GCN-based two-stage feature construction and link prediction method and device

Abstract

The invention discloses a GCN-based two-stage feature construction and link prediction method and device, and the method comprises the steps: obtaining a life cycle of a preset technical field in multi-source data, and constructing an LDA model, a technical topic co-occurrence network and a prediction model; according to the technical opportunity recognition method, a prediction model, a market demand degree, technical competitiveness and a policy matching degree are combined to construct a technical opportunity multi-dimensional evaluation system, a technical innovation direction of the technical opportunity multi-dimensional evaluation system is researched and judged by combining sensitivity analysis, and multi-dimensional index evaluation and sensitivity analysis are combined. The method can efficiently and accurately predict new technical opportunities which may occur in the future, provides more comprehensive information coverage for a technical topic co-occurrence network, avoids the problems of co-occurrence relation redundancy, key relation screening and the like, and realizes multi-dimensional attribute embedding of nodes and enhancement of model recognition capability by integrating network semantic association and topological structure features, so that a prediction model is more efficient and accurate. And similarity indexes and machine learning refinement processing are combined, so that the accuracy and robustness of the model are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of technology opportunity identification, and in particular, relates to a two-stage feature construction and link prediction method and device based on GCN. Background Art

[0002] Technological opportunities represent a series of possibilities and potentials for technological progress in a specific field or area, and they often become the key triggers for restructuring the current industry or market competitive advantage. In order to clarify potential research directions, technological opportunity identification is the first step in the technology research and development process. Technological opportunity identification refers to the process of identifying, deriving and acquiring new technological ideas with potential value in a specific technology field. This process not only requires mining patent, literature and media data, identifying complex correlation features, and predicting potential technological associations through network modeling, but also requires adapting technology, market and policy environment information to achieve breakthrough innovation and enhance competitive advantage. Therefore, it is crucial to integrate multi-source information, extract network features, and predict and evaluate network associations for the process of technological opportunity identification.

[0003] Against the backdrop of deepening technological innovation and industrial transformation, the pace of technological innovation is accelerating, and it is increasingly important to use scientific and efficient research methods to identify technological opportunities. Early technology opportunity identification methods were mainly based on qualitative analysis, such as the Delphi method, scenario planning, and analytic hierarchy process, but the analysis results were often limited by expert experience and knowledge domains. With the rapid development of information technology, a large amount of scientific and technological literature, including papers and patents, has condensed a large amount of technical data for analysis. A large number of scholars have begun to propose quantitative technology opportunity identification methods based on text mining, and have identified technology opportunities through paper and patent text analysis, and have produced a large number of related research results. For example, by conducting community analysis and main path analysis on patent citation networks, the direction of technological development in the field of autonomous vehicles was identified; IPC co-occurrence networks and patentee citation networks were drawn, and emerging technologies in the field of autonomous vehicles were identified by calculating various network indicators. However, most of the related studies using patent citations and IPC classification numbers have time lags or coarse-grained defects.

[0004] In this regard, a network is constructed around keywords to mine technology opportunities. For example, a method of using a keyword network combined with an ant colony algorithm to identify potential technology opportunities through the optimal path in the network; a technology keyword network is constructed based on paper data, and a graph representation learning method is used to identify potential disruptive technologies. However, due to the large number of co-occurrence relationships in the technology network based on keywords, how to solve relationship redundancy and screen key relationships is still a difficult problem in research, and the semantic associations between words are usually ignored, and the technical characteristics cannot be fully expressed. The above provides a rich perspective and method for identifying technology opportunities, but the defect that only papers and patent texts cannot directly reflect the characteristics and capabilities of the technology itself is becoming increasingly prominent, and the network nodes built solely based on scientific and technological literature have poor generalization ability and interpretability. In contrast, the theme network can usually capture semantic associations in a fine-grained manner, and the technology opportunities it represents are finer in granularity and more interpretable. At the same time, due to its dynamic and real-time nature, the rich social media data not only covers market feedback and the latest technology trends, but also can continuously track the development trajectory of technology opportunities, providing a more accurate analytical perspective for in-depth understanding of the market adaptability, user acceptance and development potential of technology.

[0005] In terms of network feature extraction, with the development of artificial intelligence technology, more and more methods are combined with natural language processing, machine learning, social network analysis, etc. Among them, methods represented by link prediction can most directly reflect the correlation between knowledge. They mainly start from the technical knowledge network in a specific field and calculate the "IPC pairs" or "keyword pairs" in the network that are most likely to generate links in the future through link prediction indicators. At present, link prediction methods can be divided into heuristic methods and representation learning-based methods. Heuristic methods mainly predict links based on the calculation of node similarity scores based on network topology structures, and can be divided into methods based on local information and methods based on path information. Related methods include AA, RA, PA, CN, Katz, Salton, etc. However, these methods only rely on local features, and their performance in sparse networks is often unsatisfactory, making it difficult to deeply mine the deep information of the network. In addition, path-based methods are highly dependent on the network and are more sensitive to noise. Representation learning methods effectively avoid the calculation method of predefined similarity between nodes. From the perspective of graph topology features and attribute information, they can be divided into decomposition-based methods, random walk-based methods, and graph neural network-based methods. Specific methods include DeepWalk, node2vec, GCN, GAE, VGAE, etc. Among them, the matrix decomposition-based method divides the adjacency matrix or similarity matrix of the graph to capture the potential relationship between nodes; the random walk-based method, such as using DeepWalk for network representation learning, uses the learned local neighborhood structure for link prediction tasks to identify technology opportunities in scientific literature; a link prediction method based on Node2vec is proposed to create a multi-dimensional label from the perspectives of internal enterprise technology endowment and external innovation environment to identify Huawei's technology opportunities. These methods can overcome the limitations of traditional methods to a certain extent, but they mainly rely on the topological characteristics of the network and often lack in-depth mining of attribute information and semantic associations between nodes. Summary of the invention

[0006] In view of this, the present invention provides a two-stage feature construction and link prediction method and device based on GCN, which combines the technology opportunity identification method of multi-dimensional indicator evaluation and sensitivity analysis to efficiently and accurately predict new technology opportunities that may appear in the future. On the one hand, multi-source data is used to provide more comprehensive information coverage for the constructed technology theme co-occurrence network, making technology opportunities more explanatory and more granular, while also avoiding problems such as co-occurrence relationship redundancy and key relationship screening. On the other hand, the two-stage feature construction link prediction model integrates the characteristics of network semantic associations and topological structures to achieve multi-dimensional attribute embedding of nodes, combined with similarity indicators and refined processing of machine learning, effectively improving the accuracy and robustness of the model, and is specifically implemented using the following technical solutions.

