An intelligent matching method and system for enterprise value-added services based on multimodal learning

Abstract

The present invention relates to the field of industrial optimization technology, and discloses an intelligent matching method and system for enterprise value-added services based on multi-modal learning. The method includes the following steps: dynamically evaluating the strong and weak chain relationships between enterprises, constructing a time-varying adjacency matrix, and updating the chain strength weights through a propagation correction mechanism; fusing the temporal characteristics of enterprise operation efficiency and innovation ability, and generating dual-channel fusion features through a cross-modal attention mechanism; constructing a three-dimensional tensor from the strong and weak chain scores, efficiency, and innovation features, and generating final fusion features through CP decomposition and multi-head attention enhancement. By extracting service features and modeling enterprise needs, and combining similarity calculation, the present invention recommends the most suitable services, and uses a clustering algorithm to identify enterprise demand patterns, enabling the recommendation strategy to be dynamically adjusted according to real-time data, reducing the need for manual intervention, improving the speed and efficiency of decision-making, enabling enterprises to quickly respond to market changes, and optimizing resource allocation.

Description

Technical Field

[0001] The present invention relates to the field of industrial optimization, and more specifically, to an intelligent matching method and system for enterprise value-added services based on multimodal learning. Background Art

[0002] Enterprises are facing fierce market competition and ever-changing customer needs, which have promoted the demand for value-added services.

[0003] Traditional value-added service recommendation systems rely on manual rules (such as SIC industry classification), which will lead to two typical problems: over-recommendation and service omission, thus causing the inability to achieve accurate matching and dynamic recommendation of enterprise service resources. Summary of the Invention

[0004] The present invention provides an intelligent matching method and system for enterprise value-added services based on multimodal learning to solve the technical problems in the related art.

[0005] The present invention provides an intelligent matching method for enterprise value-added services based on multimodal learning, including the following steps:

[0006] S100, dynamically evaluating the strong and weak chain relationships between enterprises, constructing a time-varying adjacency matrix, and updating the chain strength weights through a propagation correction mechanism;

[0007] S200, fusing the temporal characteristics of enterprise operation efficiency and innovation ability, and generating dual-channel fusion features through a cross-modal attention mechanism;

[0008] S300, constructing a three-dimensional tensor from the strong and weak chain scores, efficiency, and innovation features, and generating the final fusion features through CP decomposition and multi-head attention enhancement;

[0009] S400, constructing a dynamic decision-making model based on feature similarity calculation and demand pattern clustering, and outputting service recommendation results;

[0010] Output a recommendation list according to the service recommendation results, and provide a corresponding matching degree score for each service.

[0011] Furthermore, in S100, the specific steps are as follows:

[0012] S110, supply chain hypergraph construction: aggregating real-time transaction data streams to construct a weighted hypergraph, where the hyperedges represent multi-enterprise collaboration relationships;

[0013] S120, topological centrality calculation: calculating the betweenness centrality of enterprise nodes based on the hypergraph structure to identify the key hubs in the supply chain;

[0014] S130, policy sensitivity analysis: parsing the policy document set, extracting keywords in the technical research and development catalog, and calculating the enterprise policy matching degree;

[0015] S140, Dynamic Score Synthesis: Integrate topological importance and policy sensitivity to generate dynamic strong and weak link scores;

[0016] S150, Abnormal Link Detection: Identify sudden disconnection risks based on the sliding window mechanism.

[0017] Furthermore, in the steps of S110 - S150, the specific calculation formulas are as follows:

[0018] ;

[0019] where represents a hyperedge, represents the total transaction volume of the hyperedge within the time window, represents the variance of the trading time interval, represents the k - th time window;

[0020] ;

[0021] where represents the betweenness centrality of the enterprise node , represents the total number of the shortest hyperpaths from the enterprise node to , represents the number of the shortest hyperpaths passing through the enterprise node ;

[0022] ;

[0023] where represents the text feature weight of the enterprise node , which combines the word frequency and the inverse document frequency to reflect the uniqueness of its text information, represents the number of occurrences of the keyword in the document of the enterprise node , represents the total number of policy documents, represents the number of documents containing the keyword , represents the set of all keywords in the document of enterprise i, represents the total number of keywords;

[0024] ;

[0025] where represents the change amount of the chain structure score within the time , where represents the disconnection risk threshold, , is an indicator function, indicating the change in score.

