A big data matching recommendation system and method based on technology transfer platform

Abstract

The present invention discloses a big data matching recommendation system and method based on a technology transfer platform, which relates to the field of artificial intelligence technology and includes a data processing module, a knowledge graph construction module, an intent analysis module, a similarity calculation module, and a recommendation ranking module. Among them, the data processing module uses a domain adaptive encoder to realize the standardization of heterogeneous data; the knowledge graph construction module extracts technical entities and relationships based on a pre-trained language model; the intent analysis module extracts the core features of both supply and demand sides through a dual decoder mechanism; the similarity calculation module integrates graph structure analysis and differentiated weight allocation, and realizes dynamic compensation; the recommendation ranking module generates personalized recommendation results in combination with market potential assessment. The present invention significantly improves the intelligent matching accuracy of the technology transfer platform, realizes the precise docking of cross-domain technology supply and demand, and has important practical value.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a big data matching recommendation system and method based on a technology transfer platform. Background Art

[0002] Technology transfer platforms are infrastructure that promotes collaborative innovation between industry, academia, and research. Existing technology transfer platforms use methods such as keyword matching and text similarity calculation to connect supply and demand.

[0003] Currently, technology transfer platforms face the following deficiencies in practical application: In practice, technical information originates from multiple entities, including universities, research institutes, and enterprises. The formats and standards of the technical information provided by these entities vary, resulting in inefficient information integration. Furthermore, conventional text processing methods struggle to accurately understand the specialized terminology and implicit knowledge contained in technical documents, reducing the accuracy of matching technology suppliers and demanders. Furthermore, existing technology transfer platforms lack mechanisms for assessing the market potential of technologies, hindering the practicality of their recommendations. These issues limit the effectiveness of technology transfer platforms and hinder the efficiency of technology transfer. Summary of the Invention

[0004] The present invention provides a big data matching recommendation system and method based on a technology transfer platform, which is difficult to connect supply and demand across fields.

[0005] In view of this, a first aspect of the present invention provides a big data matching recommendation system based on a technology transfer platform, comprising: a data processing module for converting a technical information set from distributed heterogeneous data sources into a standardized vector representation using a domain adaptive encoder;

[0006] A knowledge graph construction module, which is used to extract technical entities and relationships in standardized vector representations through pre-trained large-scale language models to build a multi-domain fusion knowledge graph;

[0007] The intent parsing module is used to parse the intent of both the supply and demand sides based on a multi-domain fusion knowledge graph, using a dual decoder mechanism to extract core features and generate intent representation vectors;

[0008] A similarity calculation module is used to calculate the multi-dimensional similarity between intent representation vectors using a multi-level heterogeneous graph network that integrates graph structure analysis and a differentiated weight allocation mechanism, and to implement dynamic similarity compensation for feature-sparse regions.

[0009] The recommendation ranking module is used to identify the matching relationship between technology suppliers and demanders based on multi-dimensional similarity, conduct comprehensive scoring and ranking based on market potential evaluation indicators, and generate personalized recommendation results.

[0010] Optionally, the processing flow of the data processing module includes:

[0011] Preprocess the technical information sets in distributed heterogeneous data sources and construct a multi-level technical ontology index structure;

[0012] Process structured data, semi-structured data, and unstructured data in technical information sets, extracting technical terminology sets, functional description sets, and performance indicator sets;

[0013] Construct a domain-adaptive encoder, including a domain weight parameter matrix, a cross-domain mapping function, and a semantic preservation loss function;

[0014] Training a domain-adaptive encoder to encode a set of technical terms, a set of functional descriptions, and a set of performance metrics into initial vector representations;

[0015] The initial vector representation is subjected to dimensionality reduction processing to obtain a standardized vector, and a bidirectional traceable mapping mechanism from technical description text to standardized vector is established.

[0016] Optionally, the processing flow of the knowledge graph construction module includes:

[0017] The standardized vector is converted into technical description text through a bidirectional traceability mapping mechanism, and technical entity recognition is performed through a pre-trained large-scale language model to extract the technical entity set;

[0018] Construct a domain terminology corresponding structure, perform domain-adaptive transformation on the technical entity set, and form a cross-domain technical terminology mapping relationship;

[0019] Through multi-level technical relationship analysis, the explicit and implicit relationships between technical entities are extracted from the technical entity set to form a technical relationship set;

[0020] Use the technical entity set as nodes and the technical relationship set as edges to build the initial graph structure, and assign weight values to the nodes and edges in the initial graph structure;

[0021] The initial graph structure is adjusted according to the feedback information of the preset technology matching samples, and inference relationships are added to the low-density areas of the initial graph structure to form a multi-domain fusion knowledge graph.

[0022] Optionally, the processing flow of the intent parsing module includes:

[0023] Construct an intent-capability dual decoder, which includes a technology supply-side decoding unit and a demand-side decoding unit;

[0024] Extracting a first subgraph structure associated with the technology supplier from the multi-domain fusion knowledge graph, inputting the first subgraph structure into the technology supplier decoding unit, and generating a supplier capability representation vector;

[0025] A second subgraph structure associated with the demander is extracted from the multi-domain fusion knowledge graph, and the second subgraph structure is input into the demander decoding unit to generate a demander intention representation vector.

[0026] Optionally, the processing flow of the intent parsing module also includes:

[0027] The technology function decomposition and reorganization algorithm is used to decompose the supplier's capability representation vector, obtain a set of basic functional units, and construct a functional unit dependency graph;

[0028] Decompose the demander's intention representation vector to obtain a set of basic demand units and construct a demand unit priority graph;

[0029] The mapping relationship between the basic functional unit set and the basic demand unit set is calculated to obtain the function-demand matching matrix. Based on the function-demand matching matrix, the supplier's capability representation vector and the demander's intention representation vector are fused and reorganized to generate the intention representation vector.

