Intelligent docking method for technology transfer based on hypergraph-knowledge graph fusion network

Abstract

The application provides a technology transfer intelligent docking method based on a hypergraph-knowledge graph fusion network, comprising the following steps: accessing data; extracting structural features based on a hypergraph; calculating semantic association features based on a knowledge graph; enhancing structural features in a semantic guided manner; matching technology demands and achievements; dynamically recommending and sorting based on supply-demand potential; obtaining a training model by maximizing a target function; and outputting intelligent docking. The application proposes a technology transfer intelligent docking method based on a hypergraph-knowledge graph fusion network, uses a hypergraph to model complex technology supply-demand relations, uses a knowledge graph to supplement field semantics, enhances structural features in a semantic guided manner, realizes deep interaction of features, and uses semantic coverage and structural satisfaction to perform reasoning matching and dynamic recommendation based on supply-demand potential.

Description

Technical Field

[0001] This invention relates to the technical field of knowledge graph construction, and in particular to a method for intelligent technology transfer based on a hypergraph-knowledge graph fusion network. Background Technology

[0002] In the global wave of innovation-driven development, technology transfer has evolved from sporadic, accidental occurrences into a core engine driving economic growth. The fundamental purpose of technology transfer is to break down the barriers between innovation and industry. For universities and research institutions, it is a crucial path to transform papers and patents into real productivity, realizing social value and economic returns. For enterprises, especially SMEs, it is an important way to acquire cutting-edge technologies, address R&D shortcomings, and achieve leapfrog development. For nations, technology transfer is a strategic cornerstone for optimizing resource allocation, accelerating industrial upgrading, and building core competitiveness. Efficient technology transfer relies on increasingly diverse technological means. Traditional technology transfer methods include patent licensing and technology transfer contracts. Today, the development of big data and artificial intelligence is reshaping the technology transfer process. Intelligent algorithms can accurately match technology suppliers and demanders, online platforms have built a global technology trading market, and digital twin and simulation technologies allow technology acquirers to complete virtual verification before investing heavily, greatly reducing industrialization risks. Furthermore, spin-off companies, joint R&D, and technology-for-equity partnerships have made technology transfer more diverse and in-depth.

[0003] Technology transfer service organizations, positioned at the apex of the technology transfer network between supply and demand, can accelerate the exchange of information, resources, and technology between institutions. Internet-based online matching and transaction platforms have constructed a technology transfer service model that transcends time and space and aggregates massive amounts of information, reducing the search and transaction costs of technology transfer. Currently, the integration of online and offline services based on the internet has taken initial shape, but its intelligence is relatively insufficient, and service efficiency and quality need improvement. Utilizing big data, artificial intelligence, and other means to empower technology evaluation and matching, facilitate the integration of online and offline services, and drive innovation in technology transfer models requires in-depth research and active practice.

[0004] Chinese invention patent application CN113779264A discloses a transaction recommendation method based on a patent supply and demand knowledge graph. The steps of this invention are: (1) Attribute system construction: patent data is collected based on the IncoPat patent database, and an attribute index system affecting patent transactions is constructed based on literature research; (2) Patent transaction AHN construction: a patent transaction AHN including three types of nodes, three types of relationships and multiple attributes is constructed; (3) Representation learning of patent transaction AHN: including obtaining the multidimensional Gaussian distribution of nodes based on neural networks, generating node sequences based on meta-path random walks, and obtaining the low-dimensional vector space of each node by using skip-gram and KL divergence to measure the difference between the Gaussian distributions of different nodes; (4) Top-k recommendation of the target organization. The disadvantage of this method is that it only uses knowledge graph technology and does not use hypergraph technology. This results in the method only utilizing the speech information of the knowledge graph and lacking the structural information of the hypergraph. The essential feature of "demand is a collection of multiple technical points and there are complex semantic relationships between technologies" in the docking process lacks sufficient representation. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention proposes an intelligent technology transfer matching method based on a hypergraph-knowledge graph fusion network. This method utilizes a hypergraph to model complex technology supply and demand relationships, supplements domain semantics with a knowledge graph, and employs a hypergraph neural network and a graph convolutional network for joint feature extraction and alignment. Finally, it leverages feature similarity for hypergraph-path-based reasoning and matching.