[0007] In a first aspect, the present invention provides a two-stage feature construction and link prediction method based on GCN, comprising the following steps:

[0008] Acquire the life cycle of a preset technical field in multi-source data, and construct an LDA model of the life cycle, wherein the LDA model includes a document topic model based on probability and statistical methods;

[0009] Constructing a technical theme co-occurrence network of the preset technical field according to the LDA model, and constructing a two-stage feature construction link prediction model based on GNC according to the technical theme co-occurrence network, wherein the two-stage feature construction link prediction model based on GNC includes node embedding and edge feature aggregation based on GCN;

[0010] A multidimensional evaluation system for technology opportunities is constructed by combining the prediction model, market demand, technological competitiveness and policy matching, and the technological innovation direction of the multidimensional evaluation system for technology opportunities is determined in combination with sensitivity analysis to complete the two-stage feature construction and link prediction.

[0011] In a second aspect, the present invention further provides a two-stage feature construction and link prediction device based on GCN, which is applied to the above-mentioned two-stage feature construction and link prediction method based on GCN, including:

[0012] An LDA model building unit, used to obtain the life cycle of a preset technical field in multi-source data and build an LDA model of the life cycle, wherein the LDA model includes a document topic model based on probability and statistical methods;

[0013] A prediction model construction unit, used to construct a technical theme co-occurrence network of the preset technical field according to the LDA model, and to construct a two-stage feature construction link prediction model based on GNC according to the technical theme co-occurrence network, wherein the two-stage feature construction link prediction model based on GNC includes node embedding and edge feature aggregation based on GCN;

[0014] The two-stage feature construction and link prediction unit is used to construct a multi-dimensional evaluation system for technology opportunities by combining the prediction model, market demand, technology competitiveness and policy matching, and to determine the technology innovation direction of the multi-dimensional evaluation system for technology opportunities by combining sensitivity analysis to complete the two-stage feature construction and link prediction.

[0015] The present invention provides a two-stage feature construction and link prediction method and device based on GCN, which obtains the life cycle of a preset technical field in multi-source data and constructs an LDA model of the life cycle, constructs a technical theme co-occurrence network of the preset technical field according to the LDA model, and constructs a two-stage feature construction link prediction model based on GNC according to the technical theme co-occurrence network, constructs a multi-dimensional evaluation system of technical opportunities in combination with the prediction model, market demand, technical competitiveness and policy matching, and judges the technical innovation direction of the multi-dimensional evaluation system of technical opportunities in combination with sensitivity analysis to complete the two-stage feature construction and link prediction, and combines multi-dimensional indicator evaluation and sensitivity analysis to construct a link prediction model based on GNC. The technology opportunity identification method can efficiently and accurately predict new technology opportunities that may appear in the future. It uses multi-source data to provide more comprehensive information coverage for the constructed technology theme co-occurrence network, making technology opportunities more interpretable and more granular, while also avoiding problems such as co-occurrence relationship redundancy and key relationship screening. The two-stage feature construction link prediction model integrates the characteristics of network semantic associations and topological structures to achieve multi-dimensional attribute embedding of nodes and enhancement of model recognition capabilities. Combined with similarity indicators and refined processing of machine learning, it effectively improves the accuracy and robustness of the model, provides multi-dimensional expression for technology opportunities, and further improves the accuracy, effectiveness and practicality of the identification process. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 It is a flow chart of the two-stage feature construction and link prediction method based on GCN of the present invention;

[0018] Figure 2 It is a logical structure diagram of the method of the present invention;

[0019] Figure 3 A two-stage feature construction flow chart for GCN-based link prediction of the present invention;

[0020] Figure 4 It is a schematic diagram of the construction of the Topic-Topic adjacency matrix of the present invention;

[0021] Figure 5 Evaluate the potential technical opportunities of the invention into a visualization diagram;

[0022] Figure 6 A schematic diagram of the life cycle division of new energy vehicle technology of the present invention;

[0023] Figure 7 This is a schematic diagram of the co-occurrence network of the embryonic technical themes of the present invention;

[0024] Figure 8 A schematic diagram of a co-occurrence network of technical themes in the development period of the present invention;

[0025] Fig. 9 This is a schematic diagram of the co-occurrence network of the growth-stage technical themes of the present invention;

[0026] Fig.10 This is a schematic diagram of the co-occurrence network of mature technical themes of the present invention;

[0027] Fig.11 It is a schematic diagram of the ROC curve and PR curve of the model on the mature network of the present invention and the optimal threshold point;

[0028] Fig.12 It is a word cloud diagram of each functional module of the present invention;

[0029] Fig.13 A schematic diagram of a potential technology opportunity distribution diagram of the present invention;

[0030] Fig.14 It is a schematic diagram of the single factor sensitivity analysis of the present invention;

[0031] Fig.15 It is a schematic diagram of the potential technology opportunity impact analysis diagram of the present invention;

[0032] Fig.16 It is a structural block diagram of the two-stage feature construction and link prediction device based on GCN of the present invention. DETAILED DESCRIPTION

[0033] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0034] By comprehensively considering the internal and external environmental factors of the organization, designing multidimensional evaluation indicators and conducting sensitivity analysis, we can provide a multidimensional expression of technological opportunities and further improve the accuracy, effectiveness and practicality of the identification process, aiming to provide a theoretical reference for innovative entities to carry out technological innovation activities scientifically.

[0035] See also Figure 1 The present invention provides a two-stage feature construction and link prediction method based on GCN, comprising the following steps:

[0036] S1: Obtaining the life cycle of a preset technical field in multi-source data, and constructing an LDA model of the life cycle, wherein the LDA model includes a document topic model based on probability and statistical methods;

[0037] S2: constructing a technical theme co-occurrence network of the preset technical field according to the LDA model, and constructing a two-stage feature construction link prediction model based on GNC according to the technical theme co-occurrence network, wherein the two-stage feature construction link prediction model based on GNC includes node embedding and edge feature aggregation based on GCN;

[0038] S3: Construct a multidimensional evaluation system for technology opportunities by combining the prediction model, market demand, technology competitiveness and policy matching, and determine the technological innovation direction of the multidimensional evaluation system for technology opportunities by combining sensitivity analysis to complete the two-stage feature construction and link prediction.

