A Deep Learning-Based Intelligent Scheduling Method and System for Technology Incubation and R&D Resources
By constructing a competitive perception triplet resource graph and a competitive diffusion graph attention network, the problem of low resource allocation efficiency in technology incubation is solved, global optimal resource matching is achieved, and the adaptability and efficiency of resource scheduling are improved.
Patent Information
- Application Number
- CN202611148854.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-31
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies fail to effectively model the competitive relationship among multiple enterprises for the same scarce resource in technology incubation, resulting in low resource allocation efficiency and difficulty in achieving globally optimal allocation.
A deep learning-based approach is used to construct a competition-aware triplet resource graph. Through a competition diffusion graph attention network and a multi-relationship attention mechanism, the resource competition matching features are quantified. The global resource allocation matrix is solved through multi-objective iterative optimization, and a global resource scheduling scheme is output.
It improves the adaptability and overall efficiency of resource allocation for technology incubation and R&D, enabling the global optimal resource matching in complex competitive scenarios and providing feasible decision support.
Smart Images

Figure CN123117593A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of innovation and entrepreneurship incubation service technology, and in particular relates to a method and system for intelligent scheduling of scientific and technological incubation and R&D resources based on deep learning. Background Technology
[0002] Technology incubation is a crucial link in transforming scientific and technological innovation achievements into industrialization, playing a vital role in the national innovation system by nurturing technology-based enterprises, commercializing research results, and building an innovation ecosystem. The efficiency of incubation resource allocation directly determines the growth rate and success rate of science and technology enterprises. Currently, the allocation of incubation resources mainly relies on simple information matching and resource recommendations based on human experience. This model works when the incubation scale is small and resource supply is sufficient, but with the rapid expansion of technology incubation, its limitations are becoming increasingly apparent. Especially within the incubation ecosystem, high-quality incubation resources (such as renowned industry mentors, large-scale experimental and testing equipment, and special funding and policy application quotas) remain scarce. When multiple science and technology enterprises simultaneously demand the same scarce resource, existing technologies struggle to achieve optimal global allocation.
[0003] However, current technologies employ an independent matching paradigm, failing to model the competitive relationships among multiple enterprises vying for the same scarce resource. For instance, the intelligent matching method for business incubators proposed in patent CN113538115B and the innovation and entrepreneurship incubation resource matching system proposed in patent CN121278095A, while indirectly considering resource scarcity by dynamically adjusting matching thresholds, essentially still calculate the matching degree independently for each demand and then recommend it based on score ranking. They fail to model the conflict resolution mechanism and global optimization allocation strategy when multiple enterprises simultaneously compete for the same scarce resource, resulting in low global configuration efficiency. Summary of the Invention
[0004] Therefore, it is necessary to provide a method and system for intelligent scheduling of science and technology incubation and R&D resources based on deep learning to address the above-mentioned technical problems, which can improve the allocation efficiency of science and technology incubation and R&D resources.
[0005] Firstly, this application provides a method for intelligent scheduling of technology incubation and R&D resources based on deep learning, including:
[0006] S101. Perform data structuring processing on the acquired multi-source heterogeneous data to obtain enterprise demand feature vectors and resource attribute feature vectors; wherein, the multi-source heterogeneous data includes multi-dimensional attribute data of resource demanders and attribute description data of resource suppliers.
[0007] S102. Construct a competition-aware triplet resource graph based on the enterprise demand feature vector and resource attribute feature vector. The competition-aware triplet resource graph includes three types of nodes: enterprise nodes, resource nodes, and competition relationship representation nodes. Enterprise nodes and resource nodes are connected by demand edges, nodes of the same type are connected by association edges, and multiple enterprise nodes are connected to the same competition relationship representation node by competition edges.
[0008] S103. Input the competitive perception triplet resource graph into the competitive diffusion graph attention network, and calculate the attention weights for different types of edges through the type-aware multi-relation attention mechanism to obtain the attention coefficients of each edge.
[0009] S104. Based on the attention coefficients of each edge, feature propagation is carried out along the competing edges through a competitive diffusion message passing mechanism. During the message passing process, a competitive diffusion inhibition factor is introduced to control the attenuation of the feature propagation intensity through the nodes representing the competitive relationship, thereby obtaining the node feature representation.
[0010] S105. Based on the node characteristic representation, calculate the competition quantification index of each resource corresponding to each resource node; wherein, the competition quantification index characterizes the urgency of supply and demand matching of each resource under different enterprise node competition environments;
[0011] S106. Construct an initial global allocation matrix based on node feature representation; wherein, the initial global allocation matrix is a two-dimensional matrix with enterprises as rows and resources as columns, and the elements in the initial global allocation matrix represent the initial allocation tendency between the corresponding enterprise and the corresponding resource;
[0012] S107. Construct a multi-objective allocation optimization objective based on the competition quantification index, and perform differentiable iterative optimization on the initial global allocation matrix to obtain the optimized global resource allocation matrix.
[0013] S108. Based on the optimized global resource allocation matrix, output the global resource scheduling scheme.
[0014] In one embodiment, a competition-aware triplet resource map is constructed based on the enterprise demand feature vector and the resource attribute feature vector, including:
[0015] S201. Based on the similarity between the enterprise demand feature vector and the resource attribute feature vector, determine the weight of the demand edge between the enterprise node and the resource node to obtain the enterprise-resource matching relationship with demand edge weights.
[0016] S202. Based on the semantic similarity between enterprise demand feature vectors, establish association edges between different enterprise nodes; based on the functional similarity between resource attribute feature vectors, establish association edges between different resource nodes to obtain a node relationship topology with similar association edges.
[0017] S203. Based on the enterprise-resource matching relationship with demand edge weights, when multiple enterprise nodes have a demand for the same resource node, a corresponding competition relationship representation node is created for the resource node, resulting in the competition relationship representation node creation result. Based on the competition relationship representation node creation result, each participating enterprise node is connected to the competition relationship representation node through competition edges, resulting in a competition relationship topology with competition edges.
[0018] S204. Generate the initial feature vector of the node representing the competitive relationship based on the weighted average of the resource attribute feature vector and the enterprise demand feature vector of each participating enterprise node.
[0019] S205. Based on the enterprise-resource matching relationship with demand edge weights, the node relationship topology with similar related edges, and the competitive relationship topology with competitive edges, the graph structure is integrated to obtain the competition-aware triplet resource graph.
[0020] In one embodiment, a competition-aware triplet resource graph is input into a competition-diffusion graph attention network. Attention weights are calculated for different types of edges using a type-aware multi-relation attention mechanism to obtain the attention coefficients for each edge, including:
[0021] S301. Configure independent attention parameter vectors for the three edge types: demand edge, associated edge, and competing edge.
[0022] S302. For each edge in the competitive sensing triplet resource graph, select the corresponding attention parameter vector according to the edge type, and calculate the attention coefficient of the edge by performing the inner product operation between the attention parameter vector and the node feature vector; the formula for calculating the attention coefficient is:
[0023] ;
[0024] in, The superscript represents the attention coefficient between node u and node v in layer l. This represents the vector transpose operation. Indicates the relationship with the edge edge type The corresponding attention parameter vector, Indicates the first The linear transformation matrix of the layer, and Representing nodes respectively and nodes In the The feature vector of the layer, This represents a vector concatenation operation, where LeakyReLU is the activation function.
[0025] S303. Traverse all edges in the competitive perception triplet resource graph, match each edge with the attention parameter vector corresponding to the edge type and calculate the attention coefficient to obtain the attention coefficient of each edge containing all edges.
[0026] In one embodiment, based on the attention coefficients of each edge, features are propagated along the competing edges through a competitive diffusion message passing mechanism. During message passing, a competitive diffusion suppression factor is introduced to attenuate the feature propagation intensity passing through nodes representing competing relationships, resulting in node feature representations, including:
[0027] S401. Aggregate the features of each enterprise node connected to the competitive relationship node after attention weighting to obtain the updated feature representation of the competitive relationship node;
[0028] S402. Determine the value of the competitive diffusion inhibition factor based on the number of enterprise nodes connected to the same competitive relationship node; wherein, the more enterprise nodes there are, the larger the value of the competitive diffusion inhibition factor.
[0029] S403. When updating the features of enterprise nodes through competing edges, based on the competitive relationship representing the updated feature representation of the node and the competitive diffusion suppression factor, multiplicative suppression is applied to the features transmitted through the competing edges to obtain the feature representation of the enterprise node after competitive diffusion suppression; the calculation formula for the feature representation of the enterprise node after competitive diffusion suppression is as follows:
[0030] ;
[0031] in, This represents the feature component that enterprise node u obtains through the competing edge at layer l+1. Let represent the attention coefficient between enterprise node u and node c, indicating their competitive relationship. For the learnable linear transformation matrix on the enterprise node side, For enterprise nodes In the The feature vector of the layer, Representing nodes in a competitive relationship In the The feature vector of the layer, σ is a competitive diffusion inhibition factor, and σ is a nonlinear activation function;
[0032] S404. Based on the feature representation of the enterprise node after competitive diffusion suppression, integrate the features of each node transmitted through non-competitive edges and the features transmitted through competitive edges to obtain the node feature representation.