[0006] This invention provides a method for intelligent technology transfer based on a hypergraph-knowledge graph fusion network, including data access and the following steps:

[0007] Step 1: Extracting structural features based on the hypergraph;

[0008] Step 2: Calculate semantic association features based on knowledge graph;

[0009] Step 3: Semantic-guided structural feature enhancement;

[0010] Step 4: Matching technical needs with deliverables;

[0011] Step 5: Dynamic recommendation ranking based on supply and demand potential;

[0012] Step 6: Use historically successful connection cases as positive samples and unrelated technology combinations in the knowledge graph as negative samples to maximize the objective function and obtain the trained model.

[0013] Step 7: Intelligent docking output.

[0014] Preferably, the access data includes the following sub-steps:

[0015] Step 01: Access enterprise requirement documents, university / research institute achievement databases, and industry technical dictionaries and standards;

[0016] Step 02: Each demand exceeds the edge e i Includes a set of technology vertices v j The hypergraph H = (V, E) is composed of the set of entities V in the enterprise technology field and the set of hyperedges E.

[0017] Step 03: Construct the underlying knowledge graph G=(V,R) from the set of entities V in the enterprise technology field and the semantic relationships R between entities;

[0018] Where V = {v1, v2, ..., v} j ,...,v n}, E={e1,e2,...,e i ,...,e m}, where n is the number of vertices and m is the number of hyperedges.

[0019] In any of the above solutions, step 1 preferably includes the following sub-steps:

[0020] Step 11: Propagate vertex information using the hypergraph Laplacian matrix and define the incidence matrix. If vertex v i Belongs to demand e j Then H ij =1, otherwise 0;

[0021] Step 12: Calculate the degree matrix D of the hyperedge. e The degree matrix D of the vertices v The formula is:

[0022]

[0023]

[0024] Among them, H ij H represents the value at position (I,j) in the correlation matrix. ji This represents the value at position (j,i) in the correlation matrix;

[0025] Step 13: Calculate the normalized hypergraph Laplacian matrix L, using the formula:

[0026]

[0027] Where I is the identity matrix;

[0028] The formula for hypergraph convolutional layer propagation is:

[0029]

[0030] in, Let l be the feature matrix of the vertices in the l-th layer. Here are the trainable parameters, and σ is the activation function;

[0031] Step 14: After all layers of convolution are completed, the structural feature matrix of all vertices is X. struct .

[0032] In any of the above solutions, step 2 preferably includes the following sub-steps:

[0033] Step 21: Use a relational graph convolutional network to extract vertex features from the knowledge graph G, obtaining the feature h of vertex v. v ;

[0034] Step 22: Utilize the semantic information of the knowledge graph to semantically associate the features of vertex v with feature u. v calculate;

[0035] Step 23: After the calculation is completed, the semantic association feature matrix Usem of all vertices is obtained.

[0036] In any of the above schemes, the preferred method is that the formula for calculating the semantic association feature is:

[0037]

[0038] in, Let W be the set of vertices adjacent to vertex v under relation r. r Let h be the weight matrix of relation r. s Let be the characteristics of vertex s that is adjacent to vertex v.

[0039] Preferably, in any of the above schemes, the semantically guided structural feature enhancement method includes the following sub-steps:

[0040] Step 31: Calculate semantic association features U sem For structural feature X Struct The contribution of each dimension is used to generate semantic association features U through linear transformation. sem For structural feature X Struct query vector Q SemStruct and bond vector K Struct The formula is:

[0041]

[0042]

[0043] in, and The weight matrix is ​​a learnable matrix;

[0044] Step 32: Calculate the attention weight α for the structural feature channel dimension. Struct The formula is:

[0045]

[0046] Where ⊙ represents element-wise multiplication, and d is the feature dimension;

[0047] Step 33: Calculate the structural features for semantic association feature enhancement The formula is:

[0048] .

[0049] In any of the above schemes, step 4 preferably includes treating the technical requirements as a query vector, the technical supply as candidate vertices, and calculating the hyperedge-vertex matching degree, including the following sub-steps:

[0050] Step 41: Let e ​​be the hyperedge corresponding to the technology requirement Q of a certain enterprise. Q It contains a set of key technology vertices v Q Let v be the set of vertices corresponding to the technical achievement T to be matched. T v Q ⊆V, v T ⊆ V;

[0051] Step 42: Calculate the matching score S(Q,T) based on the dual constraints of semantic coverage and demand satisfaction.

[0052] In any of the above schemes, the method for calculating the semantic coverage preferably includes:

[0053] Demand beyond the edge e Q Features are aggregated into a single vector :

[0054]

[0055] Calculation result feature vector With demand feature vector cosine similarity Sim Semantic The formula is:

[0056]

[0057] MeanPooling is the average pooling operation, Z i This is a feature of fusion.