[0026] Furthermore, in S200, the specific steps are as follows:

[0027] S210, Construction of Efficiency Feature Map: Map the efficiency index matrix into a weighted graph structure, where the edge weights reflect the index similarity;

[0028] S220, Temporal Graph Convolution on Innovation: Perform temporal graph convolution on the innovation factor tensor to capture dynamic innovation patterns;

[0029] S230, Cross-Modal Attention Alignment: Establish a cross-modal attention mechanism between efficiency features and innovation features;

[0030] S240, Dynamic Coupling Analysis: Integrate dual-channel features and calculate the coupling coefficient;

[0031] S250, Multi-Scale Feature Normalization: Perform hierarchical normalization on the integrated features.

[0032] Furthermore, in the steps of S210 - S250, the specific calculation formulas are as follows:

[0033] ;

[0034] where represents the efficiency association weight between the i-th unit and the j-th unit, represents the Gaussian kernel bandwidth, , is the efficiency index vector of enterprise node , , represents the square of the Euclidean norm of the vector, is the efficiency index vector of enterprise node ;

[0035] ;

[0036] where , respectively represent the l-th and (l + 1)-th innovation feature matrices, represents the time window , , represents the cross-layer propagation matrix, , represents the innovation factor tensor, represents at time point the innovation feature map;

[0037] ;

[0038] ;

[0039] where represents the original attention score of the \(i\)-th enterprise and the \(j\)-th enterprise based on efficiency and innovation ability, represents the original attention score of the \(j\)-th and the \(k\)-th enterprises based on efficiency and innovation ability, represents the feature dimension, aligned with the three-dimensional fusion step, , represents the -th enterprise and the -th enterprise's association strength between efficiency and innovation, represents the dot product operation between two vectors, represents the time series feature of efficiency of the \(i\)-th enterprise, represents the time series feature of innovation of the \(j\)-th enterprise;

[0040] ;

[0041] ;

[0042] where represents the learning parameter for calculating the fusion coefficient, , represents the dual-channel fusion feature; represents the concatenation result of the efficiency time series feature and the innovation time series feature; represents element-wise multiplication, represents the fusion coefficient calculated through the sigmoid function, represents the innovation time series feature, represents the efficiency time series feature;

[0043] ;

[0044] where represents the standard value of the industrial cluster feature to which it belongs, represents the mean value of the industrial cluster feature to which it belongs, represents the numerical stability constant, represents the feature vector of the \(i\)-th enterprise after standardization, represents the original feature vector of the \(i\)-th enterprise.

[0045] Furthermore, in S300, the specific steps are as follows:

[0046] S310, heterogeneous tensor construction: Integrate the scoring ranges of strong and weak chains, dual-channel fusion features, and innovation time series features to construct a three-dimensional feature tensor;

[0047] S320, Core Factorization: Perform CP decomposition on the three-dimensional tensor to extract cross-dimensional shared features;

[0048] S330, Spatiotemporal Dynamic Calibration: Align the time axes of multi-source features using dynamic time warping;

[0049] S340, Multi-Head Feature Enhancement: Strengthen key feature dimensions through the multi-head attention mechanism;

[0050] S350, Fusion Vector Generation: Synthesize the final three-dimensional feature vector and perform normalization.

[0051] Furthermore, in the steps of S310 - S350, the specific calculation formulas are as follows:

[0052] ;

[0053] where represents the three-dimensional feature tensor, represents the total number of enterprise nodes, represents the dimension inherited from cross-modal attention, represents the enterprise node the scoring range of strong and weak chains, represents the i-th dual-channel fusion feature, represents the i-th innovative time series feature;

[0054] ;

[0055] where represents the decomposition rank, , represents the mode importance weight, represents the strong and weak chain scoring weight, represents the dual-channel fusion feature weight, represents the innovative time series feature weight;

[0056] ;

[0057] where represents the aligned multi-source features, and respectively represent the feature points or data at time steps and under, , is the index, represents the alignment penalty coefficient, is the calibration path, represents the square of the Frobenius norm of the policy change;

[0058] ;

[0059] ;

[0060] where represents the weight matrix of the output layer, represents the k-th value matrix that maps the input to the value space, represents the k-th query matrix, represents the k-th key matrix, represents the number of attention heads, , , represents the dimension of the input vector, and the dimension of each head is ;

[0061] ;