[0033] In one embodiment, based on node feature representation, the competition quantification index for each resource corresponding to each resource node is calculated, including:
[0034] S501. Based on the feature representations of each enterprise node connected to the competitive relationship representation node, calculate the competitiveness probability distribution of each enterprise node in the competition corresponding to the competitive relationship representation node; wherein, the competitiveness probability distribution is obtained by calculating the similarity between the feature representation of the enterprise node and the feature representation of the competitive relationship representation node and then normalizing it.
[0035] S502. Calculate the quantitative indicators of competition based on the probability distribution of competitiveness; the formula for calculating the quantitative indicators of competition is:
[0036] ;
[0037] in, Representing competitive relationships, nodes are represented. Corresponding quantitative indicators of competition This indicates a node connected to a competing node. The number of enterprise nodes, This indicates that node j, representing enterprise node, is in a competitive relationship with node j. The corresponding probability of competitiveness in the competition.
[0038] Secondly, this application also provides a deep learning-based intelligent scheduling system for technology incubation and R&D resources, used to implement the aforementioned deep learning-based intelligent scheduling method for technology incubation and R&D resources. The system includes:
[0039] The data structuring module is used to perform data structuring processing on the acquired multi-source heterogeneous data to obtain enterprise demand feature vectors and resource attribute feature vectors; among which, the multi-source heterogeneous data includes multi-dimensional attribute data of resource demanders and attribute description data of resource suppliers.
[0040] The graph construction module is used to construct a competition-aware triplet resource graph based on enterprise demand feature vectors and resource attribute feature vectors. The competition-aware triplet resource graph includes three types of nodes: enterprise nodes, resource nodes, and competition relationship representation nodes. Enterprise nodes and resource nodes are connected by demand edges, nodes of the same type are connected by association edges, and multiple enterprise nodes are connected to the same competition relationship representation node by competition edges.
[0041] The multi-relation attention module is used to input the competition-aware triplet resource graph into the competition diffusion graph attention network. It calculates the attention weights for different types of edges through a type-aware multi-relation attention mechanism to obtain the attention coefficients of each edge.
[0042] The competitive diffusion feature propagation module is used to propagate features along the competitive edges based on the attention coefficients of each edge through a competitive diffusion message passing mechanism. During the message passing process, a competitive diffusion inhibition factor is introduced to control the attenuation of the feature propagation intensity through the nodes representing the competitive relationships, thereby obtaining the node feature representation.
[0043] The competition quantification index calculation module is used to calculate the competition quantification index of each resource corresponding to each resource node based on the node characteristic representation; among which, the competition quantification index represents the urgency of supply and demand matching of each resource in different enterprise node competition environments;
[0044] The initial allocation matrix construction module is used to construct an initial global allocation matrix based on node feature representations. The initial global allocation matrix is a two-dimensional matrix with enterprises as rows and resources as columns. The elements in the initial global allocation matrix represent the initial allocation tendency between the corresponding enterprise and the corresponding resource.
[0045] The multi-objective optimization solution module is used to construct multi-objective allocation optimization objectives based on competition quantification indicators, and to perform differentiable iterative optimization on the initial global allocation matrix to obtain the optimized global resource allocation matrix.
[0046] The scheduling scheme output module is used to output a global resource scheduling scheme based on the optimized global resource allocation matrix.
[0047] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the aforementioned method for intelligent scheduling of technology incubation and R&D resources based on deep learning.
[0048] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the aforementioned method for intelligent scheduling of technology incubation and R&D resources based on deep learning.
[0049] The aforementioned intelligent scheduling method and system for R&D resources in technology incubation based on deep learning extracts structured features from heterogeneous supply and demand data to construct a competitive perception triplet resource graph that integrates supply and demand relationships with enterprise competition. It then relies on a graph deep learning model equipped with a multi-relationship attention mechanism and a competition diffusion suppression and control mechanism to quantify resource competition matching features and allocation tendencies. Through multi-objective iterative optimization, it completes the global resource allocation solution. This technical solution overcomes the problems of R&D resource scheduling in technology incubation failing to effectively characterize complex competitive relationships and achieving globally optimal resource matching and scheduling under competitive perception, thus improving the adaptability and overall efficiency of global R&D resource scheduling in complex competitive scenarios. Attached Figure Description
[0050] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0051] Figure 1 A flowchart illustrating a deep learning-based intelligent scheduling method for technology incubation and R&D resources is provided in an embodiment of the present invention.
[0052] Figure 2 This is a schematic diagram of the structure of a deep learning-based intelligent scheduling system for scientific and technological incubation and R&D resources, provided as an embodiment of the present invention. Detailed Implementation
[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0054] First, a brief introduction to the terms used in the embodiments of this application will be given.
[0055] The Competition Diffusion Graph Attention Network (CD-GAT) is a deep graph neural network model proposed in this invention. It introduces a type-aware multi-relationship attention mechanism and a competitive diffusion suppression mechanism on the basis of graph attention network, so as to realize explicit modeling of competitive relationships and feature propagation control in heterogeneous graph structures.
[0056] The Competition-Aware Triple Resource Graph is a heterogeneous graph data structure constructed in this invention. It consists of three types of nodes: enterprise nodes, resource nodes, and nodes representing competition relationships, as well as three types of edges: demand edges, association edges, and competition edges. It explicitly transforms the competition relationships of multiple entities for the same resource into a computable graph topology.
[0057] The global resource scheduling scheme is the final output of the method in this application. It refers to the set of resource allocation decisions obtained through global optimization calculation after comprehensively considering the resource needs of all active enterprises, the supply capacity of all available resources, and the competitive state faced by each resource. Unlike the serial recommendation results based on independent matching in existing technologies, the global resource scheduling scheme examines the allocation relationship between all enterprises and all resources from a global perspective after the global allocation matrix optimization solution is completed. This ensures that the allocation decision of each scarce resource is made under the premise of considering all competitive needs, rather than performing a local optimal match for a single enterprise in isolation. This scheme not only includes the final resource list obtained by each enterprise, but also implies the global optimization basis behind the allocation decision, providing implementable and traceable decision support for the resource allocation execution of incubators.
[0058] Based on the above definitions, the implementation environment of the reinforcement learning-based intelligent energy consumption optimization control method for environmental protection equipment provided in this application embodiment will be described. Indicatively, this implementation environment is the internal local area network environment of a technology incubation platform. At the hardware level, the system consists of at least one GPU-accelerated server, one application server, and a database cluster, interconnected via a gigabit Ethernet switch. The GPU-accelerated server deploys NVIDIA GPU computing cards to handle offline training and online inference computation tasks for the CD-GAT model. The application server is responsible for running business logic, receiving scheduling task requests through a Web service interface, and displaying scheduling schemes to the administrator. The database cluster can adopt a hybrid architecture of relational database and graph database. The relational database stores enterprise basic information tables, resource information tables, and allocation record tables, while the graph database stores node and edge data of the competition-aware triplet resource graph.
[0059] In terms of data interaction, each business system may include an incubation management system, an investment and financing service platform, an equipment management platform, and a policy database. These systems will aggregate multi-dimensional attribute data of incubated enterprises, such as team size, technology field, development stage, funding needs, and / or historical resource usage records, as well as attribute description data of suppliers such as mentor resources, investment and financing resources, equipment resources, and / or policy resources, to the data access layer of the application server through RESTful API interfaces.
[0060] The daily application scenario of this method within the technology incubation platform is as follows: the administrator initiates a resource scheduling task request through the web interface, the application server triggers the data acquisition module to pull the latest data from various business systems and perform structured processing, and then performs graph construction, CD-GAT feature inference, competitive quantification index calculation, initial allocation matrix construction and differentiable iterative optimization solution, and finally displays the global resource scheduling scheme on the web interface, including the specific resource list allocated to each enterprise, allocation priority ranking and supplementary suggestions for unmet needs.
[0061] For example, the intelligent scheduling method for technology incubation and R&D resources based on deep learning provided in this application embodiment can also be applied to other application scenarios. This is only an example and does not limit the specific application scenarios.
[0062] like Figure 1 As shown, Embodiment 1 provides a method for intelligent scheduling of R&D resources for technology incubation based on deep learning. This embodiment illustrates the application of this method to a technology incubation platform in the aforementioned implementation environment. It is understood that this method can also be applied to servers, and to systems including terminals and servers, and is implemented through interaction between the terminals and servers. In this embodiment, the method includes the following steps 101 to 108:
[0063] S101. Perform data structuring processing on the acquired multi-source heterogeneous data to obtain enterprise demand feature vectors and resource attribute feature vectors; wherein, the multi-source heterogeneous data includes multi-dimensional attribute data of resource demanders and attribute description data of resource suppliers.