[0058] In any of the above solutions, the method for calculating the degree of demand satisfaction includes:

[0059] According to the vertex set vT Is it related to e? Q The vertices in the graph are closely connected in the knowledge graph G. Calculate v T For e Q Structural Satisfaction Sim struct The formula is:

[0060]

[0061]

[0062] Among them, f rel (v t , v x Let γ be the correlation function between two vertices in the knowledge graph, and let γ be the decay function dist(v). t ,v x ) for v t and v x The path distance in the graph.

[0063] In any of the above schemes, the preferred method is that the matching score S(Q,T) is calculated using the following formula:

[0064]

[0065] Where λ is the weight of semantic coverage. This is the overall demand feature vector. As the feature vector of the result, Sim Semantic For the feature vector of the result With overall demand feature vector cosine similarity, , Sim struct For the vertex set v T For the superedge e Q Structural satisfaction.

[0066] In any of the above schemes, step 5 preferably includes introducing a market dynamic factor, calculating the final recommendation score Rec(Q,T), and sorting according to the recommendation score.

[0067] In any of the above schemes, the preferred method is that the final recommendation score Rec(Q,T) is calculated using the following formula:

[0068]

[0069] Where S(Q,T) is the matching score, and H T Let β be the current market popularity of the technological achievement T, β be the adjustment coefficient, and Maturity be the technological maturity. Maturity is a maturity penalty term.Q -Maturity T |This represents the maturity gap between enterprise needs (Q) and outcomes (T).

[0070] In any of the above schemes, the preferred option is that the total loss function L of the training model is... total The calculation formula is:

[0071] L total =Rec(Q,T).

[0072] In any of the above schemes, step 7 preferably includes, for a new enterprise demand Qnew, retrieving relevant vertices in the hypergraph network, calculating the recommendation score Rec(Qnew,Ti) for all candidate outcomes Ti, outputting a Top-K recommendation list, and visually displaying the matching path between the demand and the outcome points.

[0073] This invention proposes an intelligent technology transfer matching method based on a hypergraph-knowledge graph fusion network, which breaks through the limitations of traditional keyword-based or simple knowledge graph-based methods. It utilizes hypergraphs to model complex technology supply and demand relationships, uses knowledge graphs to supplement domain semantics, and achieves intelligent matching through path reasoning under dual constraints. This method can effectively handle complex one-to-many and many-to-one scenarios in enterprise technology transfer. Attached Figure Description

[0074] Figure 1 This is a flowchart of a preferred embodiment of the intelligent technology transfer docking method based on a hypergraph-knowledge graph fusion network according to the present invention.

[0075] Figure 2 This is a schematic diagram of an embodiment of the hypergraph structure of the intelligent technology transfer docking method based on the hypergraph-knowledge graph fusion network according to the present invention. Detailed Implementation

[0076] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0077] Example 1

[0078] like Figure 1 As shown, step 1000, data access, includes the following sub-steps:

[0079] Step 1010: Access enterprise requirement documents, university / research institute achievement databases, and industry technical dictionaries and standards;

[0080] Execution step 1020: Each demand exceeds edge e i Includes a set of technology vertices v j The hypergraph H = (V, E) is composed of the set of entities V in the enterprise technology field and the set of hyperedges E.

[0081] Execution step 1030: Construct the underlying knowledge graph G=(V,R) from the set of entities V in the enterprise technology field and the semantic relationships R between entities;

[0082] Where V = {v1, v2, ..., v} j ,...,v n}, E={e1,e2,...,e i ,...,e m}, where n is the number of vertices and m is the number of hyperedges.

[0083] Step 1100 involves extracting structural features based on the hypergraph, including the following sub-steps:

[0084] Execute step 1110, propagate vertex information using the hypergraph Laplacian matrix, and define the incidence matrix. If vertex v i Belongs to demand e j Then H ij =1, otherwise 0;

[0085] Execute step 1120 to calculate the degree matrix D of the hyperedge. e The degree matrix D of the vertices v The formula is:

[0086]

[0087]

[0088] Among them, H ij H represents the value at position (I,j) in the correlation matrix. ji This represents the value at position (j,i) in the correlation matrix;

[0089] Execute step 1130 to calculate the normalized hypergraph Laplacian matrix L, using the formula:

[0090]

[0091] Where I is the identity matrix;

[0092] The formula for hypergraph convolutional layer propagation is:

[0093]

[0094] in, Let l be the feature matrix of the vertices in the l-th layer. Here are the trainable parameters, and σ is the activation function;

[0095] After performing step 1140, and completing the convolution of all layers, the structural feature matrix of all vertices is X. struct .

