is the trainable query vector in the attention mechanism; obtain the long-term static embedding representation of the business entity : ； S2.

6. Construct a consistency loss function , where is the cosine similarity is the temperature parameter 2. The collaborative recommendation method for industrial chain driven by multi-view knowledge graph according to claim 1, wherein Specifically, step S1 is as follows: S1.

1. Build an original knowledge graph based on multi-source heterogeneous data including enterprise information, business reports, and historical transaction data G : ； Among them, represents an entity set containing multiple business entities; represents a relationship set of multiple business relationships; S1.

2. Using domain expert knowledge and pre-trained models, perform multi-dimensional decomposition on the original enterprise feature matrix to extract key sub-features, where represents that this matrix belongs to a matrix space of dimension in the real number field; represents the size of the entity set ; represents the number of feature dimensions of each entity; for each view , calculate the correlation between the th feature of the th enterprise node and the business matching metric , and obtain the feature retention probability through the normalization function : ; Among them , represents the values of all entities on the th feature dimension; represents all rows; represents the th feature dimension, y is the supervision signal for business matching, ; represents the correlation score between the feature and the business objective; represents the base of the natural logarithm; Generate a binary mask matrix through independent sampling based on the feature retention probability : ; Among them indicates that in the view the th feature of the th enterprise node is retained, indicating suppression; By independent sampling with a probability of highly correlated features are selectively retained and noise and low-correlation features are suppressed, thereby obtaining entity feature matrices for different views : ; S1.

3. For the original knowledge graph for any two business entities and compute the edge weight for the relationship edge between them W ij as follows ; where conf( e i , e j ) is the edge confidence obtained from multi-source data, , U is the number of interactions; After encoding the entity description using the BERT model, the cosine similarity of the calculated feature vectors is used to measure the semantic similarity between entities; is an adjustable hyperparameter; activation function Normalize the weights to the interval [0, 1], ; x is the pre-activation value of the edge weight; Sampling probability of the edge is defined as its weight value: ; For each edge based on generate a subgraph after independent sampling through the Bernoulli distribution; repeat the sampling process to generate multiple different subgraph views: ; Adjacency matrix of each view is defined as: ； If edge (e i ,e j ) is retained in view v, then the corresponding adjacency matrix has a value of 1; otherwise, has a value of 0; S1.

4. Entity feature matrix based on specific views and the corresponding adjacency matrix , construct a multi-layer graph convolutional network GCN, and use the improved graph convolutional network to perform embedded representation learning on all business entities; first, define the initial node feature matrix : ； The layer improvement graph convolution propagation rule in the th view is as follows: ； Among them, , where is the identity matrix; is the initial node feature matrix; is the -th view's -th layer convolutional network weight matrix; is the adjacency matrix after adding self-loops to view ; is the corresponding degree matrix; and are hyperparameters that control the retention ratios of the initial features and the previous layer features, respectively; is the activation function; After layer improvement of graph convolution propagation, the entity node embedding representation of the th view is obtained : .

3. The multi-view knowledge graph-driven industrial chain collaborative recommendation method according to claim 2, wherein: In step S1.1, the multiple business entities include enterprises, products, services, and demands; the multiple business relationships include supply, cooperation, competition, and demand-supply; in step S1.2, the key sub-features include business demands, resource capabilities, and market positions; in step S1.3, the initial value ; In step S1.4, the initial value .

4. The multi-view knowledge graph-driven industrial chain collaborative recommendation method according to claim 1, wherein Specifically, step S3 is as follows: S3.

1. In the dynamic behavior modeling of business entities, for each business entity Collect its time series behavior data to form a behavior sequence , where is the behavior embedding vector of the entity at time , represents the behavior embedding vector of the business entity at each time point ; Based on information including industry classification, enterprise scale, and geographical location, introduce industry characteristic embedding , represents the dimension size of the industry characteristic embedding ; S3.

2. Apply discrete wavelet transform to each dimension of the behavior sequence to generate a low-frequency component ; generate a high-frequency component , representing the decomposition level; S3.