[0064] Specifically, multi-source heterogeneous data includes two types of data sources: The first type is multi-dimensional attribute data from resource demanders, covering textual descriptive fields and numerical indicator fields such as the team size, technology field tags, current development stage, funding needs, historical resource usage records, and technical directions of ongoing projects of incubated enterprises; the second type is attribute descriptive data from resource suppliers, covering fields such as the professional field and available guidance duration of mentors, investment amount and preferred industries of investment and financing institutions, technical specifications and available machine time of equipment, funding conditions and application deadlines of policy resources.
[0065] The data structuring process consists of two parallel pathways: a text-based data pathway and a numerical data pathway. The text-based data pathway combines the text from each field into a unified text sequence, extracts a 768-dimensional semantic feature vector using a pre-trained language model, and then reduces the dimensionality to 128 dimensions using a fully connected layer. The numerical data pathway performs Min-Max normalization on each numerical field, linearly mapping the values to the [0,1] interval, and then maps the normalized features to 128 dimensions using a learnable linear mapping matrix. The 128-dimensional feature vectors output by the two pathways are summed element-wise to obtain a unified feature vector that integrates semantic information and numerical features.
[0066] Let the total number of incubated enterprises be Then, after the above processing, we get Enterprise demand feature vector The dimension of each enterprise's demand feature vector is Let the total number of resources be... Then we get Resource attribute feature vectors The dimension of each resource attribute feature vector is also 1. .
[0067] S102. Construct a competition-aware triplet resource graph based on the enterprise demand feature vector and resource attribute feature vector. The competition-aware triplet resource graph includes three types of nodes: enterprise nodes, resource nodes, and competition relationship representation nodes. Enterprise nodes and resource nodes are connected by demand edges, nodes of the same type are connected by association edges, and multiple enterprise nodes are connected to the same competition relationship representation node by competition edges.
[0068] Specifically, the node set of the graph contains three types of nodes: enterprise nodes modeled after the various incubated companies, totaling [number missing]. There are [number] resource nodes, each with an initial feature vector obtained from S101 corresponding to the enterprise's demand feature vector; and resource nodes modeled after various R&D resources, totaling [number]. There are 1 resource node, where the initial feature vector of each resource node is the corresponding resource attribute feature vector obtained in S101; and a competition relationship representation node, used to explicitly model the competition relationship between multiple enterprises for the same resource, the number of which is equal to the number of resources simultaneously demanded by two or more enterprises, denoted as . The edge set contains three types of edges: demand edges connecting enterprise nodes and resource nodes, association edges connecting different enterprise nodes or different resource nodes, and competition edges connecting enterprise nodes and nodes representing competition relationships.
[0069] The rule for establishing the demand side is as follows: for enterprise nodes... and resource nodes Calculate the cosine similarity of their initial feature vectors. When the normalized similarity value is greater than or equal to a preset threshold, in and Demand edges are established between different enterprise nodes, with edge weights taken from the aforementioned cosine similarity value, thus forming an enterprise-resource matching relationship; the rule for establishing related edges is: for different enterprise nodes... and Calculate the cosine similarity of the enterprise demand feature vectors, and establish inter-enterprise association edges when the similarity is greater than or equal to a preset threshold; for different resource nodes and Calculate the Jaccard similarity of the set of functional labels in the feature vector of its resource attributes, and establish the connection edge between resources when the similarity is greater than or equal to the preset threshold.
[0070] For example, the creation rule for competitive relationship nodes is as follows: traverse all resource nodes, and for resource nodes that are connected by two or more enterprise nodes through demand edges. ,for Create a corresponding node representing the competition relationship. ; The initial feature vector is obtained by weighting the feature vectors of the enterprise demand of participating enterprises and then summing them with the feature vectors of resource attributes, i.e. ,in For resources The collection of all enterprise nodes that have demand. For enterprise nodes With resource nodes The demand edge weights between them For resource nodes Resource attribute feature vector, For enterprise nodes The enterprise demand feature vector; the rule for establishing competitive edges is: for nodes representing competitive relationships... ,Will All enterprise nodes in the middle compete with each other through edge competition. connect.
[0071] For example, after the graph is constructed, a competition-aware triplet resource graph is obtained. , where the set of nodes In the formula, for A set of enterprise nodes for A set of resource nodes for Each competitive relationship represents a set of nodes, and the total number of nodes is [number missing]. edge set , For the set of demand edges, For the set of associated edges, For the set of competing edges; This is the initial eigenvector matrix for all nodes.
[0072] S103. Input the competitive perception triplet resource graph into the competitive diffusion graph attention network, and calculate the attention weights for different types of edges through the type-aware multi-relation attention mechanism to obtain the attention coefficients of each edge.
[0073] Specifically, the CD-GAT network is stacked. Layers, each containing a shared set of linear transformation matrices. ,in The feature dimension is defined as follows. The type-aware multi-relation attention mechanism uses a unified attention parameter space for the three edge types, but distinguishes them by one-hot encoding of the edge types: the edge type is encoded as a three-dimensional one-hot vector, corresponding to demand edge [1,0,0], associated edge [0,1,0] and competitive edge [0,0,1], respectively.
[0074] In the In the layer, the edge Attention coefficient The calculation is performed using a type-aware attention mechanism, and the formula is as follows:
[0075]
[0076] in, Indicates the first Layer nodes and nodes Attention coefficient between, superscript This represents the vector transpose operation; Indicates the relationship with the edge edge type The corresponding attention parameter vector has a dimension of In this embodiment, It is obtained by element-wise multiplication of the basic parameter vector shared by the three edge types and the one-hot encoding vector of the edge type, i.e. ; Indicates the first The linear transformation matrix of the layer has dimension 1. ; and Representing nodes respectively and nodes In the The feature vectors of the layer have a dimension of 1. ; This represents a vector concatenation operation, resulting in a vector with dimension . LeakyReLU is a linear unit activation function with leakage correction, and its negative slope is set to 0.2.
[0077] The calculated original attention coefficients are then Softmax normalized so that the sum of the attention coefficients of all neighbor edges of each node is 1.
[0078] S104. Based on the attention coefficients of each edge, feature propagation is carried out along the competing edges through a competitive diffusion message passing mechanism. During the message passing process, a competitive diffusion inhibition factor is introduced to control the attenuation of the feature propagation intensity through the nodes representing the competitive relationship, thereby obtaining the node feature representation.
[0079] Specifically, the message passing process is divided into two phases: non-contested edge propagation and contested edge propagation. In the non-contested edge propagation phase of each layer, standard graph attention message aggregation is performed along the demand edge and the associated edge: for enterprise nodes and resource nodes, the features of all their neighboring nodes are aggregated using the attention coefficient as weight to obtain the first... The intermediate features of the layer are represented.
[0080] The competition edge propagation phase performs feature propagation along the competition edge: first, it updates the competition relationship representation node. In the Layer feature representation The features of all connected enterprise nodes are aggregated and weighted by attention; then, the features are backpropagated to the enterprise nodes through competitive edges, and a competitive diffusion inhibition factor is introduced. Attenuation control is applied to the characteristic propagation intensity. In this embodiment, a competitive diffusion inhibition factor is used. Use fixed values It does not adaptively adjust to the scale of competition. Enterprise nodes The formula for calculating the feature components obtained through competing edges is:
[0081]
[0082] in, Represents enterprise nodes In the The layer obtains feature components through competing edges; For enterprise nodes The relationship between nodes represents competition. Attention coefficient between them; The learnable linear transformation matrix on the enterprise node side has a dimension of ; For enterprise nodes In the The feature vector of the layer; Representing nodes in a competitive relationship In the The feature vector of the layer; A fixed competitive diffusion inhibitor; This is the ELU nonlinear activation function.
[0083] The above multiplication suppression term The physical meaning of reducing the feature amplitude transmitted from the competitive relationship representation node to the enterprise node during the competitive edge feature propagation process is as follows: when multiple incubated enterprises have competitive demand for the same resource, the competitive tension signal carried by the competitive relationship representation node corresponding to the resource should not be propagated back to each participating enterprise node at full intensity, but should be appropriately attenuated based on the intensity of competition, so that the feature representation of each enterprise node can reflect its relative position in the overall competition rather than absolute competitive pressure.
[0084] go through After message passing through stacked layers, the node feature representations output by each layer are concatenated to obtain the final node feature representation: for enterprise nodes Its final node features are represented as , dimension ,in The first layer is the initial feature vector; the final feature representations of resource nodes and competition relationship nodes are constructed using the same concatenation method. The final 384-dimensional vector of node feature representation integrates the semantic information and competition environment information of each node within a multi-scale neighborhood.