[0062] where represents the empirical weight, the error or variance loss after performing principal component analysis on , is the result of the core tensor mode expansion after decomposition, represents the enhanced feature to be optimized, represents the L2 regularization term for to constrain the feature scale,

[0063] Further, the specific steps in S400 are as follows:

[0064] S410, Service Feature Extraction: Extract the service feature matrix from the service database, including information such as the basic attributes and functional characteristics of the service, and perform standardization processing;

[0065] S420, Requirement Feature Modeling: Construct a requirement feature tensor to represent the multi-dimensional features of user requirements, and use a clustering algorithm to identify requirement patterns;

[0066] S430, Similarity Calculation: Calculate the similarity between the service features and the requirement features, and use cosine similarity or Euclidean distance for comparison;

[0067] S440, Decision Model Construction: Based on the similarity matrix, construct a decision model to recommend the most matching service.

[0068] Further, the specific steps in S410 - S440 are as follows:

[0069] ;

[0070] where is the feature mean, is the characteristic standard deviation, represents the service feature matrix after standardization processing, represents the service feature matrix, , where is the number of services, is the feature dimension;

[0071]

[0072] where represents the demand feature, is the user demand data set, represents the clustering algorithm, represents extracting features according to characteristics;

[0073] ;

[0074] where is the service and the demand similarity, is the standardized feature vector of the service , is the standardized feature vector of the demand ;

[0075] ;

[0076] where represents the dynamic weight, represents the business rule weight vector, , represents the temperature coefficient, is the recommendation score of the service , is the service all similarity values.

[0077] The present invention also provides an intelligent matching system for enterprise value-added services based on multimodal learning, including one or more steps in the foregoing intelligent matching method for enterprise value-added services based on multimodal learning, including:

[0078] Strong and weak chain relationship dynamic evaluation module: Dynamically evaluate the strong and weak chain relationships between enterprises and construct a time-varying adjacency matrix;

[0079] Operation efficiency and innovation time-series feature fusion module: Fusion the time-series features of the operation efficiency and innovation capabilities of enterprises to generate dual-channel fusion features;

[0080] Three-dimensional tensor construction and enhancement module: Construct a three-dimensional tensor from the strong and weak chain scores, efficiency, and innovation features, and generate the final fusion features through CP decomposition and multi-head attention enhancement;

[0081] Dynamic decision-making and service recommendation module: Based on feature similarity calculation and demand pattern clustering, a dynamic decision-making model is constructed to output service recommendation results, a recommendation list is output according to the service recommendation results, and a corresponding matching degree score is provided for each service.

[0082] The beneficial effects of the present invention are as follows:

[0083] By extracting service features and modeling enterprise requirements, and combining similarity calculation, the present invention can accurately identify enterprise requirements, thereby recommending the most suitable services. The clustering algorithm is used to identify enterprise demand patterns, enabling the recommendation strategy to be dynamically adjusted according to real-time data. The automated service recommendation process reduces the need for manual intervention, improves the speed and efficiency of decision-making, enables enterprises to quickly respond to market changes, and optimizes resource allocation. Brief Description of the Drawings

[0084] Figure 1 is a flowchart of an intelligent matching method for enterprise value-added services based on multi-modal learning proposed by the present invention;

[0085] Figure 2 is a structural block diagram of an intelligent matching system for enterprise value-added services based on multi-modal learning proposed by the present invention;

[0086] Figure 3 is a recommendation list output according to the recommended most-matched service of the present invention;

[0087] In the figure: 101, dynamic evaluation module for chain relationship; 102, fusion module for operation efficiency and innovation time sequence features; 103, three-dimensional tensor construction and enhancement module; 104, dynamic decision-making and service recommendation module. Detailed Embodiments

[0088] Now, the subject matter described herein will be discussed with reference to exemplary embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein. Without departing from the protection scope of the content of this specification, the functions and arrangements of the elements discussed can be changed. Each example can omit, substitute, or add various processes or components as needed. Additionally, the features described for some examples can also be combined in other examples.