[0030] Optionally, the processing flow of the similarity calculation module includes:

[0031] Construct a multi-level heterogeneous graph network and map the supply-side capability representation vector and the demand-side intention representation vector to corresponding nodes;

[0032] Based on the characteristics of the technical field, matching historical data and technology maturity indicators, the weight coefficients of each feature dimension are dynamically assigned to form an adaptive weight matrix;

[0033] Calculate the similarity between mapping nodes in a multi-level heterogeneous graph network in three dimensions: direct similarity, indirect correlation, and complementary adaptability, and form a multidimensional similarity vector;

[0034] Identify feature-sparse regions in the intent representation vector space and generate multi-dimensional similarity compensation vectors;

[0035] The multi-dimensional similarity vector and the multi-dimensional similarity compensation vector are integrated to generate a comprehensive similarity score. Optionally, the processing flow of the recommendation ranking module includes:

[0036] Determine the initial matching relationship set between technology suppliers and demanders based on the comprehensive similarity score, and establish a technology transfer feasibility assessment system;

[0037] Based on the technology transfer feasibility evaluation system, the technology transfer feasibility of each supply and demand pair in the preliminary matching relationship set is quantitatively scored;

[0038] A weighted fusion method is used to combine the comprehensive similarity score with the technology transfer feasibility score to generate the final matching score;

[0039] The preliminary matching relationship set is grouped and sorted within the technical field according to the final matching score to generate personalized recommendation results.

[0040] A second aspect of the present invention provides a big data matching recommendation method based on a technology transfer platform, comprising: using a domain adaptive encoder to convert a technical information set from distributed heterogeneous data sources into a standardized vector representation;

[0041] Extract technical entities and relationships from standardized vector representations using pre-trained large-scale language models to build a multi-domain fusion knowledge graph.

[0042] Based on a multi-domain fusion knowledge graph, a dual decoder mechanism is used to analyze the intentions of both supply and demand parties, extract core features, and generate intent representation vectors;

[0043] A multi-layered heterogeneous graph network that integrates graph structure analysis and differentiated weight allocation is used to calculate the multi-dimensional similarity between intent representation vectors and implement dynamic similarity compensation for feature-sparse regions.

[0044] Based on multi-dimensional similarity, the matching relationship between technology suppliers and demanders is identified, and comprehensive scoring and ranking are performed based on technology transfer feasibility indicators to generate personalized recommendation results.

[0045] The beneficial effects of the present invention are as follows: the present invention realizes the unified expression and organization of heterogeneous data through domain-adaptive encoders and multi-level ontology index structures, thereby improving the accuracy and efficiency of technical information processing; the multi-domain fusion knowledge graph constructed based on dual-threshold judgment and dynamic optimization mechanism enhances the ability to discover technical associations and improves the integrity of knowledge representation; the dual-decoder structure and hierarchical decomposition and recombination algorithm are adopted to realize accurate analysis and feature alignment of technical supply and demand intentions, and enhance the interpretability of the matching process; the designed three-dimensional similarity calculation model and dynamic compensation mechanism effectively improve the accuracy and robustness of technical matching, and enhance the system's adaptability to changes in technological development; through multi-dimensional feasibility evaluation and adaptive sorting mechanism, the accuracy and practicality of recommendation results are improved, providing a reliable basis for technology transfer decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0047] Figure 1 This is a flowchart of a big data matching recommendation system based on a technology transfer platform.

[0048] Figure 2 This is a processing flow chart of the intent parsing module of a big data matching recommendation system based on a technology transfer platform.

[0049] Figure 3 This is a processing flow chart of the similarity calculation module of a big data matching recommendation system based on a technology transfer platform. DETAILED DESCRIPTION

[0050] In order to make the purpose, features, and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0051] Example 1, with reference to Figures 1 to 3 , which is the first embodiment of the present invention, provides a big data matching recommendation system based on a technology transfer platform. The flow chart of the system is as follows Figure 1 As shown, the system includes:

[0052] The data processing module is used to convert the technical information set from distributed heterogeneous data sources into a standardized vector representation through a domain adaptive encoder.

[0053] In a preferred embodiment of the present invention, the data processing module includes the following steps: preprocessing the technical information set in the distributed heterogeneous data source, and constructing a multi-level technical ontology index structure according to the information source type, data format and domain identification, specifically including: obtaining the technical information set in the distributed heterogeneous data source, including but not limited to patent documents, scientific papers, technical reports, product manuals and industry standards, and performing data cleaning operations on the technical information set; classifying and labeling the technical information set according to the information source type; dividing the technical information set into structured data, semi-structured data and unstructured data according to the data format; extracting domain identification information from the technical information set, including technical field classification codes, keyword sets and subject tags; based on the information source type, data format and domain identification, constructing a multi-level technical ontology index structure, which includes a domain layer (reflecting the relationship between different technical fields), a concept layer (representing the core technical concepts and their associations in each field), an attribute layer (describing the specific attributes and characteristics of the technical concepts) and an instance layer (containing specific technical instances and their attribute values).