[0085] S105. Based on the node characteristic representation, calculate the competition quantification index of each resource corresponding to each resource node; wherein, the competition quantification index characterizes the urgency of supply and demand matching of each resource under different enterprise node competition environments;
[0086] Specifically, after obtaining the final 384-dimensional feature representation of all nodes, the competition quantification index corresponding to each resource node is calculated. The competition quantification index is used to quantify the urgency of supply and demand matching for each resource in a multi-enterprise competitive environment. For nodes representing competitive relationships... Its corresponding resources were Competition among companies, among which .
[0087] First, calculate the probability distribution of competitiveness among the participating firms: represented by the final characteristics of each firm node. The competitive relationship represents the final feature representation of the node. Calculate the scaled dot product similarity, and after Softmax normalization, obtain the competitiveness probability of each enterprise in the competition. Based on this, the variance-based competitive dispersion is used as the quantitative indicator of competition. The measurement is performed using the following formula:
[0088]
[0089] in, Representing competitive relationships, nodes are represented. Corresponding quantitative indicators of competition; To connect to the node representing the competition relationship The number of enterprise nodes; For enterprise nodes The final feature representation The relationship between nodes represents competition. The final feature representation The scaled dot product similarity between them, divided by the square root of the feature dimension. Keep the dot product value within a reasonable range to prevent the Softmax gradient from vanishing; For each enterprise node and The average similarity between them.
[0090] The physical meaning is the degree of dispersion of the feature similarity of competing firms in the vector space: The larger the value, the more uneven the distribution of competitiveness among participating companies, and the greater the difference in the urgency of supply and demand matching. The smaller the value, the more evenly distributed the strength among competitors. For resources needed by only a single enterprise, there are no corresponding competitive nodes, and the competition quantification index is 0, indicating that there is no competitive pressure for the resource and the urgency of supply and demand matching is the lowest.
[0091] S106. Construct an initial global allocation matrix based on node feature representation; wherein, the initial global allocation matrix is a two-dimensional matrix with enterprises as rows and resources as columns, and the elements in the initial global allocation matrix represent the initial allocation tendency between the corresponding enterprise and the corresponding resource;
[0092] Specifically, based on the 384-dimensional final feature representations of all enterprise nodes and all resource nodes obtained in S104, an initial global allocation matrix is constructed. Let there be a total of... family businesses and Resource, initial global allocation matrix Given a two-dimensional matrix with enterprises as rows and resources as columns, the first... Line 1 Column elements Characterizing the first Family Enterprises and the First The initial allocation tendency among the resources, taking values in the range [0,1].
[0093] Through the first The final characteristic representation of a home enterprise node With the The final characteristic representation of a resource node The cosine similarity is calculated as follows: The matrix elements are Softmax normalized row by row so that the sum of the elements in each row is 1, thus interpreting the elements in each row as the initial allocation probability of the various resources corresponding to the enterprise.
[0094] The constructed initial global allocation matrix reflects the initial allocation tendency between enterprises and resources based on deep feature matching, without considering competition constraints and resource capacity constraints.
[0095] S107. Construct a multi-objective allocation optimization objective based on the competition quantification index, and perform differentiable iterative optimization on the initial global allocation matrix to obtain the optimized global resource allocation matrix.
[0096] Specifically, based on the competition quantification index corresponding to each competitive relationship node obtained in S105, a multi-objective allocation optimization objective is constructed, and the initial global allocation matrix obtained in S106 is solved by differentiable iterative optimization.
[0097] For example, Gumbel-Softmax reparameterization sampling transforms discrete allocation decisions into differentiable operations, including: applying the initial global allocation matrix... Inject compliance into each line Distributed independent noise During forward propagation, the first temperature parameter Calculate the sampling results ,in, These are the elements of the allocation matrix after Gumbel-Softmax sampling; For the initial global allocation matrix, the first... Line 1 The initial allocation tendency value of the column; For injected Gumbel distribution noise; Forward propagation temperature parameters; summation symbol subscript To traverse the current enterprise The corresponding dummy variable for all resource indexes is used to iterate through all resource columns in this row for the enterprise. The exponential terms corresponding to all resources are summed to complete Softmax normalization. This makes the sampling results approximate the discrete one-hot distribution form; in backpropagation, the second temperature parameter is used. The gradient signal is calculated to make gradient propagation smoother and avoid gradient vanishing. The above decoupling temperature strategy uses a smaller temperature parameter for forward propagation. To make the output approximate discrete decision-making, backpropagation uses a larger temperature parameter. The gradient signal is smoothed to achieve a coordinated and unified approach between forward discrete decision-making and backward gradient smoothing.
[0098] Optionally, the multi-objective allocation optimization objective function consists of a weighted combination of three sub-objectives: the first term is the global matching benefit sub-objective. The first term is defined as the sum of the element-wise products of each element in the differentiable sampling allocation matrix and the enterprise-resource matching score at the corresponding position of that element; the second term is the resource utilization sub-objective. The first term is defined as the concentration penalty term for the allocation of each resource, measured by the sum of the differences between the maximum allocation value and the mean value for each resource column; the third term is the competitive equilibrium sub-objective. It is related to the competition quantification index calculated in S105.
[0099] Alternatively, the optimization problem can also introduce capacity constraints for scarce resources: a logarithmic barrier function is used to penalize the total allocation of each scarce resource, ensuring that the total allocation does not exceed the upper limit of the allocable capacity of that resource. The coefficient of the barrier term is set to... .
[0100] For example, the optimization solution can use the Adam optimizer with a learning rate of 0.001, iterating for 200 rounds to decrease the value of the global allocation objective function round by round until convergence, thus obtaining the optimized global resource allocation matrix. .
[0101] S108. Based on the optimized global resource allocation matrix, output the global resource scheduling scheme;
[0102] Specifically, based on the optimized global resource allocation matrix obtained in S107 For each row of the matrix, the resource corresponding to the column containing the maximum value is taken as the allocation result for the enterprise, and the final global resource scheduling scheme is output.
[0103] For example, the scheduling scheme may include a list of specific R&D resources allocated to each incubated enterprise, a resource matching score for each allocation, and allocation priority ranking information. After outputting the allocation results for each enterprise, the scheme may also independently mark the resources in the optimized allocation matrix that have not been allocated to any enterprise, and generate a supplementary suggestion list of unmet needs.
[0104] The aforementioned intelligent scheduling method for R&D resources in technology incubation based on deep learning extracts structured features from heterogeneous supply and demand data to construct a competitive perception triplet resource graph that integrates supply and demand relationships with enterprise competition. It then relies on a graph deep learning model equipped with a multi-relationship attention mechanism and a competition diffusion suppression and control mechanism to quantify resource competition matching features and allocation tendencies. Through multi-objective iterative optimization, it completes the global resource allocation solution. This technical solution overcomes the problems of R&D resource scheduling in technology incubation failing to effectively characterize complex competitive relationships and achieving globally optimal resource matching and scheduling under competitive perception, thus improving the adaptability and overall efficiency of global R&D resource scheduling in complex competitive scenarios.
[0105] Example 2: Based on the enterprise demand feature vector and resource attribute feature vector, a competition-aware triplet resource map is constructed, including:
[0106] S201. Based on the similarity between the enterprise demand feature vector and the resource attribute feature vector, determine the weight of the demand edge between the enterprise node and the resource node to obtain the enterprise-resource matching relationship with demand edge weights.
[0107] S202. Based on the semantic similarity between enterprise demand feature vectors, establish association edges between different enterprise nodes; based on the functional similarity between resource attribute feature vectors, establish association edges between different resource nodes to obtain a node relationship topology with similar association edges.
[0108] S203. Based on the enterprise-resource matching relationship with demand edge weights, when multiple enterprise nodes have a demand for the same resource node, a corresponding competition relationship representation node is created for the resource node, resulting in the competition relationship representation node creation result. Based on the competition relationship representation node creation result, each participating enterprise node is connected to the competition relationship representation node through competition edges, resulting in a competition relationship topology with competition edges.
[0109] S204. Generate the initial feature vector of the node representing the competitive relationship based on the weighted average of the resource attribute feature vector and the enterprise demand feature vector of each participating enterprise node.
[0110] S205. Based on the enterprise-resource matching relationship with demand edge weights, the node relationship topology with similar related edges, and the competitive relationship topology with competitive edges, the graph structure is integrated to obtain the competition-aware triplet resource graph.
[0111] Specifically, this embodiment decomposes the graph construction process in S102 of embodiment 1 into five refined sub-steps. By separately constructing and finally integrating the initial features of the nodes representing demand edges, associated edges, competing edges, and competing relationships, it achieves fine control over the topological quality of each part of the competition-aware triplet resource graph.