[0089] As Figure 1 shown, an intelligent matching method for enterprise value-added services based on multi-modal learning includes the following steps:

[0090] S100, dynamically evaluate the strong and weak chain relationships between enterprises, construct a time-varying adjacency matrix, and update the chain strength weights through a propagation correction mechanism;

[0091] In one embodiment of the present invention, the specific steps are as follows:

[0092] S110, Supply chain hypergraph construction: Aggregate real-time transaction data streams , construct a weighted hypergraph , where hyperedges represent multi-enterprise collaboration relationships;

[0093] Where represents the transaction volume of the enterprise node, represents the energy or resources of the enterprise node;

[0094] Its calculation formula is as follows:

[0095] ;

[0096] Where represents the hyperedge, represents the hyperedge The total transaction volume within the time window T, and T represents the total time window represents the variance of the transaction time interval (measuring stability), represents the k-th time window;

[0097] S120, Topological centrality calculation: Calculate the betweenness centrality of enterprise nodes based on the hypergraph structure to identify key hubs in the supply chain;

[0098] Its calculation formula is as follows:

[0099] ;

[0100] Where represents the enterprise node of the betweenness centrality, represents the enterprise node to the total number of shortest hyperpaths, represents the number of shortest hyperpaths passing through the enterprise node ;

[0101] S130, Policy sensitivity analysis: Analyze the policy document set , extract keywords in the technology research and development catalog, and calculate the policy matching degree of enterprises;

[0102] Its calculation formula is as follows:

[0103] ;

[0104] Where represents the text feature weight of the enterprise node , combined with term frequency and inverse document frequency, reflecting the uniqueness of its text information, represents the keyword in the document of enterprise i The number of occurrences, represents the total number of policy documents, represents the number of documents containing the keyword . represents the set of all keywords in the documents of enterprise i, represents the total number of keywords;

[0105] S140, Dynamic score synthesis: Integrate topological importance and policy sensitivity to generate dynamic strong - weak link scores;

[0106] Its calculation formula is as follows:

[0107]

[0108] where the balance coefficient (determined by grid search), represents the score range, ;

[0109] S150, Abnormal link detection: Identify sudden disconnection risks based on the sliding window mechanism;

[0110] Its calculation formula is as follows:

[0111] ;

[0112] where represents the change in the link structure score over time , where represents the disconnection risk threshold, , is the indicator function, represents the change in the score;

[0113] S200, Integrate the temporal characteristics of enterprise operation efficiency and innovation ability, and generate dual - channel fusion features through the cross - modal attention mechanism;

[0114] In an embodiment of the present invention, the specific steps are as follows:

[0115] S210, Construction of the efficiency feature map: Map the efficiency index matrix to a weighted graph structure , and the edge weights reflect the index similarity;

[0116] Its calculation formula is as follows:

[0117] ;

[0118] where represents the efficiency correlation weight between the i - th unit and the j - th unit, represents the Gaussian kernel bandwidth, , is the performance index vector of the enterprise , , represents the square of the Euclidean norm (i.e., L2 norm) of the vector;

[0119] S220, Innovative Temporal Graph Convolution: Perform temporal graph convolution (T-GCN) on the innovative element tensor to capture dynamic innovation patterns;

[0120] Its calculation formula is as follows:

[0121] ;

[0122] where and represent the l-th and (l + 1)-th innovative feature matrices respectively, represents the time window of the convolution kernel, , represents the cross-layer propagation matrix, , represents the innovative element tensor, represents the innovative feature map at the time point ;

[0123] S230, Cross-modal Attention Alignment: Establish a cross-modal attention mechanism between performance features and innovative features;

[0124] Its calculation formula is as follows:

[0125] ;

[0126] ;

[0127] where represents the original attention score between the i-th and j-th enterprises based on performance and innovation capabilities, represents the original attention score between the j-th and k-th enterprises based on performance and innovation capabilities, represents the feature dimension, aligned with the three-dimensional fusion step, , represents the -th and -th enterprise's association strength between performance and innovation, represents the dot product operation between two vectors; represents the innovative temporal feature of the j-th enterprise;

[0128] S240, Dynamic Coupling Analysis: Fuse dual-channel features and calculate the coupling coefficient;

[0129] Its calculation formula is as follows:

[0130] ;

[0131] ;

[0132] where represents the learning parameter for calculating the fusion coefficient, , represents the dual-channel fusion feature; represents the concatenation result of the efficiency feature and the innovation time series feature; represents element-wise multiplication, represents the fusion coefficient calculated through the sigmoid function, represents the innovation time series feature, represents the efficiency time series feature;

[0133] S250, multi-scale feature normalization: perform hierarchical normalization on the fusion feature;

[0134] Its calculation formula is as follows:

[0135] ;