[0096] Step 1200 involves calculating semantic association features based on the knowledge graph, including the following sub-steps:

[0097] Execute step 1210, using a relational graph convolutional network to extract vertex features from the knowledge graph G, obtaining the feature h of vertex v. v .

[0098] Execute step 1220, utilizing the semantic information of the knowledge graph, to semantically associate the features of vertex v in the knowledge graph with feature u. v The formula for calculating the semantic association features is as follows:

[0099]

[0100] in, Let W be the set of vertices adjacent to vertex v under relation r. r Let h be the weight matrix of relation r. s Let be the characteristics of vertex s that is adjacent to vertex v.

[0101] After executing step 1230 and completing the calculation, the semantic association feature matrix Usem of all vertices is obtained.

[0102] Step 1300 involves semantically guided structural feature enhancement, the method of which includes the following sub-steps:

[0103] Execute step 1310 to calculate semantic association feature U sem For structural feature X Struct The contribution of each dimension is used to generate semantic association features U through linear transformation. sem For structural feature X Struct query vector Q SemStruct and bond vector K Struct The formula is:

[0104]

[0105]

[0106] in, and The weight matrix is ​​a learnable matrix;

[0107] Execute step 1320 to calculate the attention weight α in the structural feature channel dimension. Struct The formula is:

[0108]

[0109] Where ⊙ represents element-wise multiplication, and d is the feature dimension;

[0110] Execute step 1330 to calculate the structural features enhanced by semantic association features. The formula is:

[0111] .

[0112] Step 1400, matching technology requirements and outcomes, involves treating technology requirements as a query vector and technology supply as candidate vertices, calculating the matching degree between hyperedges and vertices, and includes the following sub-steps:

[0113] Execute step 1410, and let the hyperedge corresponding to the technical requirement Q of a certain enterprise be e. Q It contains a set of key technology vertices v Q Let v be the set of vertices corresponding to the technical achievement T to be matched. T v Q ⊆V, v T ⊆ V.

[0114] Execute step 1420 to calculate the matching score S(Q,T) based on the dual constraints of semantic coverage and requirement satisfaction.

[0115] The method for calculating the semantic coverage includes:

[0116] Demand beyond the edge e Q Features are aggregated into a single vector :

[0117]

[0118] Calculation result feature vector With demand feature vector cosine similarity Sim Semantic The formula is:

[0119]

[0120] MeanPooling is the average pooling operation, Z i This is a feature of fusion.

[0121] The method for calculating the degree of demand satisfaction includes:

[0122] According to the vertex set v T Is it related to e? Q The vertices in the graph are closely connected in the knowledge graph G. Calculate v T For e Q Structural Satisfaction Sim struct The formula is:

[0123]

[0124]

[0125] Among them, f rel (v t , v x Let γ be the correlation function between two vertices in the knowledge graph, and let γ be the decay function dist(v). t ,v x ) for v T and v x The path distance in the graph.

[0126] The formula for calculating the matching score S(Q,T) is as follows:

[0127]

[0128] Where λ is the weight of semantic coverage. This is the overall demand feature vector. For the result feature vector, , .

[0129] Step 1500 involves dynamic recommendation ranking based on supply and demand potential, including introducing market dynamic factors, calculating the final recommendation score Rec(Q,T), and ranking according to the recommendation score. The formula for calculating the final recommendation score Rec(Q,T) is as follows:

[0130]

[0131] Where S(Q,T) is the matching score, and H T Let β be the current market popularity of the technological achievement T, β be the adjustment coefficient, and Maturity be the technological maturity. Maturity is a maturity penalty term. Q -Maturity T |This represents the maturity gap between enterprise needs (Q) and outcomes (T).

[0132] Step 1600 involves using historically successful connection cases as positive samples and unrelated technology combinations from the knowledge graph as negative samples to maximize the objective function and obtain a trained model. The total loss function of the trained model is L. total The calculation formula is:

[0133] L total =Rec(Q,T).

[0134] Step 1700 is executed, and the intelligent docking output includes the following: For the new enterprise requirement Qnew, the system retrieves relevant vertices in the hypergraph network and calculates the recommendation score Rec(Qnew,Ti) for all candidate outcomes Ti. A Top-K recommendation list is output, and the matching path between the requirement and the outcome is visualized.