[0112] In S201, for the first Enterprise demand feature vector corresponding to each enterprise node and the Resource attribute feature vector of the resource node corresponding to the resource item Weight of the demand side Calculate using the cosine similarity between the two: ,in Represents enterprise nodes With resource nodes The demand edge weights between them For enterprise demand feature vectors, For resource attribute feature vectors, and They are respectively the corresponding vectors Norm. When When the threshold is greater than or equal to the preset threshold, at the enterprise node and resource nodes Establish a demand edge between them, and Assign a weight to this edge to obtain the enterprise-resource matching relationship structure with demand edge weights.
[0113] In S202, inter-firm relationships are established by calculating the cosine similarity of the feature vectors of different firm needs: for firm nodes... and ,like If the value is greater than or equal to a preset threshold, an association edge is established between the two; the association edge between resources is established by calculating the Jaccard similarity of the resource function label set: for resource nodes and If the Jaccard similarity of their functional label sets is greater than or equal to a preset threshold, then an association edge is established between the two, thereby obtaining a node relationship topology with similar association edges.
[0114] In S203, all resource nodes are traversed. For resource nodes that meet the condition of being connected by two or more enterprise nodes through demand edges, a competition relationship representation node is created for them. The total number of competition relationship representation nodes is equal to the number of resources covered by the demand states of multiple enterprises. Then, each enterprise node participating in the competition is connected to the corresponding competition relationship representation node through competition edges to form a competition relationship topology with competition edges.
[0115] In S204, competition relationships represent nodes. initial feature vector It is composed of two parts: the weighted average of the corresponding resource attribute feature vector and the enterprise demand feature vector of each participating enterprise. ,in Representing competitive relationships, nodes are represented. The initial feature vector, for Corresponding resource nodes Resource attribute feature vector, For resources A set of enterprise nodes that have demand. This represents the number of enterprise nodes in the set. For enterprise nodes With resource nodes The demand edge weights between them For enterprise nodes The initialization of the enterprise demand feature vector is as follows: the initial features of the competitive relationship representation node include both the attribute information of the resource being competed for and the weighted average of the demand features of each participating enterprise, thus giving the competitive relationship representation node a natural property of aggregating information between the two parties.
[0116] In S205, the three topologies constructed in S201 to S204 are merged and integrated: the node set consists of all enterprise nodes. All resource nodes And all competing relationships represent nodes Composition; the edge set consists of all required edges All associated edges and all competing sides Composition; Initial characteristic matrix The competitive-aware triplet resource graph is now complete by stacking the initial feature vectors of all nodes row by row. The construction of.
[0117] This embodiment breaks down graph construction into five steps, making the creation process of each type of edge and node independently controllable. The quality of each part of the topology can be verified and optimized separately. In embodiment 1, the construction of three types of nodes and three types of edges is completed at once in S102 without explicitly separating the topology construction stage. The step-by-step construction method of this embodiment is more flexible in terms of fine-grained control of the topology structure and more precise in the quality control of each part of the graph. After the graph is constructed in this embodiment, the subsequent steps S103 to S108 are executed in the same way as in embodiment 1, that is, feature learning is performed through CD-GAT, competitive quantization index is calculated, an initial allocation matrix is constructed, differentiable iterative optimization is performed, and a global resource scheduling scheme is output.
[0118] Example 3: The competition-aware triplet resource graph is input into the competition-diffusion graph attention network. Attention weights are calculated for different types of edges using a type-aware multi-relation attention mechanism to obtain the attention coefficients for each edge, including:
[0119] S301. Configure independent attention parameter vectors for the three edge types: demand edge, associated edge, and competing edge.
[0120] S302. For each edge in the competitive sensing triplet resource graph, select the corresponding attention parameter vector according to the edge type, and calculate the attention coefficient of the edge by performing the inner product operation between the attention parameter vector and the node feature vector; the formula for calculating the attention coefficient is:
[0121]
[0122] in, This represents the attention coefficient between node u and node v in layer l. This represents the attention parameter vector corresponding to the edge type of edge (u,v). Indicates the first The linear transformation matrix of the layer, and Representing nodes respectively and nodes In the The feature vector of the layer, This represents a vector concatenation operation, where LeakyReLU is the activation function.
[0123] S303. Traverse all edges in the competitive perception triplet resource graph, match each edge with the attention parameter vector corresponding to the edge type and calculate the attention coefficient to obtain the attention coefficient of each edge containing all edges.
[0124] Specifically, this embodiment refines the type-aware multi-relation attention mechanism in S103 of Embodiment 1, and improves the edge type differentiation method from a one-hot encoding shared parameter scheme to an independent parameter scheme.
[0125] In S301, three independent attention parameter vectors are configured for each of the three edge types: demand edge, associated edge, and competing edge. And correspondingly configure three sets of independent linear transformation matrices. subscript , , These correspond to demand edge, related edge, and competing edge types, respectively.
[0126] In S302, for the edges in the graph According to its edge type Select the corresponding attention parameter vector and linear transformation matrix Then the node and nodes In the eigenvectors of the layer and Each through its corresponding Perform a linear transformation, and transform the two... Concatenating the dimensional vectors yields A dimensional vector, then combined with the attention parameter vector The inner product operation is performed, and finally, the original attention coefficient of the edge is output through the LeakyReLU activation function. The physical meaning of the above attention coefficient calculation is that different types of edges carry different semantics when propagating information. Specifically, demand edges convey the matching relationship between enterprise demand and resource supply; association edges convey the cooperation or substitutability relationship between similar entities; and competition edges convey the competitive game relationship between multiple entities for limited resources. Therefore, the attention mechanisms for these three edge types should use completely independent parameter spaces to capture their respective semantic differences, avoiding semantic interference caused by shared parameters. After calculating the original attention coefficients on each edge, a Softmax normalization operation is performed on all neighbor edges of each node, making the sum of the attention coefficients of all outgoing edges of each node equal to 1.
[0127] In S303, all edges in the graph are traversed, and for each edge, the corresponding attention parameter vector is selected according to the edge type and the attention coefficient is calculated, as in S302. Finally, a set of attention coefficients containing all edges is obtained.
[0128] Compared to the scheme in Example 1 that uses a shared basic parameter vector and element-wise multiplication of edge types with one-hot encoding, the independent attention parameter scheme in this embodiment allocates a completely independent parameter space for each edge type. The attention calculation logic of each of the three types of edges is not affected by the changes in parameters of other edge types, and can more accurately capture the semantic differences of heterogeneous edge relationships. In scenarios where demand edges and competition edges are simultaneously connected to the same enterprise node, independent parameters enable the model to assign distinctly different attention weight distributions for demand matching and competition adversarial, a fine distinction that is difficult to achieve with the shared parameter scheme in Example 1. The remaining steps S101 to S102 and S104 to S108 of this embodiment are executed in the same way as in Example 1.
[0129] Example 4: Based on the attention coefficients of each edge, feature propagation is performed along the competing edges through a competitive diffusion message passing mechanism. A competitive diffusion inhibition factor is introduced during message passing to attenuate the feature propagation intensity passing through nodes representing competing relationships, resulting in node feature representations, including:
[0130] S401. Aggregate the features of each enterprise node connected to the competitive relationship node after attention weighting to obtain the updated feature representation of the competitive relationship node;
[0131] S402. Determine the value of the competitive diffusion inhibition factor based on the number of enterprise nodes connected to the same competitive relationship node; wherein, the more enterprise nodes there are, the larger the value of the competitive diffusion inhibition factor.
[0132] S403. When updating the features of enterprise nodes through competing edges, based on the competitive relationship representing the updated feature representation of the node and the competitive diffusion suppression factor, multiplicative suppression is applied to the features transmitted through the competing edges to obtain the feature representation of the enterprise node after competitive diffusion suppression; the calculation formula for the feature representation of the enterprise node after competitive diffusion suppression is as follows:
[0133]
[0134] in, Represents enterprise nodes In the The feature components obtained by the layer through competing edges For enterprise nodes The relationship between nodes represents competition. Attention coefficient between them For the learnable linear transformation matrix on the enterprise node side, For enterprise nodes In the The feature vector of the layer, Representing nodes in a competitive relationship In the The feature vector of the layer, As a competitive diffusion inhibitor, It is a non-linear activation function;
[0135] S404. Based on the feature representation of the enterprise node after competitive diffusion suppression, integrate the features of each node transmitted through non-competitive edges and the features transmitted through competitive edges to obtain the node feature representation.
[0136] Specifically, based on the independent attention parameter scheme adopted in Example 3, this embodiment refines the competitive diffusion message passing mechanism in S104 of Example 1, and improves the value of the competitive diffusion inhibition factor from a fixed value to an adaptive generation mechanism.