[0136] where represents the standard value of the feature of the affiliated industrial cluster, represents the mean value of the feature of the affiliated industrial cluster, represents the numerical stability constant, , represents the feature vector of the i-th enterprise after normalization, represents the original feature vector of the i-th enterprise;

[0137] S300, construct a three-dimensional tensor from the strong and weak link scores, efficiency, and innovation features, and generate the final fusion feature through CP decomposition and multi-head attention enhancement;

[0138] In an embodiment of the present invention, the specific steps are as follows:

[0139] S310, heterogeneous tensor construction: integrate the score ranges of the strong and weak links , the dual-channel fusion feature and the innovation time series feature to construct a three-dimensional feature tensor;

[0140] Its calculation formula is as follows:

[0141] ;

[0142] where represents the total number of enterprise nodes, represents the dimension inherited from cross-modal attention, ;

[0143] S320, Core factor decomposition: Perform CP decomposition on the three-dimensional tensor to extract cross-dimensional shared features;

[0144] Its calculation formula is as follows:

[0145] ;

[0146] Where represents the decomposition rank, (the decomposition rank is determined by the BIC criterion), represents the pattern importance weight, , and represent the coefficients or weights in the formula, which are used for itemized calculation of the aligned fusion features. Specifically, represents the strong-weak chain scoring weight, represents the dual-channel fusion feature weight, represents the innovation time-series feature weight;

[0147] S330, Spatiotemporal dynamic calibration: Use dynamic time warping (DTW) to align the time axes of multi-source features;

[0148] Where its calculation formula is as follows:

[0149] ;

[0150] Where represents the aligned multi-source features, and respectively represent the time steps (or different perspectives) and the feature points or data at , is the index, (the alignment penalty coefficient, of the same order of magnitude as the in the strong-weak chain evaluation), is the calibration path, represents the square of the Frobenius norm of the policy change;

[0151] S340, Multi-head feature enhancement: Strengthen the key feature dimensions through the multi-head attention mechanism;

[0152] Its calculation formula is as follows:

[0153] ;

[0154] ;

[0155] Where represents the concatenation operation, Represents the weight matrix of the output layer, represents the k-th Value matrix, which maps the input to the value space, represents the k-th Query matrix, which maps the input to the query space, represents the k-th Key matrix, which maps the input to the key space, represents the number of attention heads, , , represents the dimension of the input vector, and the dimension of each head is ;

[0156] S350, Fusion vector generation: Synthesize the final three-dimensional feature vector and perform normalization;

[0157] Its calculation formula is as follows:

[0158] ;

[0159] where (empirical weight, determined by grid search), represents the error or variance loss after principal component analysis, is the result of the core tensor mode expansion after decomposition, represents the enhanced feature to be optimized, represents the L2 regularization term of constrains the feature scale,

[0160] S400, Based on feature similarity calculation and demand pattern clustering, construct a dynamic decision model to output service recommendation results;

[0161] In an embodiment of the present invention, the specific steps are as follows:

[0162] S410, Service feature extraction: Extract the service feature matrix , including information such as the basic attributes and functional characteristics of the service, and perform standardization processing;

[0163] Its calculation formula is as follows:

[0164] ;

[0165] where is the feature mean, is the feature standard deviation, represents the service feature matrix after standardization processing, represents the service feature matrix, , where is the number of services, is the feature dimension;

[0166] S420, Requirement Feature Modeling: Construct a requirement feature tensor , representing the multi-dimensional features of user requirements, and use clustering algorithms to identify requirement patterns;

[0167] Its calculation formula is as follows:

[0168] ;

[0169] where represents the requirement feature, is the user requirement data set, represents the clustering algorithm, represents extracting features according to characteristics;

[0170] S430, Similarity Calculation: Calculate the similarity between service features and requirement features, and use cosine similarity or Euclidean distance for comparison;

[0171] Its calculation formula is as follows:

[0172] ;

[0173] where is the service and the requirement similarity, is the service normalized feature vector, is the requirement normalized feature vector;

[0174] S440, Decision Model Construction: Based on the similarity matrix , construct a decision model to recommend the most matching service, usually using sorting or threshold decision;

[0175] The calculation formula of its decision model is as follows:

[0176] ;

[0177] where (dynamic weight, maintaining magnitude continuity with α = 0.6 for three-dimensional fusion), : represents the business rule weight vector, represents the feature vector output by three-dimensional fusion, is the recommended score for the service , is all the similarity values of the service .