[0039] In this embodiment, the core purpose of multi-source data representation learning is to integrate data from different sources and use their consistency and complementarity to achieve efficient data mining. In this process, machine learning methods can identify representative and interpretable technical themes in multi-source data texts and then deeply explore the technical potential and development trends in specific research fields. Technical themes, as the basic units for characterizing technical information, explain the knowledge structure of the technical field to a certain extent, and are crucial for understanding technological evolution and identifying emerging technological opportunities. Therefore, the present invention is based on multi-source data and is based on constructing a technical theme co-occurrence network, aiming to develop a set of multi-dimensional technology opportunity identification methods. Specifically, this study will adopt a two-stage feature construction and link prediction method based on graph convolutional networks (GCN) to monitor the dynamic process of technology convergence. In addition, the applicability of technical opportunities is analyzed through multi-dimensional evaluation indicators, and sensitivity analysis is used to deeply characterize the application scenarios of technical opportunities. For a detailed display of this logical structure, see Figure 2 .

[0040] It should be noted that the overall process of the proposed technology opportunity identification path can be divided into the following three parts:

[0041] Firstly, patent, literature and information data are collected, and the year and cumulative amount in the patent data set are used to fit the Logistic growth curve to determine the life cycle of the new energy vehicle technology field. The LDA model is further used to extract topics and construct the technical topic co-occurrence network of each stage.

[0042] Secondly, in the first stage, the model aims to solve the sparsity and diversity of node information by combining node clustering with comprehensive modeling of structural features, while capturing global semantic information and topological features; the second stage focuses on refining the direct connections and potential interactions between node pairs to enhance the model's ability to express edge features. By deeply characterizing the intrinsic structure and deep semantic representation of nodes and their edges, the model aims to improve the accuracy and robustness of predictions. The effectiveness of the model will be verified on networks at past stages of the technology life cycle, and will be applied to networks at the current stage to predict future technology opportunities;

[0043] Finally, by comprehensively considering key indicators such as market demand, technological competitiveness, and policy matching, we construct a multi-dimensional evaluation system for technological opportunities. Combined with sensitivity analysis, we further analyze the direction of technological innovation, reveal its performance in different application scenarios, and help organizations flexibly adjust their strategies to adapt to environmental changes.

[0044] Among them, Latent Dirichlet Allocation (LDA) is a document topic model based on probability and statistical methods. LDA constructs a Bayesian probability model with a three-layer structure of "word-topic-document" to explore potential topic structures from a large number of documents. The model can automatically extract technical topics and keywords and achieve efficient clustering of documents and terms. Therefore, in order to refine the granularity and enhance the semantic representation, the LDA model is used to extract technical topics, and then identify and analyze the technological evolution and development trends in a specific research field. The process of extracting technical topics based on the LDA model is as follows: First, determine the hyperparameters α, β and the number of topics K, randomly assign topics to each word w, and initialize the topic assignment matrix z; secondly, iteratively update through Gibbs sampling, calculate the conditional probability of each word in document D being assigned to each topic, and reassign the topic; after multiple iterations, the document-topic distribution θ and topic-word distribution φ are finally obtained. LDA needs to determine the number of topics in advance to identify text topics, and comprehensively judge the model performance under different numbers of topics by combining the perplexity index and the coherence score. The number of topics with a low perplexity index and a high coherence score is taken as the optimal number of topics K, and the LDA model is trained. The LDA clustering effect is comprehensively evaluated in combination with the opinions of domain experts, and then technical topics are extracted.

[0045] It should be understood that by obtaining the life cycle of the preset technology field in multi-source data and constructing the LDA model of the life cycle, constructing the technology theme co-occurrence network of the preset technology field according to the LDA model, and constructing a two-stage feature construction link prediction model based on GNC according to the technology theme co-occurrence network, combining the prediction model, market demand, technology competitiveness and policy matching to construct a multi-dimensional evaluation system for technology opportunities, and combining sensitivity analysis to determine the technology innovation direction of the multi-dimensional evaluation system for technology opportunities to complete the two-stage feature construction and link prediction, the technology opportunity identification method combining multi-dimensional indicator evaluation and sensitivity analysis can efficiently and accurately predict new technology opportunities that may appear in the future, using multi-source data to provide more comprehensive information coverage for the constructed technology theme co-occurrence network, making technology opportunities more interpretable and more granular, while also avoiding problems such as co-occurrence relationship redundancy and key relationship screening, the two-stage feature construction link prediction model integrates the characteristics of network semantic association and topological structure to achieve multi-dimensional attribute embedding of nodes and enhancement of model recognition ability, combined with similarity indicators and refined processing of machine learning, effectively improves the accuracy and robustness of the model, provides multi-dimensional expression for technology opportunities, and further improves the accuracy, effectiveness and practicality of the recognition process.

[0046] Optionally, obtaining the life cycle of the new energy technology field in multi-source data and constructing an LDA model of the life cycle includes:

[0047] The LDA model includes a Bayesian probability model of a three-layer structure of word-topic-document, and explores potential topic structures from documents. The LDA model is used to extract technical topics and keywords to achieve efficient clustering of documents and terms. The process of extracting technical topics by the LDA model includes:

[0048] Get the hyperparameters α, β and the number of topics K, randomly assign a topic to each word w and initialize the topic assignment matrix z;

[0049] Iterative updates are performed through Gibbs sampling to calculate the conditional probability of each word in document D being assigned to each topic, and the topics are reallocated. After multiple iterations, the document-topic distribution θ and topic-word distribution are obtained.

[0050] The perplexity index and consistency score are combined to judge the model performance under different numbers of topics. The number of topics with the lowest perplexity index and the highest consistency score is taken as the optimal number of topics K. The LDA model is trained and the opinions of domain experts are collected to comprehensively evaluate the clustering effect of the LDA model to extract the corresponding technical topics.