[0054] Furthermore, a cross-modal coding unit is used to process structured data, semi-structured data and unstructured data in the technical information set, and extract a set of technical terms, a set of function descriptions and a set of performance indicators, specifically including: constructing a cross-modal coding unit containing multiple processing modules, and using the cross-modal coding unit to process structured data, semi-structured data and unstructured data respectively; extracting technical terms and performance indicators from structured data, parsing and extracting technical information in nested structures from semi-structured data, and extracting technical information contained in text and graphics from unstructured data; through cross-modal feature alignment, integrating the information extracted from different modal data to form a unified set of technical terms, function descriptions and performance indicators.

[0055] Furthermore, a domain adaptive encoder is constructed, including a domain weight parameter matrix, a cross-domain mapping function and a semantic preservation loss function. Specifically, the domain weight parameter matrix is set to adjust the importance of features in different domains; a cross-domain mapping function is designed to map input features to a specific domain representation space. , as follows:

[0056] ;

[0057] Among them, x is the input data, d is the target domain identifier, is the basic encoding function, is the domain weight parameter matrix, is the domain bias term, is a non-linear activation function.

[0058] Construct a semantic preservation loss function L to preserve the semantic content of the original technical information sem , the formula is as follows:

[0059] ;

[0060] in, L content For content retention loss, L structure To maintain the loss of structure, L domain is the domain consistency loss, Hyperparameters for weighing different loss components.

[0061] Furthermore, the encoding parameters of the domain adaptive encoder are trained by a contrastive learning method, and the trained domain adaptive encoder is used to encode the technical terminology set, the functional description set and the performance indicator set into an initial vector representation. Specifically, the contrastive learning framework including a positive sample pair generation strategy and a negative sample selection strategy is established, in which different expressions of the same technical information or expressions of the same technology in different fields are regarded as positive sample pairs, and representative negative samples are selected by combining random sampling and hard example mining methods; the contrastive loss function is defined as introducing a marginal mechanism to improve InfoNCE, and is combined with the semantic preservation loss function to form the overall training objective L total , the formula is as follows:

[0062] ;

[0063] ;

[0064] in, To balance the weight coefficient of the two losses, is the contrast loss, is the semantic preservation loss, is the similarity gap between negative samples and positive samples, is the temperature parameter (used to control the scaling of the similarity score), and m is the margin hyperparameter (used to explicitly enforce a minimum similarity gap between positive and negative samples).

[0065] It should be noted that the similarity of positive samples is , the negative sample similarity is (in ),in and is the encoding representation of the positive sample pair, is the cosine similarity function (used to calculate the two vectors and ), is the encoding representation of the negative sample. In essence, the contrast loss function introduces a margin mechanism to force the similarity gap between positive and negative samples to be at least m, while the temperature parameter Adjust the model's sensitivity to difficult samples, and the two work together to enable the encoder to learn more discriminative feature representations.

[0066] A batch gradient descent algorithm is used to optimize the encoding parameters, including an adaptive learning rate adjustment strategy, gradient clipping to prevent gradient explosion, and an early stopping strategy to prevent overfitting. The trained domain-adaptive encoder is used to convert the set of technical terms, the set of functional descriptions, and the set of performance indicators into vector representations, and the above three vector representations are integrated into the initial vector representation.

[0067] Furthermore, the initial vector representation is subjected to dimensionality reduction processing to obtain a standardized vector, and a bidirectional traceable mapping mechanism from the technical description text to the standardized vector is established, specifically including: using nonlinear dimensionality reduction technology to reduce the dimensionality of the initial vector representation while retaining the structural information of the data; standardizing the vector after dimensionality reduction to ensure the stability of the vector distribution; constructing a forward mapping mechanism from the technical description text to the standardized vector, mapping key technical terms to the vector space; constructing a reverse parsing mechanism from the standardized vector to the technical description text, decomposing the standardized vector into a linear combination of technical features, establishing a feature-text template library, converting the decomposed features into natural language descriptions, and generating fluent technical description text.

[0068] Optionally, a consistency verification mechanism for bidirectional mapping is implemented to ensure semantics are maintained during the conversion process; a mapping relationship database is constructed to support accurate retrieval and traceability analysis of technical information.

[0069] The data processing module transforms distributed heterogeneous data into standardized vector representations. This module offers the following technical advantages: a domain-adaptive encoder enables semantic preservation and feature extraction of technical information; a standardized information organization framework based on a multi-level technical ontology index structure; and a bidirectional traceability mapping mechanism supports accurate retrieval and analysis of technical information. These advantages provide a high-quality data foundation for matching recommendations on technology transfer platforms.

[0070] The knowledge graph construction module is used to extract technical entities and relationships in standardized vector representations through pre-trained large-scale language models to build a multi-domain fusion knowledge graph.

[0071] In another preferred embodiment of the present invention, the processing flow of the knowledge graph construction module includes: converting the standardized vector into a technical description text through a bidirectional traceable mapping mechanism, and performing technical entity recognition through a pre-trained large-scale language model to extract a technical entity set; constructing a domain term corresponding structure, performing domain adaptive transformation on the technical entity set, and forming a cross-domain technical term mapping relationship, specifically including: establishing a domain classification system, and performing multi-dimensional classification and annotation on technical entities; constructing a feature vector of the technical entity based on technical function characteristics and application scenario characteristics; calculating the similarity matrix of the feature vector to obtain the correlation strength between technical entities; constructing a cross-domain term mapping relationship based on a first preset threshold and a second preset threshold, constructing a technical entity pair with a correlation strength higher than the first preset threshold as an equivalent term mapping relationship, and constructing a technical entity pair with a correlation strength between the two preset thresholds as a related term mapping relationship; based on the constructed term mapping relationship, converting the technical entity into a unified standard representation. Wherein, the first preset threshold and the second preset threshold are determined by historical matching sample analysis and expert evaluation, and the first preset threshold is greater than the second preset threshold.