[0135] Example 2

[0136] This invention constructs a knowledge graph and a supply-demand hypergraph for enterprise technologies. It utilizes the hypergraph to model complex technology supply-demand relationships, supplements domain semantics with the knowledge graph, and employs a hypergraph neural network and a graph convolutional network for joint feature extraction. Finally, it uses feature similarity for hypergraph-based path-based reasoning and matching. This invention aims to overcome the limitations of traditional keyword-based or simple knowledge graph-based methods. By using a hypergraph to model complex technology supply-demand relationships, supplementing domain semantics with the knowledge graph, and achieving intelligent matching through path reasoning under dual constraints, it effectively handles complex scenarios in enterprise technology transfer, such as one-to-many (one achievement satisfies multiple needs) and many-to-one (a complex need requires the integration of multiple achievements).

[0137] This invention includes the following steps:

[0138] 1. Data Access

[0139] By integrating enterprise requirement documents, university / research institute achievement databases, and industry technical dictionaries and standards, an underlying knowledge graph is constructed.

[0140] The set of entities in the enterprise's technology field is V = {v1, v2, ..., v...} n}

[0141] The set of supply and demand relationships is E = {e1, e2, ..., e}. m}

[0142] In H=(V,E), there is a demand hyperedge e i Includes a set of technology vertices v j That is, e i ⊆V.

[0143] The knowledge graph G=(V,R) is a semantic relationship between entities (such as belong_to, apply_to, etc.).

[0144] 2. Hypergraph-based structural feature extraction

[0145] Propagate vertex information using the Laplacian matrix of a hypergraph. Define the incidence matrix. If vertex v i Belongs to demand e j Then H ij =1, otherwise 0.

[0146] The degree matrix D of the hyperedge e The degree matrix D of the vertices v The formula is:

[0147]

[0148]

[0149] The normalized hypergraph Laplacian matrix L is defined as:

[0150]

[0151] Hypergraph convolutional layer propagation formula:

[0152]

[0153] in, It is the feature matrix of the vertices in the l-th layer. σ is a trainable parameter, and σ is the activation function.

[0154] After all layers of convolution are completed, the structural feature matrix of all vertices is X. struct .

[0155] 3. Semantic association feature calculation based on knowledge graph

[0156] To incorporate technical knowledge, we first use a relational graph convolutional network to extract vertex features from the knowledge graph G, obtaining the vertex... The characteristic is h v Secondly, by utilizing the semantic information of the knowledge graph, the vertices in the knowledge graph are analyzed. Semantic association feature calculation is performed on the features.

[0157]

[0158] in W represents the set of vertices adjacent to vertex v under relation r. r It is the weight matrix of relation r. It is an activation function. R, This represents the characteristics of the vertex s that is adjacent to vertex v.

[0159] After the calculation is completed, the semantic association feature matrix of all vertices is U. sem .

[0160] 4. Semantic-guided structural enhancement

[0161] (1) Calculate semantic association features U sem For structural feature X StructThe contribution values ​​of each dimension are used to generate query vectors and key vectors through linear transformation:

[0162]

[0163] Q SemStruct and K Struct Semantic association features U sem For structural feature X Struct The query vector and key vector.

[0164] and It is a learnable weight matrix.

[0165] Calculate the attention weights along the channel dimension of the structural features.

[0166]

[0167] ⊙ represents element-wise multiplication, and d is the feature dimension.

[0168] Computational semantic enhancement of structural features ,

[0169] .

[0170] 5. Matching of Technology Needs and Outcomes

[0171] We can treat technological requirements as a query vector and technological supplies (achievements / patents) as candidate vertices. Since requirements are often hyperedges with multiple constraints, we need to calculate the matching degree between "hyperedges and vertices".

[0172] Let e ​​be the superedge corresponding to a company's technology requirement Q. Q It contains a set of key technology vertices v Q Let v be the set of vertices corresponding to the technical achievement T to be matched. T v Q ⊆V, v T ⊆ V.

[0173] The matching score S(Q,T) is calculated under the dual constraints of semantic coverage and demand satisfaction:

[0174]

[0175] Where λ is the weight of semantic coverage. , .

[0176] (1) Semantic coverage calculation

[0177] First, the demand super-edge e QFeatures are aggregated into a single vector :

[0178]

[0179] MeanPooling represents the average pooling operation. Z i These are fusion features, which have already been calculated above.

[0180] Calculation result vector With demand feature vector Cosine similarity:

[0181]

[0182] (2) Calculate the degree of satisfaction of requirements

[0183] According to the vertex set v T Whether or not The vertices in the graph are closely connected in the knowledge graph G, and v is defined as follows. T right Structural satisfaction:

[0184]

[0185] f rel (v t , v x ) is the correlation function between two vertices in a knowledge graph:

[0186]

[0187] Here, γ() is the decay function, and the decay value is related to dist(v). t , v x Related to ) dist(v t , v x ) represents v t and v x The path distance in the graph.