[0051] In this embodiment, a technical theme co-occurrence network in the field of new energy technology is constructed according to the LDA model, including: the first stage: constructing a feature matrix from the two perspectives of semantics and structure, inputting the feature matrix together with the adjacency matrix with self-loops added into GCN for node embedding learning, and capturing global information by layer-by-layer convergence during the propagation process; the second stage: constructing variable features using binary operators to enhance the edge feature expression ability of the model, combining the edge features generated by GCN embedding with the local similarity index features to form a comprehensive edge feature matrix; defining positive links and negative links in the constructed technical theme co-occurrence network, defining the link between two themes as a positive link, and defining the link that does not exist between two themes as a negative link, and selecting an undersampling method to balance the number of positive links and negative links.

[0052] It should be noted that if Figure 3 ,Since the attribute information of nodes is often sparse (some technical topics lack clear attributes), and the topological information between nodes may also be sparse (such as the cross-relationship between technical topics in different fields), in the first stage, the model focuses on the multi-dimensional construction of node features, constructing a feature matrix from the perspectives of semantics (technical topic semantic information extracted by Word2Vec) and structure (topological attributes such as degree centrality and clustering coefficient), and then inputs this feature matrix together with the adjacency matrix with self-loops into GCN for node embedding learning. In the propagation process, global information is captured by layer-by-layer convergence, so that the embedding of each node not only contains its own features, but also reflects the semantic and structural information of its neighboring nodes, dispersing data sparsity and diversity problems.

[0053] However, relying solely on the node embedding in the first stage cannot directly describe the strength of the relationship between node pairs. Therefore, in the second stage, the edge features are constructed using binary operators to further enhance the model's edge feature expression capabilities. In addition, the model also combines the edge features generated by GCN embedding with the local similarity index features to form a comprehensive edge feature matrix, which provides the final machine learning link prediction model with high-dimensional semantic and structural dual representation input features, thereby more accurately identifying potential technical associations in the network.

[0054] Optionally, the feature matrix and the adjacency matrix with self-loops added are input into GCN for node embedding learning, including:

[0055] The construction of node features and the generation of embedded representations are performed through convolution operations to learn the connection information of neighboring nodes. The embedding vector of the node is induced according to the neighborhood attributes of the node. A graph G(V,E) is preset, where V(|V|=n) and E are the sets of nodes and edges respectively. The new k-dimensional node feature matrix is ​​calculated based on the single-layer GCN as follows:

[0056]

[0057] Among them, H (l) is the node feature matrix of the lth layer, A is the adjacency matrix with self-loops added, D is the diagonal matrix of node degrees, and W (l) is the learnable weight matrix of the lth layer, σ is the activation function;

[0058] GCN embedding includes an adjacency matrix and a feature matrix. The adjacency matrix is ​​an n×n square matrix used to represent the connection relationship between nodes in the graph. If there is an edge between two nodes, the corresponding element in the adjacency matrix is ​​represented by 1; if there is no connection between two nodes, the corresponding element in the adjacency matrix is ​​0;

[0059] The construction of the adjacency matrix includes:

[0060] Starting from the document-topic probability matrix D output by the LDA model, the matrix P is obtained by filtering low-probability topics by setting the Top-N standard and threshold;

[0061] The co-occurrence relationship between topics is determined according to the calculation formula (2) of the co-occurrence intensity, and the document-topic probability matrix D is converted into the trotter-topic matrix C;

[0062] The co-occurrence intensity is converted into a binary relationship to obtain the adjacency matrix U for GCN. When GCN performs feature propagation, the adjacency matrix is ​​updated by adding self-loops according to the characteristics of the node itself, and all elements on the diagonal of the adjacency matrix are set to 1. The calculation formula (2) is:

[0063]

[0064] Among them, W ij is the co-occurrence intensity between topic i (i∈(1,2,...,M)) and topic j (j∈(1,2,...,M)), T ai is the topic i in the a(a∈(1,2,...,N))th document, P(T ai ) is the probability value of topic i appearing in document a in the document-topic matrix, P(T aj ) is the probability value of topic j appearing in document a in the document-topic matrix.

[0065] In this embodiment, in order to enable the model to integrate the high-order structural features of the node itself and its neighbors as well as the semantic and attribute features of the node, and to improve the expressive power of the node representation, in the first stage, we will focus on the construction of node features and the generation of embedded representations. The connection information of neighboring nodes is learned through convolution operations. GCN is a multi-layer neural network that induces the embedding vector of the node according to the neighborhood attributes of the node. In order to realize GCN embedding, two key matrices need to be prepared: the adjacency matrix and the feature matrix. The adjacency matrix is ​​an n×n square matrix used to represent the connection relationship between nodes in the graph. If there is an edge between two nodes, the corresponding element in the matrix is ​​represented by 1; if there is no connection, the element is 0. In order to realize GCN embedding, two key matrices need to be prepared: the adjacency matrix and the feature matrix. The adjacency matrix is ​​an n×n square matrix used to represent the connection relationship between nodes in the graph. If there is an edge between two nodes, the corresponding element in the matrix is ​​represented by 1; if there is no connection, the element is 0. As for the construction of the adjacency matrix, first start from the document-topic probability matrix D output by the LDA model, and filter low-probability topics by setting the Top-N standard and threshold to obtain the matrix P.

[0066] It should be noted that Figure 4 The calculation process of the contribution intensity of Topic 1 and Topic 2 in matrix C is as follows: First, calculate the co-occurrence probability of Topic 1 and Topic 2 in each document by multiplying the probabilities of the corresponding topics in the original matrix. For document 2, this value is 0.34×0.36=0.1224, and for document 3, this value is 0.35×0.45=0.1575; then add the co-occurrence probabilities of all patent documents to obtain the total co-occurrence intensity of Topic 1 and Topic 2, that is, 0.1224+0.1575=0.2799.

[0067] Optionally, the process of constructing the feature matrix includes:

[0068] The Word2Vec model is used to represent words as vectors of a specified length and capture the semantic relationship between words. The capture tools include the CBOW model and the Skip-Gram model.

[0069] First, use the Skip-Gram model to extract the 32-dimensional feature vector of each technical topic node to obtain the semantic information between nodes;

[0070] The importance of nodes is evaluated by using multiple indicators that reflect the structural characteristics of nodes, such as degree centrality, closeness centrality, betweenness centrality, clustering coefficient, local clustering coefficient, node betweenness, and node efficiency;

[0071] Each feature is concatenated to form a node feature matrix and then normalized.