[0072] The pre-trained large-scale language model uses a bidirectional encoder structure based on the Transformer architecture, and processes semantic associations in technical texts through a self-attention mechanism. The model is pre-trained on a corpus containing technical literature and patent documents from multiple fields, using a masked language modeling training objective. To perform technical entity recognition tasks, the model undergoes targeted fine-tuning, uses conditional random fields to enhance sequence labeling, and incorporates domain ontology knowledge to constrain the results. The model supports the recognition of composite and nested entities in technical texts and can capture the precise meaning of professional terms through context-enhanced semantic representations. The model also incorporates adversarial training and multi-task learning strategies to enhance cross-domain recognition capabilities while maintaining the contextual relevance of entity representations, providing basic data support for subsequent term mapping and association analysis. The hierarchical attention mechanism and boundary detection algorithm are used in the processing process to improve the recognition accuracy of complex technical concepts, while the output results retain the original context information to support bidirectional traceability.

[0073] Furthermore, through multi-level technical relationship analysis, explicit and implicit relationships between technical entities are extracted from the technical entity set to form a technical relationship set, which specifically includes: building a basic technical knowledge base, which contains predefined technical relationship types and relationship patterns; extracting explicit technical relationships based on rule templates and semantic analysis, including component composition relationships, function implementation relationships and technical evolution relationships; using knowledge reasoning and deep learning methods to mine implicit technical relationships, including technical similarity relationships, technical dependency relationships and technical fusion relationships; performing credibility assessment and hierarchical classification on the identified relationships to form a structured technical relationship set.

[0074] Furthermore, the initial graph structure is constructed using the technical entity set as nodes and the technical relationship set as edges, and weight values are assigned to the nodes and edges in the initial graph structure. The initial graph structure is adjusted according to the feedback information of the preset technical matching samples, and inference relationships are added to the low-density areas of the initial graph structure to form a multi-domain fusion knowledge graph. Specifically, the following steps are performed: using the pre-labeled technical matching sample set to extract the association path characteristics between technical entity nodes; based on the frequency and intensity of technical entity associations, the weight values of the corresponding nodes and edges in the initial graph structure are updated; the node connection density of each area in the initial graph structure is calculated, and the area with node connection density lower than the preset density threshold is determined as a low-density area; in the low-density area, based on the semantic relevance of existing nodes and edges, new inference relationships are generated through transitive rules and analogical reasoning methods; the credibility of the newly generated inference relationships is evaluated, and the inference relationships with credibility higher than the preset relationship threshold are added to the initial graph structure; the above process is repeated until the initial graph structure converges to form the final multi-domain fusion knowledge graph. Among them, the node connection density is calculated by calculating the average number of connections between nodes in a unit area, the preset density threshold is determined by statistical analysis of historical data, and the preset relationship threshold is determined by expert evaluation and historical case verification.

[0075] The knowledge graph construction module of this invention offers the following technical advantages: a dual-threshold decision mechanism improves the discrimination of cross-domain technical terminology mapping; a multi-level analysis of explicit and implicit relationships expands the types of associations between technical entities; and a mechanism for supplementing inferred relationships based on low-density regions enables dynamic optimization of the knowledge structure. These improvements enhance the knowledge graph's support for technology transfer platforms.

[0076] The intent parsing module is used to parse the intentions of both supply and demand sides based on a multi-domain fusion knowledge graph, using a dual decoder mechanism to extract core features and generate intent representation vectors.

[0077] In a preferred embodiment of the present invention, the intention analysis module processing flow chart is as follows Figure 2As shown, it includes: constructing an intention-capability dual decoder, wherein the intention-capability dual decoder includes a technology supplier decoding unit and a demander decoding unit, each decoding unit includes an attention guidance layer, a semantic aggregation layer and a feature mapping layer, and the specific design is as follows: constructing a dual decoder structure, including a technology supplier decoding unit and a demander decoding unit, the two decoding units share the basic network architecture but use independent parameter spaces to process the subgraph structures associated with the technology supplier and the demander respectively, and map the node features of the multi-domain fusion knowledge graph and its associated subgraph structure to a unified semantic space; constructing an attention guidance layer, using a multi-head self-attention mechanism, through the query matrix, key matrix and The value matrix adaptively selects technical capability features and demand intention features, wherein the technology supplier decoding unit extracts core capabilities, technical advantages and application field features based on the first subgraph structure, and the demand side decoding unit extracts key requirements, technical constraints and application scenario features based on the second subgraph structure; constructs a semantic aggregation layer, adopts a multi-layer perceptron structure and a residual connection mechanism to perform local-global nonlinear transformation and fusion on the features in the subgraph structure, and maintains the stability of feature distribution through layer normalization; constructs a feature mapping layer, and adopts a learnable projection matrix to map the supplier capability representation vector and the demand intention representation vector after feature fusion to a unified semantic space of predetermined dimensions.

[0078] Preferably, the intention-capability dual decoder of the present invention has the following technical advantages: first, a dual decoder structure is adopted to replace a single decoder, and the adaptability to the differences in technology supply and demand characteristics is improved by sharing the basic architecture but maintaining independent parameter space; second, a differentiated feature extraction strategy is designed, and the attention mechanism is used to strengthen the core capability characteristics of the supplier and the key demand characteristics of the demander respectively; third, the associated subgraph structure information is integrated into the feature extraction process, and the expression ability of technology correlation is enhanced through local-global feature fusion.