[0178] In the service matching decision-making process, "recommend the most suitable service" means that based on the similarity calculation results of service features ( ) and requirement features ( ), through a sorting or threshold screening mechanism, select services from the candidate service set that meet the following conditions: )

[0179] Maximize similarity: The recommended score of service satisfies , where is the multi-dimensional similarity vector between it and the requirement;

[0180] Dynamic adaptability: The matching result needs to be aligned with the feature dimensions output by the three-dimensional feature fusion to ensure semantic space consistency;

[0181] Interpretability constraint: The final recommendation needs to satisfy (threshold , calibrated through historical interaction data) to avoid low-confidence matches.

[0182] According to the above process, the following is an example:

[0183] A manufacturing enterprise needs to optimize its cross-border logistics link. The requirement is described as "reduce transportation costs (cost weight 40%), and shorten the average delivery time to 7 days (timeliness weight 60%)".

[0184] There are 50 logistics services in the service library, and the feature dimension d = 64 (including cost, timeliness, stability, etc.);

[0185] The implementation process of recommending the most suitable service:

[0186] Service feature extraction: Extract the standardized feature matrix of the logistics service , where the cost item of the i-th service is , and the timeliness item is ;

[0187] Requirement feature modeling: Analyze the requirement to generate a weight vector , and fill in default values for other secondary feature dimensions through clustering.

[0188] Similarity calculation:

[0189] Calculate the weighted cosine similarity:

[0190] ;

[0191] Decision-making and recommendation:

[0192] If the threshold , this service satisfies ;

[0193] After global sorting, this service (ranked first);

[0194] Finally, recommend this logistics service and mark the matching dimension: "The comprehensive matching degree of cost - timeliness is 82%".

[0195] Recommended definition mapping:

[0196] In this case, the "most matching service" needs to satisfy simultaneously:

[0197] Similarity (interpretability constraint);

[0198] Align with the demand weight in the 64 - dimensional feature space (dynamic adaptability);

[0199] Through Achieve global optimality (maximize similarity).

[0200] As Figure 3 shown, according to the above - mentioned recommended most - matching service, output a recommendation list and provide a corresponding matching degree score for each service.

[0201] As Figure 2 shown, according to one or more steps in the above - mentioned intelligent matching method for enterprise value - added services based on multi - modal learning, the present invention also proposes an intelligent matching system for enterprise value - added services based on multi - modal learning, including:

[0202] Chain - relationship dynamic evaluation module 101: Dynamically evaluate the strong and weak chain relationships between enterprises and construct a time - varying adjacency matrix;

[0203] Operation - efficiency and innovation - time - series feature fusion module 102: Fuse the time - series features of the operation efficiency and innovation ability of enterprises to generate dual - channel fusion features;

[0204] Three - dimensional tensor construction and enhancement module 103: Construct a three - dimensional tensor from the strong - weak chain scores, efficiency, and innovation features, and generate the final fusion features through CP decomposition and multi - head attention enhancement;

[0205] Dynamic decision - making and service recommendation module 104: Based on feature similarity calculation and demand pattern clustering, construct a dynamic decision - making model to output service recommendation results, output a recommendation list according to the service recommendation results, and provide a corresponding matching degree score for each service.

[0206] The above has described the embodiments of this example, but this example is not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative rather than restrictive. Under the inspiration of this example, those of ordinary skill in the art can also make many forms, all of which fall within the protection scope of this example.