[0072] In the GCN model, two hidden GCNconv layers are used to process graph data. The first hidden layer is followed by a ReLU activation function to increase the nonlinear expression ability of the model. The output of the second hidden layer is directly used to calculate the loss. The loss function uses BCEWithLogitsLoss or BCEW. The loss function BCEWithLogitsLoss includes a sigmoid activation function. The sigmoid activation function is used to interpret the model output as probability. The expression of the GCN model is:

[0073] Z = AReLU(AXW (0) )W (1) (3)

[0074] in, is the normalized adjacency matrix, W (0) ∈X N×H is the trainable weight matrix of the first layer, H is the dimension of the hidden layer, W (1( is the node embedding dimension after the first layer of convolution operation, X is the node feature matrix, and Z is the feature matrix of the final output.

[0075] In this embodiment, a feature matrix is ​​constructed as the input of GCN. The Word2Vec model [35,36] can represent words as vectors of specified length and capture the semantic relationship between words. The tool includes two models: CBOW model and Skip-Gram model. First, the Skip-Gram model is used to extract the 32-dimensional feature vector of each technical topic node to obtain the semantic information between nodes; secondly, the importance of the node is evaluated, and multiple indicators reflecting the node structure characteristics are used, such as degree centrality, closeness centrality, betweenness centrality, clustering coefficient, local clustering coefficient, node betweenness and node efficiency; thirdly, the above features are concatenated to form a node feature matrix and normalized. In the GCN model of this article, two hidden GCNconv layers are used to process graph data; the first hidden layer is followed by a ReLU activation function, which helps to increase the nonlinear expression ability of the model. Since this is a binary classification problem, that is, predicting whether an edge exists, no additional activation function is added after the second hidden layer, and the output is directly used to calculate the loss. The loss function used is BCEWithLogitsLoss, which contains a sigmoid activation function, which allows the model output to be interpreted as probability.

[0076] Optionally, the features of the constructed edges are represented based on the node embedding representation obtained in the first stage, and the node embedding vector H is obtained. i Based on the representation, the binary operator Hadamard product is applied to embed the two nodes corresponding to each edge into the embedding vector h i 、hj Multiply element by element to get the edge feature matrix h ij , where the node embedding vector H i The dimension is N×D, where N is the number of nodes and the feature matrix h ij The dimension is M×D, where M is the number of node pairs;

[0077] Use multiple local similarity indicators to construct the feature matrix of the edge. The local similarity indicators are used to capture the similarity and relationship strength between nodes to measure the similarity and integration trend between technologies. The feature matrix E of the edge is calculated and each feature is normalized using the StandardScaler function. The dimension of the feature matrix E is M×F, where F is the number of node pairs with similarity and relationship strength.

[0078] The edge feature matrix h ij Concatenate with E to form a comprehensive feature matrix E combined .

[0079] In this embodiment, in order to further capture the direct relationship and interaction between node pairs and enhance the model's ability to identify technology convergence trends, at this stage, the edge features are further constructed based on the node embedding representation obtained in the first stage, aiming to improve the accuracy and reliability of link prediction by in-depth analysis of node and edge features. Technology similarity or technology relevance is an important measure of technology convergence. Ten local similarity indicators are used to further construct the edge feature matrix, including CN, Salton, Jaccard, Sorenson, HPI, HDI, LHN-1, PA, AA, and RA. The above indicators capture the similarity and relationship strength between nodes in different ways to measure the similarity and fusion trends between technologies. The edge feature matrix E can be obtained by calculation, and each feature is normalized using StandardScaler. The above edge feature matrix h ij Concatenate with E to form a comprehensive feature matrix E combined ,Prepared for the subsequent link prediction task, this comprehensive feature matrix integrates node embedding representation and multiple similarity features, covering local and global information, and providing input features with both semantic and structural dual representation for the subsequent link prediction task.

[0080] It should be noted that in order to perform link prediction, it is first necessary to define positive links and negative links for the constructed topic network. In the network, if there is a link between two topics, the link is considered a positive link; on the contrary, if there is no link, it is defined as a negative link. In fact, positive links represent the "actually existing" connections in the network, and they are real. In the face of the imbalance problem of positive and negative link data in the network, oversampling and undersampling strategies are usually adopted. Since positive links reflect the actual existing data, the undersampling method is selected to balance the number of positive links and negative links. In order to comprehensively evaluate the model performance, a variety of machine learning algorithms are used for comparative training, including support vector machine (SVM), logistic regression (LR), fully connected neural network (FCNN), back propagation neural network (BP) and K nearest neighbor (KNN). The dataset is divided into training set, validation set and test set, and the cross-validation method is used to optimize the model performance and ensure the reliability of the evaluation results.

[0081] Among them, common evaluation indicators such as accuracy, precision, recall, F1-score and area under the ROC curve (AUC-ROC) are used to comprehensively evaluate the model performance. The calculation formulas are as follows (4) to (7). Among them, M represents the number of links predicted correctly in the test set, N represents the total number of links in the test set, TP represents the number of links in the test set that are actually connected and correctly predicted, PN represents the number of links in the test set that are predicted to be connected, and TN represents the number of links in the test set that are not actually connected. By comparing the performance of different models under these evaluation indicators, the advantages and disadvantages of each model in the link prediction task are comprehensively evaluated, and finally the model with excellent performance under various indicators is selected as the link prediction model:

[0082]

[0083] Optionally, a multi-dimensional evaluation system for technology opportunities is constructed by combining the forecasting model, market demand, technology competitiveness and policy matching, including:

[0084] taking newly generated links in the document as technical opportunities and obtaining user comment data of the technical opportunities, and using a data decomposition algorithm to perform functional classification on the user comment data;

[0085] Performing sentiment analysis on the review data after the functional classification to obtain the number of negative reviews;

[0086] Calculate the similarity between the technology opportunity combination and the functional module, and complete the user comment-technology demand perception mapping through formula (8):

[0087]

[0088] Among them, w j is the similarity in the interval [0,1], r j is the number of negative comments corresponding to the functional module, MRD(·) represents the user demand intensity, and [Topic1,Topic2] represents the technology opportunity combination;

[0089] According to the document-topic probability distribution output by the LDA model, the topic with the highest corresponding probability in each document is identified, and the patent collection related to Topic 1 and Topic 2 is obtained. The number and quality of patents related to technical topics are used to evaluate the technical competitiveness of potential technical opportunities. The number of patents reflects the output probability of new technology research and development and innovative results, and the quality of patents is reflected by the number of patent citations.