[0079] Furthermore, a first subgraph structure associated with the technology supplier is extracted from the multi-domain fusion knowledge graph, and the first subgraph structure is input into the technology supplier decoding unit to identify the core capabilities, technical advantages and application fields of the technology supplier, and generate a supplier capability representation vector; a second subgraph structure associated with the demander is extracted from the multi-domain fusion knowledge graph, and the second subgraph structure is input into the demander decoding unit to identify the demander's key needs, technical constraints and application scenarios, and generate a demander intention representation vector.

[0080] Furthermore, the technical function decomposition and reorganization algorithm is used to decompose the supplier's capability representation vector to obtain a set of basic functional units and construct a functional unit dependency graph, which specifically includes: using a recursive hierarchical algorithm to divide the functional hierarchy according to the principle of coarse to fine functional granularity, and extracting basic functional units including core functions, performance functions, scenario functions and extended functions, where the recursive hierarchical algorithm adopts a top-down hierarchical clustering method; extracting characteristic parameters for each basic functional unit, constructing a functional unit feature matrix and mapping it to a unified representation space; analyzing the hierarchical relationship, preconditions and mutual exclusion constraints between basic functional units, and establishing a directed acyclic graph structure to represent the functional unit dependency relationship; performing topological sorting based on the functional unit dependency graph, identifying the critical path, and determining the minimum functional unit set, where the minimum functional unit set refers to the set of all functional units on the critical path.

[0081] Furthermore, the demander's intention representation vector is decomposed to obtain a set of basic demand units, and a demand unit priority graph is constructed, specifically including: using a hierarchical decomposition algorithm to divide the hierarchy according to the demand type, extracting basic demand units including core requirements, performance requirements, scenario requirements and extended requirements; extracting attribute features for each basic demand unit, constructing a demand unit feature matrix and mapping it to a unified representation space; analyzing the priority relationship, dependency conditions and conflict constraints between basic demand units, and establishing a directed weighted graph structure to represent the demand unit priority relationship; performing hierarchical sorting based on the demand unit priority graph, identifying the critical demand path, and determining the core demand unit set, where the core demand unit set refers to the set of all demand units on the critical demand path.

[0082] Preferably, the decomposition and reorganization algorithm of the present invention has the following technical effects: through the construction of a recursive hierarchical functional decomposition algorithm and a functional unit dependency graph, a refined expression and structured description of technical capabilities are achieved, and the recognition accuracy of technical features is improved; a demand priority modeling method based on a weighted graph is adopted to identify key demand paths through hierarchical sorting, providing a quantitative basis for demand analysis and enhancing the accuracy of demand matching; the supply-side functional decomposition and the demand-side demand decomposition adopt a unified modeling framework to facilitate the subsequent alignment and matching of supply and demand features.

[0083] Furthermore, the mapping relationship between the basic function unit set and the basic requirement unit set is calculated to obtain the function-requirement matching matrix, and the supplier's capability representation vector and the demander's intention representation vector are fused and reorganized based on the function-requirement matching matrix to generate an intention representation vector, specifically including: calculating the semantic similarity and structural similarity between the basic function unit and the basic requirement unit in the unified representation space, and weighting the semantic similarity and structural similarity based on preset weights to obtain a comprehensive matching degree; constructing a function-requirement matching matrix based on the comprehensive matching degree, normalizing the function-requirement matching matrix, and filtering the low-correlation mapping relationship in the function-requirement matching matrix according to a preset matching threshold; using the function-requirement matching matrix to group the minimum function unit set and the core requirement unit set, extracting high-matching function-requirement pairs, and using the attention mechanism to weightedly fuse the function-requirement pairs to generate an intention representation vector; constructing an interpretable path for the intention representation vector, including recording the mapping relationship corresponding to the function-requirement matching matrix and the weight information in the weighted fusion process.

[0084] Among them, semantic similarity is obtained by calculating the cosine distance of unit feature vectors in the unified representation space, and structural similarity is obtained by calculating the weighted sum of the hierarchical distance and path overlap of corresponding nodes in the functional unit dependency graph and the requirement unit priority graph; the preset weights are set according to the importance of semantic similarity and structural similarity in the expression of technical features, and the matching threshold is determined based on the significance level of the function-requirement mapping relationship; the extraction process of high-matching function-requirement pairs is achieved by executing the bipartite graph maximum weight matching algorithm on the function-requirement matching matrix; the attention mechanism calculates the attention weight based on the comprehensive matching of the function-requirement pair, and performs weighted summation on the functional unit feature vector and the requirement unit feature vector; the interpretability path includes the mapping link from the functional unit to the requirement unit, the corresponding matching value and the attention weight distribution during the fusion process.

[0085] The similarity calculation module is used to calculate the multi-dimensional similarity between intent representation vectors using a multi-level heterogeneous graph network that integrates graph structure analysis and differentiated weight allocation mechanism, and implement dynamic similarity compensation for feature sparse areas.

[0086] Specifically, the similarity calculation module processing flow chart is as follows: Figure 3As shown, it includes: constructing a multi-level heterogeneous graph network, mapping the supplier's capability representation vector and the demander's intention representation vector to the corresponding nodes, specifically including: using a three-layer heterogeneous network structure to construct a multi-level heterogeneous graph network, including a technical field layer, a technical function layer, and a technical parameter layer, each layer of nodes represents the technical characteristics of different levels, and the nodes are connected through corresponding association relationships; in the technical field layer, a field node set is constructed based on the technical field ontology, an edge set is established according to the association relationship between fields, and the field characteristics of the supplier and the demander are mapped to the corresponding nodes; in the technical function layer, a function node set is constructed based on the function unit dependency graph and the demand unit priority graph, and the function node set is constructed according to the association relationship between functions. The dependency relationship establishes an edge set, and maps the functional characteristics of the supplier and demander to the corresponding nodes; at the technical parameter layer, a parameter node set is constructed based on functional attributes, technical indicators and constraints, an edge set is established according to the correlation between parameters, and the parameter characteristics of the supplier and demander are mapped to the corresponding nodes; the inter-layer connection relationship is constructed, the correspondence and influence relationship between nodes at different levels in the three-layer network are analyzed, an inter-layer edge set is established, and the inter-layer edges are given initial weights based on technical knowledge. The inter-layer connection relationship includes mapping edges from the technical field layer to the technical function layer and mapping edges from the technical function layer to the technical parameter layer, where the mapping edge weights are determined based on the statistical co-occurrence frequency in the professional knowledge base.