Claims

1. An intelligent matching method for enterprise value-added services based on multimodal learning, characterized in that, It includes the following steps: S100, Dynamically evaluate the strong and weak chain relationships among enterprises, construct a time-varying adjacency matrix, and update the chain strength weights through a propagation correction mechanism; The specific steps are as follows: S110, Supply chain hypergraph construction: Aggregate real-time transaction data streams to construct a weighted hypergraph, where hyperedges represent multi-enterprise collaboration relationships; ； Among them represents a hyperedge, represents the total transaction volume of the hyperedge within the time window, represents the variance of the trading time interval, represents the k-th time window, represents the total time window, represents the hyperedge weight; S120, Topological centrality calculation: Calculate the betweenness centrality of enterprise nodes based on the hypergraph structure to identify key hubs in the supply chain; ； Among them represents the betweenness centrality of the enterprise node ; represents the total number of the shortest hyperpaths from the enterprise node to ; represents the number of the shortest hyperpaths passing through the enterprise node . S130, Policy sensitivity analysis: Analyze policy document sets, extract keywords for technology research and development catalogs, and calculate the policy matching degree of enterprises; ； Among them represents the text feature weight of the enterprise node , which combines the word frequency and inverse document frequency to reflect the uniqueness of its text information represents the enterprise node in the document the number of occurrences of the keyword represents the total number of policy documents represents the number of documents containing the keyword ; represents the set of all keywords in the document of enterprise i represents the total number of keywords S140, Dynamic score synthesis: Integrate topological importance and policy sensitivity to generate dynamic strong and weak chain scores; Among them represents the scoring range of enterprise nodes ; S150, Abnormal link detection: Identify sudden disconnection risks based on a sliding window mechanism; ； wherein represents the change amount of the chain structure score within the time wherein represents the broken chain risk threshold, , is an indicator function, represents the change amount of the score; S200, Integrate the temporal features of enterprise operation efficiency and innovation ability, and generate dual-channel fusion features through a cross-modal attention mechanism; S300, Construct a three-dimensional tensor from the strong and weak chain scores, efficiency, and innovation features, and generate final fusion features through CP decomposition and multi-head attention enhancement; S400, Based on feature similarity calculation and demand pattern clustering, construct a dynamic decision-making model to output service recommendation results; Output a recommendation list according to the service recommendation results, and provide a corresponding matching degree score for each service.

2. The intelligent matching method for enterprise value-added services based on multimodal learning according to claim 1, wherein In S200, the specific steps are as follows: S210, Efficiency feature map construction: Map the efficiency index matrix into a weighted graph structure, and the edge weights reflect index similarity; S220, Innovation temporal graph convolution: Perform temporal graph convolution on the innovation factor tensor to capture dynamic innovation patterns; S230, Cross-modal attention alignment: Establish a cross-modal attention mechanism between efficiency features and innovation features; S240, Dynamic coupling analysis: Integrate dual-channel features and calculate the coupling coefficient; S250, Multi-scale feature normalization: Perform hierarchical normalization processing on the fusion features.

3. The intelligent matching method for enterprise value-added services based on multi-modal learning according to claim 2, wherein In the steps of S210, the specific calculation formula is as follows: ； wherein represents the efficiency correlation weight between the i-th unit and the j-th unit, represents the Gaussian kernel bandwidth, , is the efficiency index vector of enterprise node , , represents the square of the Euclidean norm of the vector; In the steps of S220, the specific calculation formula is as follows: ； Among them and represent the l-th and (l + 1)-th innovation feature matrices respectively, represents the time window of the convolutional kernel, , represents the cross-layer propagation matrix, , represents the innovation factor tensor, represents the innovation feature map at the time point ; In the steps of S230, the specific calculation formula is as follows: ； ； Among them represents the original attention score of the \(i\)-th enterprise and the \(j\)-th enterprise based on efficiency and innovation ability represents the original attention score of the \(j\)-th enterprise and the \(k\)-th enterprise based on efficiency and innovation ability represents the feature dimension, aligned with the three-dimensional fusion step , represents the -th enterprise and the -th enterprise's correlation strength between efficiency and innovation represents the dot product operation between two vectors represents the efficiency time-series feature of the \(i\)-th enterprise represents the innovation time-series feature of the \(j\)-th enterprise; In the steps of S240, the specific calculation formula is as follows: ； wherein represents the learning parameter for calculating the fusion coefficient, , represents the dual-channel fusion feature; represents the concatenation result of the efficacy time-series feature and the innovation time-series feature; represents element-wise multiplication, represents the fusion coefficient calculated through the sigmoid function, represents the innovation time-series feature, represents the efficacy time-series feature; In the steps of S250, the specific calculation formula is as follows: ； wherein represents the characteristic standard value of the affiliated industrial cluster, represents the characteristic mean value of the affiliated industrial cluster, represents the numerical stability constant, represents the characteristic vector of the i-th enterprise after standardization, represents the characteristic vector of the original i-th enterprise.