[0090] By counting the number of patents and the number of citations corresponding to the patent set, the mapping between patents and technological competitiveness is completed through formula (9):

[0091]

[0092] Among them, P[Topic1,Topic2] represents the number of patents related to Topic1 and Topic2, Cit[Topic1,Topic2] represents the number of patent citations of Topic1 and Topic2, and TCD represents technological competitiveness.

[0093] In this embodiment, since the newly generated links are regarded as technological opportunities, but the actual value and feasibility of these opportunities need to be further evaluated, a potential technological opportunity identification method based on multi-dimensional index evaluation is further proposed, which will comprehensively consider the mutual adaptability between technological opportunities and market demand environment, technological attributes, and industrial policy environment, aiming to help enterprises grasp market dynamics more accurately, optimize technological strategic layout, and effectively utilize policy advantages to enhance the forward-looking and leading nature of industrial layout. The technology, market and policy opportunity window will provide enterprises with innovative resource support and improve the level of technological innovation risk-taking, promote enterprises to carry out innovative breakthroughs in key core technology fields and achieve sustainable development. Therefore, potential technological opportunities [Topic1, Topic2] are evaluated from three perspectives: market requirement degree (MRD), technological competitiveness (TCD), and policy matching degree (PMD).

[0094] It should be noted that in production practice, the market often leads the technology direction with keen technical perception. The rich comment data on the Internet not only expresses the user group's concern about product functions, but also reflects the degree to which the current product satisfies user needs. The negative feedback of users on a product or function is essentially caused by deep-seated technical deficiencies. Therefore, analyzing these feedbacks can reveal users' needs and expectations for technical improvements, thereby indicating the direction of technical research and development and innovation for enterprises. The present invention mines the market demand for potential technical opportunities based on user comment data. The technical competitiveness of potential technical opportunities is evaluated by the number and quality of patents related to technical topics. The number of patents reflects the output frequency of new technology research and development and innovation results, while the quality of patents is reflected by the number of patent citations. First, according to the document-topic probability distribution output by the LDA model, the topic with the highest corresponding probability in each document is identified and attributed to the topic, thereby obtaining the patent collection related to Topic 1 and Topic 2, further counting the number and number of citations of these patents, and completing the mapping of patent-technical competitiveness through formula (9).

[0095] Optionally, the policy match is used to measure the degree of attention paid to a specific technology field, the granularity of support, and the impact of the policy environment on technological innovation, so as to measure the key indicators of the fit between the policy environment and technological development. By constructing a mapping relationship between policy texts and technological topics, the granularity of policy support for potential technological opportunities is evaluated, including:

[0096] Preprocess the policy text and extract feature words, and calculate the TF-IDF value of the feature words;

[0097] According to the potential technology opportunities, the sum of the TF-IDF values ​​of the feature words related to the technology topic in the policy text is counted to quantify the policy matching degree. The corresponding calculation formula (10) is:

[0098]

[0099] Where n is the number of policy texts, KW topici The feature word set representing the technical topic i, TF-IDF (word, P j ) indicates that the feature word word is in the policy text P j The value of TF-IDF in .

[0100] In this embodiment, the policy matching degree reflects the degree of attention paid to a specific technology field, the strength of support, and the impact of the policy environment on technological innovation. As a key indicator for measuring the fit between the policy environment and technological development, it directly affects the R&D input and output of technological innovation. By constructing a mapping relationship between policy texts and technical topics, the degree of policy support for potential technological opportunities is evaluated. First, the policy text is preprocessed and feature words are extracted, and the TF-IDF value of the feature words is calculated; then, for each pair of potential technological opportunities, the sum of the TF-IDF values ​​of the feature words related to these technical topics in the policy text is statistically calculated; in order to quantify the policy matching degree and reflect the potential impact of the policy on technological innovation.

[0101] Optionally, the technical innovation direction of the multi-dimensional evaluation system of technical opportunities is determined in combination with sensitivity analysis to complete two-stage feature construction and link prediction, including:

[0102] The mean values ​​of the indicators corresponding to market demand, technological competitiveness and policy matching are used as thresholds, and potential technological opportunities are divided into four types to obtain the basis for classification;

[0103] Construct a three-dimensional indicator evaluation model based on the classification criteria, and visualize various technical opportunities;

[0104] The CRITIC algorithm is used to weight the indicators corresponding to market demand, technological competitiveness and policy matching, and the comprehensive evaluation score of each pair of potential technological opportunities is calculated by weighted summation. The corresponding formula (11) is:

[0105] Tod=w1MRD+w2TCD+w3PMD (11)

[0106] Among them, Tod represents the comprehensive evaluation score;

[0107] A single-factor sensitivity analysis is adopted, and the weights w1, w2, and w3 of the indicators corresponding to market demand, technological competitiveness, and policy matching are slightly disturbed. The evaluation results before and after the disturbance are compared to identify the indicators that are most sensitive to the results.

[0108] In this embodiment, the average values ​​of market demand, technological competitiveness and policy matching indicators are used as thresholds, and potential technological opportunities can be divided into four types. The division basis and combination types are specifically shown in Table 1.

[0109] Table 1 Meaning of identification results of various types of technology opportunities

[0110]

[0111] "Hotspot technology opportunities" are in a comprehensive advantage position with strong market demand, high technological competitiveness, and strong policy support. Such technology opportunities are usually the focus and development direction of the current market, and have high commercial potential and development prospects; "Market demand technology opportunities" have high market demand and technological competitiveness, but low policy matching, and need to cope with the challenges brought about by insufficient policy support; "Policy-driven technology opportunities" have high market demand and policy matching, but relatively insufficient industry activity. Such technology opportunities are often in the key core links of the industrial technology chain and require continuous investment to cope with technological frontier challenges and achieve bottleneck breakthroughs. Since it is still in the early stages of development, there is a high degree of uncertainty, but it also exhibits the characteristics of an emerging field; "Potential technology opportunities" have high technological competitiveness and policy matching. Although there is currently a lack of market demand, they still show high development potential under the industry's innovation momentum and policy support. According to the classification basis in Table 1, various types of technology opportunities are visualized by constructing a three-dimensional indicator evaluation model, such as Figure 5 As shown, the x-axis represents the market demand index, the y-axis represents the technological competitiveness index, the z-axis represents the policy matching index, and the red (A) (computer display color), green (B), yellow (C), and blue (D) scatter points represent hot spot type, market demand type, policy-driven type, and future opportunity type technological opportunities, respectively.