[0087] Preferably, a multi-level heterogeneous graph network is used to represent technical features, and the technical features of both the supply and demand sides are hierarchically mapped at the technical field layer, technical function layer, and technical parameter layer, and inter-layer associations are established, breaking through the expression limitations of a single feature dimension in existing technologies and improving the integrity of the expression of technical features.

[0088] Furthermore, based on the characteristics of the technical field, matching historical data and technology maturity indicators, combined with the technology life cycle assessment function, the weight coefficients of each feature dimension are dynamically allocated to form an adaptive weight matrix, wherein the technical field characteristics include technology innovation attributes, application scenario scope and technology development direction, which are extracted according to the node attributes of the technical field layer in the multi-level heterogeneous graph network; matching historical data includes the distribution pattern and importance score of technical features in historical successful matching cases, which are statistically analyzed based on the historical matching records of the technical function layer and the technical parameter layer; technology maturity indicators include the degree of technology realization, verification completeness and application popularity, which are obtained by analyzing the node correlation in the multi-level heterogeneous graph network; the technology life cycle assessment function is constructed based on the S-shaped curve of technology development, and the technology maturity indicators are mapped into time-series weight factors; the weighted combination method is used to fuse the above features, and the adaptive weight coefficients of the features of each dimension are calculated according to the feature importance and time-series weight factors, and the weight coefficients are organized into a weight matrix corresponding to the multi-level heterogeneous graph network structure.

[0089] Furthermore, the similarities in three dimensions, namely direct similarity, indirect correlation and complementary adaptability, between mapping nodes in a multi-level heterogeneous graph network are calculated to form a multidimensional similarity vector, specifically including: calculating direct similarity, based on the feature vectors of corresponding nodes in the multi-level heterogeneous graph network, using cosine similarity to calculate the degree of node feature matching between the supply and demand sides in the technical field layer, technical function layer and technical parameter layer; calculating indirect correlation, based on the path connection relationship between nodes in the multi-level heterogeneous graph network, calculating the reachability strength between the nodes of the supply and demand sides by weighting the path distance, and combining the inter-layer connection weights to obtain the degree of correlation; calculating complementary adaptability, based on the combined characteristics of the nodes of the supply and demand sides in the multi-level heterogeneous graph network, by analyzing the dependency completeness of the functional nodes and the constraint satisfaction of the parameter nodes, evaluating the technical matching potential of the supply and demand sides; normalizing the direct similarity, indirect correlation and complementary adaptability, and combining them to form a multidimensional similarity vector that characterizes the degree of technical matching between the supply and demand sides.

[0090] Furthermore, a local density clustering method is used to identify feature-sparse areas in the intent representation vector space, and a multidimensional similarity compensation vector is generated based on the distribution characteristics of historical successful matching cases, specifically including: calculating the local density value of each vector point in the intent representation vector space, including calculating the number of samples within a specified radius with each vector point as the center, and determining the feature distribution density of the area based on the number of samples; setting a feature distribution density threshold to divide the vector space into regions, marking the regions with local density values less than or equal to the feature distribution density threshold as feature-sparse regions, and marking the regions with local density values greater than the feature distribution density threshold as feature-dense regions; screening historical successful matching cases in the feature-dense regions, and extracting the feature distribution laws of these cases, including statistical features such as feature mean, variance and correlation; based on the feature distribution law, performing feature compensation on the vector points in the feature-sparse regions to generate a multidimensional similarity compensation vector, wherein the generation of the compensation vector comprehensively considers the regional density difference, feature distribution variance and feature distribution of adjacent dense regions.

[0091] It's important to note that when the local density values of all regions in the vector space are greater than the feature distribution density threshold, the entire vector space is marked as a feature-dense region, and the resulting multidimensional similarity compensation vector is a zero-compensation vector (all elements are zero). This indicates that feature compensation is unnecessary, and the zero-compensation vector does not change the original similarity value when integrated with the multidimensional similarity vector, but maintains the consistency and integrity of the processing flow.

[0092] Furthermore, a nonlinear weighted fusion method is used to integrate the multidimensional similarity vector and the multidimensional similarity compensation vector, where the weight coefficient is determined by minimizing the error function of historical matching samples to generate a comprehensive similarity score.

[0093] Preferably, the similarity calculation module of the present invention achieves the following technical effects: a dynamic weight allocation mechanism based on the technology life cycle is designed to enhance the adaptability of similarity calculation to technology development trends; a three-dimensional similarity calculation model is constructed to reflect the explicit matching degree of technical features through direct similarity, to mine potential technical associations through indirect correlation, and to evaluate the potential for technical integration through complementary adaptability, thereby improving the recognition accuracy of technical matching; a dynamic compensation method for feature-sparse areas is proposed to solve the problem of similarity calculation deviation caused by uneven feature distribution and improve the accuracy of similarity calculation.