[0188] 6. Dynamic recommendation ranking based on supply and demand potential

[0189] To address the issue of technology matching but difficulty in transfer, a market dynamics factor is introduced. The final recommendation score Rec(Q,T) is defined as follows:

[0190]

[0191] H T The current market popularity of technological achievement T can be calculated based on recent transaction frequency or search volume. The matching score S(Q,T) has already been calculated.

[0192] Maturity: Technological maturity.

[0193] Maturity penalty, the maturity gap between firm needs (Q) and outcomes (T) | Maturity Q -Maturity T The larger the value, the more difficult the transfer becomes; therefore, a negative exponential penalty is used.

[0194] β is the adjustment coefficient.

[0195] 7. Model Training

[0196] Using historically successful integration cases as positive samples and unrelated technology combinations from the knowledge graph as negative samples, the total loss function is designed as follows:

[0197] L total =Rec(Q,T)

[0198] The trained model is obtained by maximizing the objective function.

[0199] 8. Intelligent docking output

[0200] For a new enterprise requirement Qnew, the system retrieves relevant vertices in the hypergraph network and calculates the recommendation score Rec(Qnew,Ti) for all candidate outcomes Ti. It outputs a Top-K recommendation list and visualizes the matching paths between requirements and outcomes.

[0201] This invention combines graph neural networks and hypergraph neural networks, elevating single technical indicator queries to structured semantic queries based on hypergraphs. This method overcomes the limitations of traditional keyword-based or simple knowledge graphs (containing only entities and binary relations). It utilizes hypergraphs to model complex technology supply and demand relationships, supplements domain semantics with knowledge graphs, and achieves intelligent matching through path reasoning under dual constraints. This effectively handles complex scenarios in enterprise technology transfer, such as one-to-many (one achievement satisfies multiple needs) and many-to-one (a complex need requires the integration of multiple achievements).

[0202] This innovative approach separates the "multi-faceted supply and demand structure relationship" and the "technological semantic relationship" into two independent graphs. A hypergraph represents the multi-faceted supply and demand relationship where "one demand corresponds to multiple technological points," while a knowledge graph G represents the semantic relationships between technologies, such as synonyms, hierarchical relationships, and causal relationships. Existing technologies either merge the hypergraph and knowledge graph into a single hypergraph knowledge graph or model all relationships using only a single graph structure. This separate modeling approach better reflects the essential characteristic of technology supply and demand matching: "demand is a collection of multiple technological points, and there are complex semantic relationships between technologies."

[0203] Cross-reconstruction mechanism: The core challenge of structural features (higher-order topological relations) extracted from hypergraphs and semantic association features (attributes / semantics of entities and relations) extracted from knowledge graphs lies in the fact that structural features are local and discrete topological representations, while semantic features are global and continuous logical representations. There is a semantic alignment gap between the two. Therefore, instead of directly concatenating features, we use structural features to reconstruct semantic features and semantic features to reconstruct structural features, thereby forcing the model to learn the aligned representation of structure and semantics.

[0204] A dual-constraint matching score calculation method: This innovative method calculates matching scores based on both semantic coverage and requirement satisfaction. Semantic coverage measures the semantic consistency between the deliverable and the overall requirement, while requirement satisfaction measures the degree to which the deliverable meets the structural requirements for each technical point. This method effectively solves the problem of matching score calculation in technology docking scenarios.

[0205] The dynamic recommendation ranking method based on supply and demand potential energy comprehensively considers the matching degree of enterprise technology needs and candidate technologies, as well as the market popularity and maturity of technologies. It innovatively designs a dynamic recommendation ranking score based on supply and demand potential energy. In particular, the maturity gap is designed as an exponential decay term, which can more significantly punish matching results with excessive maturity gaps. This method is more in line with the actual technology transformation needs of enterprises and solves the problem of matching technologies but being difficult to transfer.

[0206] Example 3

[0207] Hypergraph H=(V,E) Figure 2 As shown, the set of hyperedges is E = {e1, e2, e3}, and the set of vertices is V = {v1, v2, v3, v4, v5, v6, v7}. A hyperedge e i Contains a set of vertices v j That is, e i ⊆V, i can be 1, 2, or 3. e1 contains v1 and v2. e2 contains v3, v4, v5, and v2. e3 contains v6, v4, v5, and v7.

[0208] The corresponding correlation matrix is

[0209] .