[0137] In S401, competition relationships represent nodes. In the The feature update of a layer is calculated by aggregating the attention-weighted features of all enterprise nodes connected to it: ,in Representing competitive relationships, nodes are represented. In the Layer feature representation, To represent nodes in a competitive relationship The set of all enterprise nodes connected by competing edges. For enterprise nodes The relationship between nodes represents competition. The attention coefficients between them were calculated using the independent attention parameter scheme of Example 3. Let the competitive relationship be represented by the learnable linear transformation matrix on the node side, with dimension . , For enterprise nodes In the The feature vector of the layer, For ELU nonlinear activation function, the physical meaning of this aggregation operation is: the competitive relationship means that the feature update of the node is the weighted fusion result of the features of all participating enterprises under the guidance of attention weight, which reflects the overall characteristics of the resource competition environment.
[0138] In S402, the competition relationship output from S401 represents the updated feature representation of the node. Input a parameterized adaptive inhibition factor generation network to compute the competitive diffusion inhibition factor: ,in It is a competitive diffusion inhibition factor, taking values in the interval (0,1). The learnable linear transformation vector used to generate the suppression factor has dimensions and Consistent, As a bias scalar, this adaptive mechanism does not require pre-setting an explicit functional relationship between the number of enterprise nodes and the inhibition factor, but instead represents the competitive relationship as a feature after node updates. Mapped to inhibition strength via the sigmoid function: when When it contains information about large-scale competition, that is, when many companies are competing and the intensity of demand varies, The modulus tends to increase after [the process]. The projected scalar value increases accordingly, and the sigmoid output approaches 1, thus achieving a greater suppression strength. When the competition scale is small, the suppression factor tends to a smaller value, retaining more information about the competition relationship.
[0139] In S403, enterprise nodes The feature components obtained through competing edges are as described above. The enterprise node feature update formula is calculated, where Generated by the adaptive mechanism of S402. This multiplication suppression term... The physical meaning is: when competition is intense, competitive signals are significantly attenuated during propagation, and the model tends to make allocation decisions based on the enterprise's own demand characteristics and resource matching degree, avoiding excessive interference from competitive confrontation signals; when competition is mild, more competitive signals are preserved, providing reference competitive situation information for allocation decisions.
[0140] In S404, the enterprise node will be in the... The layer obtains the first-order feature of each node by summing the intermediate feature components aggregated through non-competitive edges (demand edges and association edges) and the feature components transmitted through competitive edges and adaptively suppressed and decayed element-wise. Layer feature representation; then the feature representations output from layers 0 to 2 (a total of three layers) are concatenated along the dimensional direction, i.e., for enterprise nodes. Its final characteristics are expressed as The dimension is 384, and the resource nodes and competition relationship representation nodes are constructed in the same way to form the final feature representation.
[0141] Compared to the fixed competitive diffusion inhibitory factor in Example 1, the adaptive inhibitory factor mechanism in this example allows the inhibitory strength to be dynamically adjusted according to the scale of competition: when there are many participating companies and their demands vary greatly, a stronger inhibitory factor is automatically applied to prevent the spread of competitive characteristics from overwhelming the individual demand characteristics of companies; when the scale of competition is small and the demands of participating companies are similar, the inhibitory strength is automatically reduced to preserve the structured information of the competitive environment. This adaptive design avoids the problems of insufficient or excessive inhibitory factor suppression that may occur in different competitive scenarios. The remaining steps of this example are the same as in Example 3.
[0142] Example 5: Based on the node feature representation, calculate the competition quantification index for each resource corresponding to each resource node, including:
[0143] S501. Based on the feature representations of each enterprise node connected to the competitive relationship representation node, calculate the competitiveness probability distribution of each enterprise node in the competition corresponding to the competitive relationship representation node; wherein, the competitiveness probability distribution is obtained by calculating the similarity between the feature representation of the enterprise node and the feature representation of the competitive relationship representation node and then normalizing it.
[0144] S502. Calculate the quantitative indicators of competition based on the probability distribution of competitiveness; the formula for calculating the quantitative indicators of competition is:
[0145]
[0146] in, Representing competitive relationships, nodes are represented. Corresponding quantitative indicators of competition This indicates a node connected to a competing node. The number of enterprise nodes, This indicates that node j, representing enterprise node, is in a competitive relationship with node j. The corresponding probability of competitiveness in the competition.
[0147] Specifically, in this embodiment, the method of calculating the competitive quantification index in S105 of embodiment 1 is replaced from variance competitive dispersion to competitive entropy index based on information entropy.
[0148] In S501, the competitive relationship is represented by the node. Its connection has Given that the final feature representation and competition relationship representation of each enterprise node have been obtained by concatenating the multi-layer features of S104, the enterprise nodes... In competition Competitive probability in Calculated using a scaled dot product attention mechanism and normalized using Softmax: ,in For enterprise nodes The final feature representation has a dimension of 384. Representing nodes in a competitive relationship The final feature representation has a dimension of 384. The dot product of two feature vectors measures the compatibility between firm characteristics and competitive environment characteristics. The scaling factor, divided by the square root of the feature dimension, stabilizes the dot product value within a reasonable range, preventing the gradient from approaching zero due to the exponential operation of Softmax when the feature dimension is large. The summation in the denominator also covers contention. All The number of companies participating in the competition makes the probability distribution... satisfy Competitive probability The physical meaning is: in resources In a competitive environment, enterprises The degree of relative competitive advantage compared to other competing companies The larger the size, the more likely the enterprise is to be successful. The higher the degree of matching with the competitive environment of the resource, the more obvious the relative advantage in the competition.
[0149] In S502, the competition entropy is calculated based on the above-mentioned competitiveness probability distribution: Competition entropy The range of values is When the competition entropy reaches its maximum value When, it means When the competitive entropy is at its minimum of 0, it indicates that one firm has an absolute competitive advantage with a probability of 1, while the competitiveness of other firms is negligible, resulting in a unipolar monopoly. The physical meaning of competitive entropy is a specialized application of information entropy in information theory within the context of resource competition: it not only quantifies the intensity of competition (the higher the theoretical upper limit of competitive entropy, the more participating firms), but also characterizes the equilibrium of competition. Specifically, a more uniform probability distribution results in a higher competitive entropy, indicating a competition closer to perfect competition; a more concentrated probability distribution results in a lower competitive entropy, indicating a competition closer to oligopoly.
[0150] By simultaneously applying competitive entropy to the competitive equilibrium sub-objective of multi-objective optimization in S107, the optimization process can maximize global matching benefits while penalizing allocation schemes with excessive competition concentration, thereby promoting the convergence of allocation results towards a more balanced resource competition pattern and achieving long-term fairness and global optimization in the allocation of R&D resources for science and technology innovation enterprises in the incubation platform.
[0151] Compared to the variance-based competitive dispersion index in Example 1, variance only measures the distribution of the similarity of features among competing firms in the vector space, failing to characterize the relative competitive structure and strength / weakness relationships between firms. In contrast, competitive entropy, based on a complete probability distribution, captures the structural characteristics of competition, such as multiple strong firms coexisting, one dominant firm with many weaker firms, and balanced competition, providing a more comprehensive and refined characterization of the competitive state. The remaining steps S101 to S104 and S106 to S108 of this embodiment are consistent with Example 1.
[0152] The following section provides a unified description of the specific specifications and training strategies for the CD-GAT network model, and this content applies to all the above embodiments.
[0153] For example, CD-GAT networks are stacked together. Layers, each layer has a feature dimension of 1 for both input and output. In Example 1, each layer contains a shared set of linear transformation matrices and a basic attention parameter vector that interacts with edge type one-hot encoding; in Examples 3 to 5, each layer contains three independent sets of type-aware parameter matrices: demand edge correspondence. and The associated edge corresponds to and Competitive edge corresponds to and Linear transformation matrix on the enterprise node side Dimensions The competitive relationship represents the linear transformation matrix on the node side. Dimensions The Dropout ratio between all layers is uniformly set to 0.1 to prevent overfitting; all edge attention calculations, message aggregations, and nonlinear transformations within each layer share the same feature dimension. Skip connections are constructed between layers through feature concatenation: the node features of the input layer (layer 0), layer 1, and layer 2 are concatenated along the dimensional direction to obtain the final feature representation of each node, with a dimension of [dimensional value missing]. .
[0154] The scoring network MLP_score is used to calculate the enterprise-resource matching score in the initial global assignment matrix. Its structure is as follows: The input layer receives a 384-dimensional concatenated vector of the final feature representations of the enterprise and the final feature representations of the resource, with a total input dimension of 768. After passing through the first fully connected layer, the 768 dimensions are mapped to 256 dimensions. A non-linear transformation is performed through the ReLU correction linear unit activation function. Then, the 256 dimensions are mapped to 64 dimensions through the second fully connected layer. After passing through the second ReLU activation function, the 64 dimensions are finally mapped to a 1-dimensional scalar output by the third fully connected layer, which serves as the logit value of the matching score. Then, the sigmoid function is used to map it to the (0,1) interval.