[0094] The recommendation ranking module is used to comprehensively score and sort supply and demand matching pairs based on multi-dimensional similarity, combined with technology maturity and market potential evaluation indicators, and output a personalized recommendation list.

[0095] Specifically, the processing flow of the recommendation ranking module includes: calculating the similarity mean and standard deviation based on the comprehensive similarity distribution of historical matching samples, setting the matching recognition benchmark threshold, and dividing the similarity interval into three matching intervals: high, medium, and low based on the threshold.

[0096] Furthermore, based on the comprehensive similarity score, the preliminary matching relationship set between the technology supplier and the demander is determined. Specifically: a normalized calculation is performed on the comprehensive similarity scores of the technology supply and demand parties, and a preliminary screening is performed using the set matching identification benchmark threshold to construct a candidate matching set; based on the matching distribution variance and skewness of the supply and demand matching pairs in the candidate matching set in each dimension, an adaptive adjustment coefficient is calculated, and the matching identification benchmark threshold is adjusted and optimized in real time; the adjusted matching identification benchmark threshold is applied to perform a secondary screening of the candidate matching set, and finally the preliminary matching relationship set between the supply and demand parties is determined.

[0097] Furthermore, using historical data on technology transfer, a technology transfer feasibility evaluation system is established that includes conversion difficulty indicators, implementation cycle indicators and investment scale indicators; based on the evaluation system, the technology transfer feasibility of the supply and demand matching pairs in the preliminary matching relationship set is quantitatively scored, specifically including: statistical analysis of the conversion difficulty, implementation cycle and investment scale of historical technology transfer projects, and establishment of a hierarchical scoring standard based on technical fields and scale levels; combining the actual situation of both the technology supply and demand sides, evaluating the conversion difficulty from the dimensions of technical complexity, resource and equipment matching degree, and technical team reserves; evaluating the implementation cycle indicators based on project cycle and milestone planning; evaluating the investment scale indicators by comprehensively considering the human, material and financial needs; determining the weight coefficient of each indicator according to the success rate distribution of historical projects, and performing weighted calculation to obtain a comprehensive score for the feasibility of technology transfer.

[0098] Furthermore, a weighted fusion method is used to combine the comprehensive similarity score with the technology transfer feasibility score to generate a final matching score; based on the final matching score, the preliminary matching relationship set is grouped and sorted within the technology field to generate personalized recommendation results.

[0099] Furthermore, this embodiment also provides a big data matching and recommendation method based on a technology transfer platform, including: using a domain adaptive encoder to convert a technical information set from a distributed heterogeneous data source into a standardized vector representation; extracting technical entities and relationships in the standardized vector representation through a pre-trained large-scale language model to construct a multi-domain fusion knowledge graph; based on the multi-domain fusion knowledge graph, using a dual decoder mechanism to analyze the intentions of both technology supply and demand parties, extract core features and generate intention representation vectors; using a multi-level heterogeneous graph network that integrates graph structure analysis and differentiated weight allocation mechanism to calculate the multi-dimensional similarity between intention representation vectors, and implement dynamic similarity compensation for feature sparse areas; based on the multi-dimensional similarity, the matching relationship between technology suppliers and demanders is identified, and combined with the technology transfer feasibility indicators for comprehensive scoring and ranking to generate personalized recommendation results.

[0100] In summary, the present invention realizes the unified expression and organization of heterogeneous data through domain-adaptive encoders and multi-level ontology index structures, thereby improving the accuracy and efficiency of technical information processing; the multi-domain fusion knowledge graph constructed based on dual-threshold judgment and dynamic optimization mechanism enhances the ability to discover technical associations and improves the integrity of knowledge representation; the dual-decoder structure and hierarchical decomposition and recombination algorithm are adopted to realize the precise analysis and feature alignment of technical supply and demand intentions, and enhance the interpretability of the matching process; the designed three-dimensional similarity calculation model and dynamic compensation mechanism effectively improve the accuracy and robustness of technical matching, and enhance the system's adaptability to changes in technological development; through multi-dimensional feasibility evaluation and adaptive sorting mechanism, the accuracy and practicality of recommendation results are improved, providing a reliable basis for technology transfer decisions.

[0101] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A big data matching recommendation system based on a technology transfer platform, characterized in that: include: A data processing module, configured to convert a technical information set from distributed heterogeneous data sources into a standardized vector representation using a domain-adaptive encoder; A knowledge graph construction module, configured to extract technical entities and relationships in the standardized vector representations through a pre-trained large-scale language model, and construct a multi-domain fusion knowledge graph; An intent parsing module, which is used to parse the intentions of both the supply and demand sides based on the multi-domain fusion knowledge graph using a dual decoder mechanism, extract core features and generate an intent representation vector; A similarity calculation module is used to calculate the multi-dimensional similarity between intent representation vectors using a multi-level heterogeneous graph network that integrates graph structure analysis and a differentiated weight allocation mechanism, and to implement dynamic similarity compensation for feature-sparse regions. A recommendation ranking module is used to identify the matching relationship between technology suppliers and demanders based on the multi-dimensional similarity, perform comprehensive scoring and ranking based on market potential evaluation indicators, and generate personalized recommendation results; The processing flow of the similarity calculation module includes: Construct a multi-level heterogeneous graph network and map the supply-side capability representation vector and the demand-side intention representation vector to corresponding nodes; Based on the characteristics of the technical field, matching historical data and technology maturity indicators, the weight coefficients of each feature dimension are dynamically assigned to form an adaptive weight matrix; Calculate the similarity between mapping nodes in a multi-level heterogeneous graph network in three dimensions: direct similarity, indirect correlation, and complementary adaptability, and form a multidimensional similarity vector; Identify feature-sparse regions in the intent representation vector space and generate multi-dimensional similarity compensation vectors; The multidimensional similarity vector is integrated with the multidimensional similarity compensation vector to generate a comprehensive similarity score.