4. An intelligent matching method for enterprise value-added services based on multimodal learning according to claim 3, characterized in that, In S300, the specific steps are as follows: S310, Heterogeneous tensor construction: Integrate the scoring ranges of strong and weak chains, dual-channel fusion features, and innovation temporal features to construct a three-dimensional feature tensor; S320, Core factor decomposition: Perform CP decomposition on the three-dimensional tensor to extract cross-dimensional shared features; S330, Spatiotemporal dynamic calibration: Align the time axes of multi-source features using dynamic time warping; S340, Multi-head feature enhancement: Strengthen key feature dimensions through a multi-head attention mechanism; S350, Fusion vector generation: Synthesize the final three-dimensional feature vector and perform normalization.

5. The intelligent matching method for enterprise value-added services based on multimodal learning according to claim 4, wherein In the steps of S310, the specific calculation formula is as follows: ； Among them represents the three-dimensional feature tensor represents the total number of enterprise nodes represents the dimension inherited from cross-modal attention represents the enterprise node the scoring range of strong and weak chains represents the i-th dual-channel fusion feature represents the i-th innovation time series feature; In the steps of S320, the specific calculation formula is as follows: ； Among them represents the decomposition rank, , represents the importance weight, represents the strong-weak chain scoring weight, represents the dual-channel fusion feature weight, represents the innovation timing feature weight; In the steps of S330, the specific calculation formula is as follows: ； Among them represents the aligned multi-source features, and respectively represent the feature points or data at time step and ; , is the index, represents the alignment penalty coefficient, is the calibration path, represents the square of the Frobenius norm of the policy change; In the steps of S340, the specific calculation formula is as follows: ； ； where represents the weight matrix of the output layer, represents the k-th value matrix that maps the input to the value space, represents the k-th query matrix, represents the k-th key matrix, represents the number of attention heads, , , represents the dimension of the input vector, and the dimension of each head is ; In the steps of S350, the specific calculation formula is as follows: Among them represents the empirical weight, the error or variance loss after performing principal component analysis on and is the core tensor modal expansion result after decomposition represents the enhanced features to be optimized represents the L2 regularization term for, used to constrain the feature scale represents the feature vector of the three-dimensional fusion output 6. The intelligent matching method for enterprise value-added services based on multimodal learning according to claim 5, wherein The specific steps in S400 are as follows: S410, Service Feature Extraction: Extract the service feature matrix from the service database, including the basic attributes and functional characteristic information of the service, and perform standardization processing; S420, Requirement Feature Modeling: Construct a requirement feature tensor to represent the multi-dimensional features of user requirements, and use a clustering algorithm to identify requirement patterns; S430, Similarity Calculation: Calculate the similarity between the service features and the requirement features, and use cosine similarity or Euclidean distance for comparison; S440, Decision Model Construction: Based on the similarity matrix, construct a decision model to recommend the most matching service.

7. An intelligent matching method for enterprise value-added services based on multimodal learning according to claim 6, characterized in that In the step of S410, the specific calculation formula is as follows: ； where is the feature mean value, is the feature standard deviation, represents the service feature matrix after standardization processing, represents the service feature matrix, , where is the number of services, is the feature dimension; In the step of S420, the specific calculation formula is as follows: ； Among them represents a demand feature is the user demand data set represents a clustering algorithm represents extracting features according to characteristics In the step of S430, the specific calculation formula is as follows: ； wherein is the service and the requirement similarity of, is the service standardized eigenvector of, is the requirement standardized eigenvector of; In the step of S440, the specific calculation formula is as follows: ； Among them represents the dynamic weight, represents the business rule weight vector, , represents the temperature coefficient, is the recommendation score for service , is all the similarity values for service .

8. An intelligent matching system for enterprise value-added services based on multimodal learning, characterized in that, The steps in a method for intelligent matching of enterprise value-added services based on multi-modal learning as described in any one of claims 1-7 include: Chain Relationship Dynamic Evaluation Module: Dynamically evaluate the strong and weak chain relationships between enterprises and construct a time-varying adjacency matrix; Operation Efficiency and Innovation Time-Series Feature Fusion Module: Fuse the time-series features of the operation efficiency and innovation capabilities of enterprises to generate dual-channel fusion features; Three-Dimensional Tensor Construction and Enhancement Module: Construct a three-dimensional tensor from the strong and weak chain scores, efficiency, and innovation features, and generate the final fusion features through CP decomposition and multi-head attention enhancement; Dynamic Decision and Service Recommendation Module: Based on feature similarity calculation and requirement pattern clustering, construct a dynamic decision model to output service recommendation results, output a recommendation list according to the service recommendation results, and provide a corresponding matching score for each service.

Images