[0210] Example 4

[0211] AI technology knowledge graph in the form of triples

[0212] (1) Domain level

[0213] (Artificial intelligence, including machine learning)

[0214] (Artificial intelligence, including deep learning)

[0215] (Artificial intelligence, including knowledge graphs)

[0216] (Artificial intelligence includes robotics and perception)

[0217] (2) Machine learning paradigm

[0218] (Machine learning is divided into supervised learning)

[0219] (Machine learning is divided into unsupervised learning)

[0220] (Machine learning, divided into reinforcement learning)

[0221] (Supervised learning, including algorithms and logistic regression)

[0222] (Supervised learning, including algorithms and support vector machines)

[0223] (Supervised learning, including algorithms and decision trees)

[0224] (Unsupervised learning, including algorithms, K-Means clustering)

[0225] (Reinforcement learning, including algorithms, Q-Learning)

[0226] (3) Deep learning architecture

[0227] (Deep learning, with its infrastructure, neural networks)

[0228] (Neural network, variant, Convolutional Neural Network CNN)

[0229] (Neural network, variant, recurrent neural network RNN)

[0230] (Neural network, variant, Transformer)

[0231] (Transformer, model derivation, BERT)

[0232] (Transformer, model derivation, GPT)

[0233] (4) Generative AI

[0234] (Deep learning, with its cutting-edge branches, generative AI)

[0235] (Generative AI, applied technology, diffusion model)

[0236] (Generative AI, applied technology, Generative Adversarial Networks (GANs))

[0237] (Generative AI, applied technology, large language model LLM)

[0238] (5) Knowledge graph technology

[0239] (Knowledge graph, with core technologies, knowledge representation)

[0240] (Knowledge graph, with core technologies, knowledge extraction)

[0241] (Knowledge extraction, with sub-tasks, named entity recognition)

[0242] (Knowledge extraction, with subtasks, relation extraction)

[0243] (Knowledge representation, representation methods, graph embedding)

[0244] (6) Application association

[0245] (Autonomous driving, technology-dependent, object detection)

[0246] (Autonomous driving, technology-dependent, reinforcement learning)

[0247] (Intelligent question answering relies on technology and large language models)

[0248] (Intelligent question answering, technology-dependent, knowledge graph)

[0249] (Recommendation systems, dependency technologies, random forests)

[0250] (AIGC, Dependency Technology, Diffusion Model).

[0251] To better understand the present invention, specific embodiments have been described in detail above, but these are not intended to limit the invention. Any simple modifications made to the above embodiments based on the technical essence of the present invention still fall within the scope of the present invention. Each embodiment in this specification focuses on its differences from other embodiments; similar or identical parts between embodiments can be referred to mutually. For system embodiments, since they basically correspond to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

Claims

1. A technology transfer intelligent docking method based on a hypergraph-knowledge graph fusion network, including data access, characterized in that, It also includes the following steps: Step 1: Extracting structural features based on the hypergraph; Step 2: Calculate semantic association features based on knowledge graph; Step 3: Semantic-guided structural feature enhancement; Step 4: Matching technical needs with deliverables; Step 5: Dynamic recommendation ranking based on supply and demand potential; Step 6: Use historically successful connection cases as positive samples and unrelated technology combinations in the knowledge graph as negative samples to maximize the objective function and obtain the trained model. Step 7: Intelligent docking output.

2. The intelligent technology transfer matching method based on a hypergraph-knowledge graph fusion network as described in claim 1, characterized in that, The access data includes the following sub-steps: Step 01: Access enterprise requirement documents, university / research institute achievement databases, and industry technical dictionaries and standards; Step 02: Each demand exceeds the edge e i Includes a set of technology vertices v j The hypergraph H = (V, E) is composed of the set of entities V in the enterprise technology field and the set of hyperedges E. Step 03: Construct the underlying knowledge graph G=(V,R) from the set of entities V in the enterprise technology field and the semantic relationships R between entities; Where V = {v1, v2, ..., v} j ,...,v n }, E={e1,e2,...,e i ,...,e m }, where n is the number of vertices and m is the number of hyperedges.

3. The intelligent technology transfer matching method based on a hypergraph-knowledge graph fusion network as described in claim 2, characterized in that, Step 1 includes the following sub-steps: Step 11: Propagate vertex information using the hypergraph Laplacian matrix and define the incidence matrix. If vertex v i Belongs to demand e j Then H ij =1, otherwise 0; Step 12: Calculate the degree matrix D of the hyperedge. e The degree matrix D of the vertices v The formula is: ； ； Step 13: Calculate the normalized hypergraph Laplacian matrix L, using the formula: ； The formula for hypergraph convolutional layer propagation is: ； Step 14: After all layers of convolution are completed, the structural feature matrix of all vertices is X. struct ; Among them, H ij H represents the value at position (I,j) in the correlation matrix. ji Let I be the value at position (j,i) in the incidence matrix, where I is the identity matrix. Let l be the feature matrix of the vertices in the l-th layer. σ represents the trainable parameters, and σ is the activation function.