2. The big data matching recommendation system based on the technology transfer platform according to claim 1 is characterized in that: The processing flow of the data processing module includes: Preprocess the technical information sets in distributed heterogeneous data sources and construct a multi-level technical ontology index structure; Processing structured data, semi-structured data, and unstructured data in the technical information set to extract a set of technical terms, a set of functional descriptions, and a set of performance indicators; Construct a domain-adaptive encoder, including a domain weight parameter matrix, a cross-domain mapping function, and a semantic preservation loss function; Training the domain adaptive encoder to encode the set of technical terms, the set of functional descriptions, and the set of performance indicators into initial vector representations; The initial vector representation is subjected to dimensionality reduction processing to obtain a standardized vector, and a bidirectional traceable mapping mechanism from the technical description text to the standardized vector is established.

3. The big data matching recommendation system based on the technology transfer platform according to claim 1 is characterized in that: The processing flow of the knowledge graph construction module includes: The standardized vector is converted into technical description text through a bidirectional traceability mapping mechanism, and technical entity recognition is performed through a pre-trained large-scale language model to extract the technical entity set; Constructing a domain terminology corresponding structure, performing domain adaptive transformation on the technical entity set, and forming a cross-domain technical terminology mapping relationship; By multi-level technical relationship analysis, explicit and implicit relationships between technical entities are extracted from the technical entity set to form a technical relationship set; Using the technical entity set as nodes and the technical relationship set as edges, constructing an initial graph structure, and assigning weight values to the nodes and edges in the initial graph structure; The initial graph structure is adjusted according to feedback information of preset technology matching samples, and inference relationships are added to low-density areas of the initial graph structure to form a multi-domain fusion knowledge graph.

4. The big data matching recommendation system based on the technology transfer platform according to claim 1 is characterized in that: The processing flow of the intent parsing module includes: Constructing an intent-capability dual decoder, wherein the intent-capability dual decoder includes a technology supply-side decoding unit and a demand-side decoding unit; Extracting a first subgraph structure associated with a technology supplier from the multi-domain fusion knowledge graph, inputting the first subgraph structure into the technology supplier decoding unit, and generating a supplier capability representation vector; A second subgraph structure associated with the demander is extracted from the multi-domain fusion knowledge graph, and the second subgraph structure is input into the demander decoding unit to generate a demander intention representation vector.

5. The big data matching recommendation system based on the technology transfer platform according to claim 1 is characterized in that: The processing flow of the intent parsing module also includes: The technology function decomposition and reorganization algorithm is used to decompose the supplier's capability representation vector, obtain a set of basic functional units, and construct a functional unit dependency graph; Decompose the demander's intention representation vector to obtain a set of basic demand units and construct a demand unit priority graph; Calculate the mapping relationship between the basic function unit set and the basic requirement unit set to obtain a function-requirement matching matrix, and fuse and reorganize the supplier's capability representation vector and the demander's intention representation vector based on the function-requirement matching matrix to generate an intention representation vector.

6. The big data matching recommendation system based on the technology transfer platform according to claim 1 is characterized in that: The processing flow of the recommendation ranking module includes: Determine the initial matching relationship set between technology suppliers and demanders based on the comprehensive similarity score, and establish a technology transfer feasibility assessment system; Quantitatively scoring the technology transfer feasibility of each pair of supply and demand parties in the preliminary matching relationship set based on the technology transfer feasibility evaluation system; A weighted fusion method is used to combine the comprehensive similarity score with the technology transfer feasibility score to generate the final matching score; The preliminary matching relationship set is grouped and sorted within the technical field according to the final matching score to generate personalized recommendation results.

7. A big data matching recommendation method based on a technology transfer platform, characterized in that: include: A domain-adaptive encoder is used to convert technical information sets from distributed heterogeneous data sources into standardized vector representations; Extracting technical entities and relationships in the standardized vector representation through a pre-trained large-scale language model to construct a multi-domain fusion knowledge graph; Based on the multi-domain fusion knowledge graph, a dual decoder mechanism is used to analyze the intentions of both the supply and demand sides, extract core features and generate an intention representation vector; A multi-layered heterogeneous graph network that integrates graph structure analysis and differentiated weight allocation is used to calculate the multi-dimensional similarity between intent representation vectors and implement dynamic similarity compensation for feature-sparse regions. Based on the multi-dimensional similarity, the matching relationship between technology suppliers and demanders is identified, and a comprehensive score and ranking are performed in combination with the technology transfer feasibility index to generate personalized recommendation results; The dynamic similarity compensation for the feature sparse area includes: Construct a multi-level heterogeneous graph network and map the supply-side capability representation vector and the demand-side intention representation vector to corresponding nodes; Based on the characteristics of the technical field, matching historical data and technology maturity indicators, the weight coefficients of each feature dimension are dynamically assigned to form an adaptive weight matrix; Calculate the similarity between mapping nodes in a multi-level heterogeneous graph network in three dimensions: direct similarity, indirect correlation, and complementary adaptability, and form a multidimensional similarity vector; Identify feature-sparse regions in the intent representation vector space and generate multi-dimensional similarity compensation vectors; The multidimensional similarity vector is integrated with the multidimensional similarity compensation vector to generate a comprehensive similarity score.

Images