3. Aggregate the hidden states at each scale and input them into the multi-head attention network MHA with the industry feature embedding as the query vector to guide the model to focus on the time series components most relevant to the industry features; calculate the weights through normalization using the softmax function and obtain the comprehensive time series representation through weighted fusion ; S3.

4. Adopt a multi-head attention mechanism driven by industry characteristics: First, represent the comprehensive timing and embed it with industry characteristics for combination to calculate the attention weights . Subsequently, introduce the value mapping matrix , linearly project the weighted multi-scale representation, and aggregate the outputs of all attention heads at all time steps to obtain the short-term dynamic comprehensive representation of the business entity at the current moment . The calculation method is as follows: ； is the number of attention heads, is the dimension of the dynamic representation; S3.

5. Treat the representations at adjacent times and as a positive sample pair to construct a positive guidance for behavior continuity for the multi-scale hidden state specific to the time step; meanwhile, form a negative sample pair with the representation of the current time and the competing entity to enhance the discrimination ability of the dynamic behavior between entities, that is , where represents the set of entities competing with entity ; its time loss function is ； wherein is a temperature hyperparameter; S3.

6. Use industry-similar entities as positive samples , where represents the set of entities belonging to the same industry as entity ; meanwhile, select entities with significant industry differences as negative samples , where represents the set of entities belonging to different industries from entity ; Loss function for cross-entity contrastive learning is defined as follows: 。 5. The multi-view knowledge graph-driven industrial chain collaborative recommendation method according to claim 4, wherein: In step S3.2, the low-frequency component is regarded as the 0th-order scale, denoted as ; an independent temporal convolutional network model TCN is designed for each component to extract features: ; wherein is the hidden representation at each scale; In step S3.4, in the initial implementation, the number of attention heads is set to 4 to balance the computational complexity and the ability to capture different temporal features; meanwhile, the dimension of the dynamic representation is set to match that of the static embedding dimension , that is .

6. The multi-view knowledge graph-driven industrial chain collaborative recommendation method according to claim 4, wherein Specifically, step S4 is as follows: S4.

1. Combine the long-term stable static representation with the dynamic representation reflecting recent behaviors ; where it is derived from the multi-view knowledge graph and convolutional propagation; while the dynamic behavior response of the enterprise is captured through time-series data and the attention mechanism; Adopt a learnable gated network to adaptively adjust the fusion ratio of static and dynamic features: ； wherein and are projection parameter matrices, is a parameter matrix for projecting static features ; is a parameter matrix for projecting dynamic features ; is a bias term; is an activation function that normalizes the output value to the interval [0, 1]. Each element represents the importance of the enterprise's stable business advantages in the corresponding dimension; the larger the value, the more this dimension depends on static features, and vice versa, the more it focuses on recent dynamics; Subsequently, the two features are weighted and fused through element-wise multiplication to obtain the final comprehensive representation of the industrial chain business entity : ； wherein represents element-by-element multiplication; S4.

2. Construct a multi-party matching scoring mechanism based on the comprehensive representation; for any two enterprises with potential business synergy and , calculate their fused comprehensive representations and respectively, and construct a matching score : ； Among them, is the Sigmoid activation function, which is used to normalize the output to the interval [0, 1]; represents the multi-layer perceptron function, and its form is as follows: ； Among them, represents a vector concatenation operation, , is the weight matrix, , is the bias term; S4.

3. For the target enterprise , select a group of enterprises from the candidate enterprise set as follows: Based on pairwise matching scores , the attention mechanism is used to quantify the target enterprise and the candidate enterprise The interaction importance between them is calculated by exponential normalization operation to obtain the attention weights: ; Among them, represents the enterprise in the business collaboration for the target enterprise in the relative importance; After obtaining the attention weights further use these weights to perform weighted aggregation on the comprehensive representations of each enterprise in the candidate enterprise set to generate the context representation of the target enterprise : ; Among them, reflects the overall interaction relationship between the target enterprise and the candidate enterprise set; Combined with the comprehensive representation of the target enterprise and the context representation , calculate its final business recommendation score: ; Among them, represents a vector concatenation operation, is a learnable weight matrix, is a bias term.

Images