4. The intelligent technology transfer matching method based on a hypergraph-knowledge graph fusion network as described in claim 3, characterized in that, Step 2 includes the following sub-steps: Step 21: Use a relational graph convolutional network to extract vertex features from the knowledge graph G, obtaining the feature h of vertex v. v ; Step 22: Utilize the semantic information of the knowledge graph to semantically associate the features of vertex v with feature u. v The formula for calculating the semantic association features is as follows: ； Step 23: After the calculation is completed, the semantic association feature matrix Usem of all vertices is obtained; in, Let W be the set of vertices adjacent to vertex v under relation r. r Let h be the weight matrix of relation r. s Let be the characteristics of vertex s that is adjacent to vertex v.

5. The intelligent technology transfer matching method based on a hypergraph-knowledge graph fusion network as described in claim 4, characterized in that, The semantically guided structural feature enhancement method includes the following sub-steps: Step 31: Calculate semantic association features U sem For structural feature X Struct The contribution of each dimension is used to generate semantic association features U through linear transformation. sem For structural feature X Struct query vector Q SemStruct and bond vector K Struct The formula is: ； ； Step 32: Calculate the attention weight α for the structural feature channel dimension. Struct The formula is: ； Step 33: Calculate the structural features for semantic association feature enhancement The formula is: ； in, and Let be the learnable weight matrix, ⊙ denotes element-wise multiplication, and d be the feature dimension.

6. The intelligent technology transfer matching method based on a hypergraph-knowledge graph fusion network as described in claim 5, characterized in that, Step 4 involves treating the technology requirement as a query vector and the technology supply as candidate vertices, and calculating the matching degree between the hyperedge and the vertex, including the following sub-steps: Step 41: Let e ​​be the hyperedge corresponding to the technology requirement Q of a certain enterprise. Q It contains a set of key technology vertices v Q Let v be the set of vertices corresponding to the technical achievement T to be matched. T v Q ⊆V, v T ⊆ V; Step 42: Calculate the matching score S(Q,T) based on the dual constraints of semantic coverage and requirement satisfaction. The formula for calculating the matching score S(Q,T) is as follows: ； Where λ is the weight of semantic coverage. This is the overall demand feature vector. As the feature vector of the result, Sim Semantic For the feature vector of the result With overall demand feature vector cosine similarity, , Sim struct For the vertex set v T For the superedge e Q Structural satisfaction.

7. The intelligent technology transfer matching method based on a hypergraph-knowledge graph fusion network as described in claim 6, characterized in that, The method for calculating the semantic coverage includes: Demand beyond the edge e Q Features are aggregated into a single vector : ； Calculation result feature vector With demand feature vector cosine similarity Sim Semantic The formula is: ； MeanPooling is the average pooling operation, Z i This is a feature of fusion.

8. The intelligent technology transfer matching method based on a hypergraph-knowledge graph fusion network as described in claim 7, characterized in that, The method for calculating the degree of demand satisfaction includes: According to the vertex set v T Is it related to e? Q The vertices in the graph are closely connected in the knowledge graph G. Calculate v T For e Q Structural Satisfaction Sim struct The formula is: ； ； Among them, f rel (v t , v x Let γ be the correlation function between two vertices in the knowledge graph, and let γ be the decay function dist(v). t , v x ) for v t and v x The path distance in the graph.

9. The intelligent technology transfer matching method based on a hypergraph-knowledge graph fusion network as described in claim 8, characterized in that, Step 5 includes introducing market dynamic factors, calculating the final recommendation score Rec(Q,T), and sorting the data according to the recommendation score. The formula for calculating the final recommendation score Rec(Q,T) is as follows: ； Where S(Q,T) is the matching score, and H T Let β be the current market popularity of the technological achievement T, β be the adjustment coefficient, and Maturity be the technological maturity. Maturity is a maturity penalty term. Q -Maturity T |This represents the maturity gap between enterprise needs (Q) and outcomes (T).

10. The intelligent technology transfer matching method based on a hypergraph-knowledge graph fusion network as described in claim 9, characterized in that, The total loss function L of the training model total The calculation formula is: L total =Rec(Q,T)。

Images