[0112] It should be noted that in order to further identify the trend of technological integration in key core technologies, refine the application scenarios of technological opportunities, and enhance the adaptability of strategies, the above indicators are weighted based on the CRITIC method. This method takes into account the comparison strength (i.e., standard deviation) and mutual conflict (i.e., correlation) of indicators, and can be more objectively weighted based on the information attributes of the indicators themselves, and the comprehensive evaluation score of each pair of potential technological opportunities is calculated by weighted summation. Single-factor sensitivity analysis is used to make small perturbations to the weights of market demand, technological competitiveness, and policy matching indicators, and compare the evaluation results before and after the perturbations to identify the indicators that are most sensitive to the results. This can not only verify the robustness and reliability of the model, but also provide more accurate and powerful support for the identification of technological integration trends in key core technologies. In addition, the results of sensitivity analysis can also help to better understand the role of each indicator in technological integration and provide a reference for policy formulation and strategic planning.

[0113] The present invention also provides an empirical analysis process, which is as follows:

[0114] (1) Data collection and preprocessing

[0115] Under the general trend of new energy and intelligent automobiles around the world, my country, as the world's largest automobile market, is becoming a strategic focus for major multinational auto companies, especially in the field of new energy vehicles. In order to identify new trends in technological development, grasp new trends in market consumption, and respond to new models of industrial development, the empirical analysis part of this invention selects patents, literature and social media data in the field of new energy vehicles for relevant explanations.

[0116] Using a certain patent database and a certain National Bureau of Statistics patent database as data sources, and "new energy vehicles or fuel cell vehicles or plug-in hybrid vehicles or hybrid vehicles or pure electric vehicles or electric vehicles" as keywords, patents that have been rejected, revoked, substantive reviewed, and disclosed inventions are excluded, and a total of 33,752 data on patent inventions in the field of new energy vehicles from 2010 to 2023 are obtained after merging; at the same time, a certain CNKI database is used as the data source, and the search keywords are the same as the patent search. The journals are limited to SCI, CSCD, CSSCI, and EI. After removing duplicate and blank irrelevant data, a total of 9,461 journal article data are obtained; in addition, using Python crawler technology, 12,771 pieces of social media information data on the field of new energy vehicles from the Electric State website were obtained.

[0117] (2) Life cycle division

[0118] This paper uses patent data sets to fit the Logistic life cycle curve through the grid search method according to the patent year and the cumulative number of patents. The grid width of parameter k is set to 0.01, and the grid width of parameter L is set to 1000. Through this method, the optimal parameters K = 1.1, t0 = 2016.14, c = 8316.42 are obtained. The fitting results are as follows Figure 6 As shown in the figure, based on the S-shaped curve and technology life cycle characteristics of new energy vehicle technology, its development process is divided into the following four stages: the embryonic stage, the development stage, the growth stage and the mature stage.

[0119] (3) Construction of technical topic co-occurrence network

[0120] In order to improve the complementarity and accuracy of multi-source data topic identification, a technical topic network construction method with differentiated preprocessing, source-based modeling, and step-by-step fusion is proposed. First, the patent and literature abstract texts and information data in each stage of the data set are preprocessed by noise filtering, definition of stop words, reserved word list, word segmentation, etc. Since patent texts usually contain a wealth of technical terms and literature tends to use academic vocabulary, the reserved word list is defined by referring to the standard "Electric Vehicle Terminology" and expert opinions; the information data involves news information and industry dynamics with strong timeliness, and focuses on extracting keywords related to consumer demand, market trends, and policy orientation. Secondly, the LDA model is independently run on papers, patents, and data to extract the topic keywords of each type of data. In addition, the Word2vec embedding technology is used to vectorize the topics of different data sources, and similar topics are aggregated based on cosine similarity calculation, and the comprehensive topic keywords and their distribution are regenerated. Finally, the technical topic network of each stage is constructed to integrate the advantages of each data source and avoid the interference that may be caused by direct fusion, thereby improving the comprehensiveness and consistency of the topic.

[0121] Based on the document-topic probability distribution output by the LDA model, the top three topics in each document are selected, and low co-occurrence topics with probability values ​​less than 0.2 are filtered out. The technical topic co-occurrence network of each stage is constructed according to the above co-occurrence intensity calculation formula (1). The number of nodes and edges in each stage of the network is shown in Table 2. Then, Gephi software is used for visualization, as shown in Figures 7 to 10 shown.

[0122] Table 2 Statistics of network structure at each stage

[0123]

[0124] (4) Link prediction model verification and evaluation

[0125] The present invention is based on the two-stage feature construction of GCN, uses a combination of multiple feature types for binarization processing to obtain link prediction results, and compares and verifies the prediction effects of the model with multiple benchmark methods, and analyzes the technical opportunity link prediction results.

[0126] 1. Two-stage feature construction based on GCN

[0127] 1) GCN node embedding. First, the preprocessed text is used as the training dataset of the Word2Vec model. Specifically, the Skip-Gram model is selected, the dimension of the word vector is set to 32, the window size is 5, the number of negative samples is 20, and 10 epochs are traversed to extract the semantic features of the topic text. Then, these semantic features are combined with the structural features of five network nodes such as degree centrality and closeness centrality to form a node feature matrix. Furthermore, the node feature matrix is ​​used as input together with the adjacency matrix with self-loops added to train a graph convolutional network (GCN) model built based on the PyTorch library. The model contains two GCNConv layers, Relu is used as the activation function of the first layer, and the second layer is followed by a binary cross entropy loss function (BCEWithLogitsLoss). 30 epochs are set during the training process, the Adam optimizer is used, and the learning rate is set to 0.01 to obtain the node embedding representation H i , the specific data shape and parameter settings are shown in Table 3 below.

[0128] Table 3 Data shape and parameter settings

[0129]

[0130] 2) Edge feature aggregation. Define 10 local similarity index calculation formulas to measure the similarity and association strength between nodes. After calculating these features, use StandardScaler to normalize them to eliminate the dimensionality effect between different features, and splice them into an initial edge feature matrix E. Secondly, use the Hadamard Product, that is, element-by-element multiplication, to combine the embedding vectors of the two nodes corresponding to each edge to obtain a new edge feature matrix h. ij , in order to effectively integrate the semantic and structural information of the node into the feature representation of the edge. Finally, the initial edge feature matrix E is combined with the new edge feature matrix h ij Splice to form a comprehensive feature matrix E combined .