[0155] Each learnable parameter matrix is initialized with weights using Xavier initialization, meaning each element in the weight matrix starts with a mean of 0 and ends with a mean of 0. Sampling is performed on a truncated normal distribution with standard deviation, where and These represent the input and output dimensions of the weight matrix, respectively, and the bias term is initialized to a zero vector or a zero scalar. Preferably, the optimizer is the Adam optimizer, with an initial learning rate of 0.001. First-order moment attenuation coefficient Second-order moment attenuation coefficient Furthermore, the weight coefficients of the global matching benefit sub-objective in multi-objective optimization. Weighting coefficients of the resource utilization sub-objective Weight coefficients of the competitive equilibrium sub-objective The logarithmic barrier coefficient of scarce resource capacity constraints .
[0156] The training process employs a three-stage strategy, including: The first stage is the CD-GAT network pre-training stage, lasting 500 rounds. All parameters of the global allocation optimizer are frozen, and edge prediction is used as the pre-training task. Specifically, the input is a competition-aware triplet resource graph. After CD-GAT outputs the node feature representations of each enterprise node and resource node, a bilinear decoder predicts the probability of a demand edge between an enterprise and a resource. The training objective is to minimize the binary classification cross-entropy loss. This stage only updates all learnable parameters of CD-GAT. The second stage is the allocation optimizer training stage, also lasting 500 rounds. All parameters of the CD-GAT network are frozen, and only the parameters of the global allocation optimizer are trained. The global allocation objective function, composed of a weighted combination of global matching benefit sub-objectives, resource utilization sub-objectives, and competition equilibrium sub-objectives, serves as the optimization objective. An additional KL divergence regularization term is added, where the KL divergence between the distribution of the differentiable sampled allocation matrix and the distribution of the allocation matrix before each round of updates is added as a regularization penalty to the loss function to prevent drastic oscillations in the allocation matrix during iterative optimization. The third stage is the end-to-end joint fine-tuning stage, which involves 300 training rounds: simultaneously unfreezing the CD-GAT network and allocating all parameters of the optimizer to the joint loss function. As the optimization objective, among which The edge prediction loss in the first stage, Distribute the loss globally in the second phase. To assess the fairness loss based on the approximation of the Shapley value, an approximate Shapley value for each firm in each competitive relationship is calculated using a parameterized marginal contribution estimation network. The difference between the actual distribution revenue obtained by each firm and its Shapley value is used as a fairness penalty.
[0157] Throughout all three stages, the batch size is uniformly set to 32. The learning rate decays exponentially by a decay factor of 0.95 every 100 training epochs. If the loss value on the validation set does not decrease for 50 consecutive epochs, an early stopping mechanism is triggered to terminate training, retaining the model checkpoint with the lowest validation loss. After the above training process is completed, the CD-GAT model is directly deployed on a GPU-accelerated server for feature inference, and the optimizer is deployed on the application server to perform online optimization. The entire system does not require any additional training computations.
[0158] In summary, the deep learning-based intelligent scheduling method for technology incubation and R&D resources provided in this application obtains enterprise demand feature vectors and resource attribute feature vectors by structuring multi-source heterogeneous data containing multi-dimensional attribute data of resource demanders and attribute description data of resource suppliers. Based on this, a competition-aware triplet resource graph is constructed. This graph introduces three types of nodes: enterprise nodes, resource nodes, and competition relationship representation nodes for explicitly modeling competitive relationships. Demand edges connect enterprises to resources, association edges connect similar nodes, and competition edges connect multiple enterprises to the same competition relationship representation node, thus explicitly transforming the competition relationship of multiple entities for the same scarce resource into a computable graph topology. This graph is then input into a competition diffusion graph attention network, and a type-aware multi-relationship attention mechanism is used to configure independent values for demand edges, association edges, and competition edges. Attention parameters and attention weights are calculated to distinguish the differentiated semantic information carried by different relationship types. Simultaneously, a competitive diffusion message passing mechanism is used to propagate features along competitive edges, and a competitive diffusion inhibition factor is introduced during message passing to adaptively attenuate the feature propagation intensity through nodes representing competitive relationships. This results in node feature representations that accurately reflect the true state of each enterprise in a competitive environment. Furthermore, competitive quantitative indicators for each resource are calculated based on the node feature representations to characterize the urgency of supply and demand matching for each resource under different enterprise competitive environments. An initial global allocation matrix is constructed to characterize the initial allocation tendency between enterprises and resources. Based on the competitive quantitative indicators, a multi-objective allocation optimization objective is constructed, and the initial global allocation matrix is solved through differentiable iterative optimization to obtain the optimized global resource allocation matrix, from which a global resource scheduling scheme is output. This technical solution can coordinate the allocation relationships of all enterprises and all resources from a global perspective, rather than performing isolated local optimal matching for individual needs, thereby improving the adaptability and overall efficiency of global R&D resource scheduling in complex competitive scenarios.
[0159] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.
[0160] Based on the same inventive concept, this application also provides a deep learning-based intelligent scheduling system for technology incubation and R&D resources to implement the aforementioned deep learning-based intelligent scheduling method for technology incubation and R&D resources. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more deep learning-based intelligent scheduling system embodiments for technology incubation and R&D resources provided below can be found in the limitations of the deep learning-based intelligent scheduling method for technology incubation and R&D resources described above, and will not be repeated here.
[0161] In one exemplary embodiment, such as Figure 2 As shown, a deep learning-based intelligent scheduling system 10 for technology incubation and R&D resources is provided to achieve the following:
[0162] The data structuring module 11 is used to perform data structuring processing on the acquired multi-source heterogeneous data to obtain enterprise demand feature vectors and resource attribute feature vectors; wherein, the multi-source heterogeneous data includes multi-dimensional attribute data of resource demanders and attribute description data of resource suppliers.
[0163] The graph construction module 12 is used to construct a competition-aware triplet resource graph based on enterprise demand feature vectors and resource attribute feature vectors. The competition-aware triplet resource graph includes three types of nodes: enterprise nodes, resource nodes, and competition relationship representation nodes. Enterprise nodes and resource nodes are connected by demand edges, nodes of the same type are connected by association edges, and multiple enterprise nodes are connected to the same competition relationship representation node by competition edges.
[0164] The multi-relation attention module 13 is used to input the competition-aware triplet resource graph into the competition diffusion graph attention network, and calculate the attention weights for different types of edges through the type-aware multi-relation attention mechanism to obtain the attention coefficients of each edge.
[0165] The competitive diffusion feature propagation module 14 is used to propagate features along the competitive edges based on the attention coefficients of each edge through a competitive diffusion message passing mechanism. During the message passing process, a competitive diffusion inhibition factor is introduced to control the attenuation of the feature propagation intensity through the nodes representing the competitive relationship, so as to obtain the node feature representation.
[0166] The competition quantification index calculation module 15 is used to calculate the competition quantification index of each resource corresponding to each resource node based on the node characteristic representation; wherein, the competition quantification index represents the urgency of supply and demand matching of each resource under different enterprise node competition environments;
[0167] The initial allocation matrix construction module 16 is used to construct an initial global allocation matrix based on node feature representations; wherein, the initial global allocation matrix is a two-dimensional matrix with enterprises as rows and resources as columns, and the elements in the initial global allocation matrix represent the initial allocation tendency between the corresponding enterprise and the corresponding resource;
[0168] The multi-objective optimization solution module 17 is used to construct a multi-objective allocation optimization objective based on the competition quantification index, and to perform differentiable iterative optimization on the initial global allocation matrix to obtain the optimized global resource allocation matrix.
[0169] The scheduling scheme output module 18 is used to output a global resource scheduling scheme based on the optimized global resource allocation matrix.
[0170] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the deep learning-based intelligent scheduling method for technology incubation and R&D resources as described above.
[0171] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.
[0172] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0173] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.
Claims
1. A deep learning-based intelligent scheduling method for technology incubation R&D resources, characterized in that, The method includes: S101. Perform data structuring processing on the acquired multi-source heterogeneous data to obtain enterprise demand feature vectors and resource attribute feature vectors; wherein, the multi-source heterogeneous data includes multi-dimensional attribute data of resource demanders and attribute description data of resource suppliers. S102. Construct a competition-aware triplet resource graph based on the enterprise demand feature vector and the resource attribute feature vector; wherein, the competition-aware triplet resource graph includes three types of nodes: enterprise nodes, resource nodes, and competition relationship representation nodes. Enterprise nodes and resource nodes are connected through demand edges, nodes of the same type are connected through association edges, and multiple enterprise nodes are connected to the same competition relationship representation node through competition edges. S103. Input the competitive sensing triplet resource graph into the competitive diffusion graph attention network, and calculate the attention weights for different types of edges through a type-aware multi-relation attention mechanism to obtain the attention coefficients of each edge. S104. Based on the attention coefficients of each edge, feature propagation is performed along the competing edges through a competitive diffusion message passing mechanism. During the message passing process, a competitive diffusion inhibition factor is introduced to control the attenuation of the feature propagation intensity passing through the nodes representing the competitive relationship, thereby obtaining the node feature representation. S105. Based on the node feature representation, calculate the competition quantification index of each resource corresponding to each resource node; wherein, the competition quantification index characterizes the urgency of supply and demand matching of each resource under different enterprise node competition environments; S106. Based on the node feature representation, construct an initial global allocation matrix; wherein, the initial global allocation matrix is a two-dimensional matrix with enterprises as rows and resources as columns, and the elements in the initial global allocation matrix represent the initial allocation tendency between the corresponding enterprise and the corresponding resource; S107. Based on the competition quantification index, construct a multi-objective allocation optimization objective, and perform differentiable iterative optimization on the initial global allocation matrix to obtain the optimized global resource allocation matrix. S108. Output the global resource scheduling scheme based on the optimized global resource allocation matrix.
2. The method of claim 1, wherein, The step of constructing a competition-aware triplet resource map based on the enterprise demand feature vector and the resource attribute feature vector includes: S201. Based on the similarity between the enterprise demand feature vector and the resource attribute feature vector, determine the weight of the demand edge between the enterprise node and the resource node to obtain the enterprise-resource matching relationship with demand edge weights. S202. Based on the semantic similarity between the enterprise demand feature vectors, establish the association edges between different enterprise nodes; based on the functional similarity between the resource attribute feature vectors, establish the association edges between different resource nodes, and obtain a node relationship topology with similar association edges. S203. Based on the enterprise-resource matching relationship with demand edge weights, when multiple enterprise nodes have a demand for the same resource node, a corresponding competition relationship representation node is created for the resource node to obtain the competition relationship representation node creation result; based on the competition relationship representation node creation result, each enterprise node participating in the competition is connected to the competition relationship representation node through the competition edge to obtain the competition relationship topology with competition edge. S204. Generate the initial feature vector of the node representing the competitive relationship based on the weighted average of the resource attribute feature vector and the enterprise demand feature vector of each participating enterprise node. S205. Based on the enterprise-resource matching relationship with demand edge weights, the node relationship topology with similar associated edges, and the competitive relationship topology with competitive edges, the graph structure is integrated to obtain the competition-aware triplet resource graph.
3. The method of claim 1, wherein, The step of inputting the competition-aware triplet resource graph into the competition-diffusion graph attention network, and calculating the attention weights for different types of edges through a type-aware multi-relation attention mechanism to obtain the attention coefficients of each edge includes: S301. Configure independent attention parameter vectors for the three edge types: demand edge, associated edge, and competing edge. S302. For each edge in the competitive sensing triplet resource graph, select the corresponding attention parameter vector according to the edge type, and calculate the attention coefficient of the edge by performing an inner product operation between the attention parameter vector and the node feature vector; the formula for calculating the attention coefficient is: ; wherein, denotes the attention coefficient between node u and node v in the lth layer, the superscript denotes vector transposition operation, denotes the edge type of the edge corresponding to the attention parameter vector, denotes the attention parameter vector corresponding to the edge denotes the linear transformation matrix of the lth layer, denotes the feature vector of node and node in the lth layer, denotes the feature vector of node in the lth layer, denotes the feature vector of node denotes vector splicing operation, and LeakyReLU is an activation function. S303. Traverse all edges in the competitive sensing triplet resource graph, match each edge with the attention parameter vector corresponding to the edge type and calculate the attention coefficient to obtain the attention coefficient of each edge containing all edges.
4. The method of claim 3, wherein, Based on the attention coefficients of each edge, features are propagated along the competing edges through a competitive diffusion message passing mechanism. During message passing, a competitive diffusion inhibition factor is introduced to attenuate the feature propagation intensity passing through the nodes representing the competing relationships, resulting in node feature representations, including: S401. Aggregate the attention-weighted features of each enterprise node connected to the competitive relationship representation node to obtain the updated feature representation of the competitive relationship representation node; S402. Determine the value of the competitive diffusion inhibition factor based on the number of enterprise nodes connected to the same competitive relationship representation node; wherein, the more enterprise nodes there are, the larger the value of the competitive diffusion inhibition factor. S403. When updating the features of the enterprise node through the competing edge, based on the competition relationship representing the updated feature representation of the node and the competition diffusion suppression factor, multiplicative suppression is applied to the features transmitted through the competing edge to obtain the feature representation of the enterprise node after competition diffusion suppression; the calculation formula for the feature representation of the enterprise node after competition diffusion suppression is: ; in, This represents the feature component that enterprise node u obtains through the competing edge at layer l+1. Let be the attention coefficient between the enterprise node u and the competing relationship representing node c. For the learnable linear transformation matrix on the enterprise node side, For enterprise nodes In the The feature vector of the layer, The nodes representing the competitive relationship In the The feature vector of the layer, σ is the competitive diffusion inhibition factor, and σ is the nonlinear activation function; S404. Based on the feature representation of the enterprise node after competitive diffusion suppression, the features of each node transmitted through non-competitive edges and the features transmitted through competitive edges are integrated to obtain the node feature representation.
5. The method of claim 1, wherein, The step of calculating the competition quantification index of each resource corresponding to each resource node based on the node feature representation includes: S501. Based on the feature representations of each enterprise node connected to the competitive relationship representation node, calculate the competitiveness probability distribution of each enterprise node in the competition corresponding to the competitive relationship representation node; wherein, the competitiveness probability distribution is obtained by calculating the similarity between the feature representation of the enterprise node and the feature representation of the competitive relationship representation node and then normalizing it. S502. Calculate the competition quantification index based on the aforementioned competitiveness probability distribution; the calculation formula for the competition quantification index is as follows: ; wherein, represents a competitive relationship representation node corresponding to the competitive quantification indicator, represents the number of enterprise nodes connected to the competitive relationship representation node corresponding to the competitive quantification indicator, represents the competitive power probability of enterprise node j in the competition corresponding to the competitive relationship representation node corresponding to the competitive quantification indicator.
6. A deep learning-based intelligent scheduling system for technology incubation R&D resources, characterized in that, The system includes: The data structuring module is used to perform data structuring processing on the acquired multi-source heterogeneous data to obtain enterprise demand feature vectors and resource attribute feature vectors; wherein, the multi-source heterogeneous data includes multi-dimensional attribute data of resource demanders and attribute description data of resource suppliers; The graph construction module is used to construct a competition-aware triplet resource graph based on the enterprise demand feature vector and the resource attribute feature vector. The competition-aware triplet resource graph includes three types of nodes: enterprise nodes, resource nodes, and competition relationship representation nodes. Enterprise nodes and resource nodes are connected by demand edges, nodes of the same type are connected by association edges, and multiple enterprise nodes are connected to the same competition relationship representation node by competition edges. The multi-relation attention module is used to input the competition-aware triplet resource graph into the competition-diffusion graph attention network, and calculate the attention weights for different types of edges through a type-aware multi-relation attention mechanism to obtain the attention coefficients of each edge. The competitive diffusion feature propagation module is used to propagate features along the competitive edges based on the attention coefficients of each edge through a competitive diffusion message passing mechanism. During the message passing process, a competitive diffusion inhibition factor is introduced to control the attenuation of the feature propagation intensity passing through the nodes representing the competitive relationship, thereby obtaining the node feature representation. The competition quantification index calculation module is used to calculate the competition quantification index of each resource corresponding to each resource node based on the node feature representation; wherein, the competition quantification index characterizes the urgency of supply and demand matching of each resource under different enterprise node competition environments; An initial allocation matrix construction module is used to construct an initial global allocation matrix based on the node feature representation; wherein, the initial global allocation matrix is a two-dimensional matrix with enterprises as rows and resources as columns, and the elements in the initial global allocation matrix represent the initial allocation tendency between the corresponding enterprise and the corresponding resource; The multi-objective optimization solution module is used to construct a multi-objective allocation optimization objective based on the competition quantification index, and to perform differentiable iterative optimization on the initial global allocation matrix to obtain the optimized global resource allocation matrix. The scheduling scheme output module is used to output a global resource scheduling scheme based on the optimized global resource allocation matrix. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6. When the processor executes the computer program, it implements the method of any one of claims 1 to 5.
8. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.
Citation Information
Patent Citations
Intelligent matching methods, systems, devices, and readable storage media for startup incubators
CN113538115B
Innovation and entrepreneurship incubation resource matching system
CN121278095A