[0131] 2. Experimental process and comparative verification

[0132] The link prediction model proposed in this paper cannot be compared and verified with other benchmark methods on public datasets because it involves multi-source data fusion and proprietary technology theme co-occurrence networks. Considering that technology opportunity identification requires link prediction for specific fields, and then in-depth discussion and analysis of technology evolution and innovation paths, comparative verification is carried out on the technology theme co-occurrence network in the growth stage to illustrate the effectiveness of the model. Furthermore, training and prediction are carried out on the current mature network to obtain technology opportunities that may appear in the future.

[0133] Table 4 shows the results of ML link prediction. Six machine learning models were used: random forest (RF), logistic regression (LR), k-nearest neighbor (KNN), support vector machine (SVM), fully connected neural network (FCNN), and naive Bayes (NB). To ensure accuracy, the networks in the development and growth stages were divided into training sets, test sets, and validation sets in a ratio of 6:2:2, where the test set was used for parameter optimization and the test set was used for performance evaluation. Ten random repeated tests were performed on each network, and the average value of each indicator was taken as the final result. The bold in the table indicates the best performance in the corresponding data set. It can be found that random forest (RF) showed the highest performance in both stages of the network, indicating that it has good generalization ability in the link prediction task, so this model was selected as the final mature network link prediction model.

[0134] Table 4 Link prediction algorithm performance comparison

[0135]

[0136] In order to deeply understand the impact of different feature types on the link prediction task, the performance of various machine learning models in the link prediction task was systematically analyzed and the F1 scores were calculated. According to the data in Table 5, when link prediction depends only on GCN node features, the overall performance of the model is not ideal. However, when the model predicts links based only on similarity metrics, the performance is improved by 10% on average. Furthermore, when edge relationships (i.e., GCN node features) are incorporated into the traditional similarity metrics, the overall performance of the model is significantly enhanced, and the performance of some models is improved by about 10%. Specifically, the random forest (RF) model performs best in the feature combination of comprehensive GCN node features and similarity features, with F1 scores above 0.9 on all stages of the network. This result highlights the powerful ability of RF in integrating multi-source features and capturing complex patterns. In addition, in the growing network, the logistic regression (LR), support vector machine (SVM) and naive Bayes (NB) models also show good performance under this feature combination; while in the developing network, the performance of the LR, FCNN and NB models is more prominent. However, in the growing network, k-nearest neighbor (KNN) and fully connected neural network (FCNN), and in the developing network, KNN and SVM performed best when using only similarity features. Overall, the model combining edge features and similarity measures showed the highest performance, emphasizing the importance of using multiple features in practical applications.

[0137] Table 5 Model performance under different input features

[0138]

[0139] 3. Technical opportunities and link prediction

[0140] The results of the above comparative experiments verify the feasibility and effectiveness of the two-stage feature construction link prediction model based on graph convolutional network (GCN) in the technical topic co-occurrence network. The model is used for further training and prediction in the mature network, keeping the basic parameter settings unchanged. After multiple iterations of optimization, the ROC curve and PR curve of the model on the test set are as follows: Fig.11 shown.

[0141] like Fig.11 As shown in Figure 2, the AUC value of the model on the test set of the mature network reached 0.9677, and the AP value was 0.9490, which shows that the model has excellent performance. The best threshold determined by the Youden index on the ROC curve is 0.62. Fig.11 The data of the mature network are input into the model. After two-stage feature construction, the model is combined with the random forest (RF) model to predict the links according to the threshold of 0.62. The number of links in the original network is 477. The remaining 513 pairs of technical topics that have not been connected are predicted to have 73 pairs of new links. Links with prediction scores greater than or equal to the threshold are new links, and those less than the threshold are predicted to disappear, thus obtaining the final prediction results of the model.

[0142] (5) Evaluation of potential technology opportunities

[0143] 1. Three-dimensional evaluation index system

[0144] For the market demand (MRD), firstly, the review data of the top three new energy vehicle models (HO, DM, TS) in April 2024 were crawled from a certain car platform, and a total of 67,850 reviews were obtained after merging and deduplication. Then, the review data was decomposed to decompose a single review data involving multiple functions into multiple short review sentences involving only one function, and the review data was functionally classified according to Table 6.

[0145] Table 6 Main functional modules of new energy vehicles

[0146]

[0147] Next, the Chinese sentiment analysis library SnowNLP is used to perform sentiment analysis on the comment sentences to obtain the number of negative comments on each functional module, i.e., the user demand intensity r, and the TF-IDF algorithm is used to extract the keywords of each functional module. The keywords of each functional module and the user demand intensity are shown in Figure 2. Fig.12 shown.

[0148] Since the self-trained Word2Vec model using a small amount of corpus may lead to large errors, the Elmo pre-trained language model

[47] of Harbin Institute of Technology is used to vectorize the topic-word distribution output by LDA, that is, the keywords of the technical topic nodes and the keywords of each functional module, and calculate the similarity w between each technical topic node and the functional module. Finally, the MRD of each potential technology opportunity is obtained according to formula (3).

[0149] For technical competitiveness (TCD), the patent set of each technical topic node is matched according to the document-topic probability distribution output by LDA, and TCD is obtained by formula (4).

[0150] For the policy matching degree (PMD), the policy document library was used as the data source to collect 112 policy documents and interpretations related to new energy vehicles. After preprocessing operations such as word segmentation and stop word removal, the TF-IDF method was used to extract keywords in the text, and the PMD was obtained through formula (5).

[0151] Each index is normalized by Min-Max to eliminate the dimension. Some calculation results are shown in Table 7.

[0152] Table 7 Calculation results of various indicators (partial)

[0153]

[0154] Taking the mean of each indicator in Table 1 as the threshold, the mean of market demand is 0.494, the mean of technological competitiveness is 0.3148, and the mean of policy matching is 0.4206. According to the identification rules in Table 3, the above 73 pairs of potential technology opportunities are visualized, as shown in Figure 3. Fig.13 shown.

[0155] 2. Interpretation of potential technological opportunities

[0156] Based on the three-dimensional evaluation indicators, a total of 34 pairs of technology opportunities of four types were identified. The following is an interpretation of the identification results to further understand the potential value of technology opportunities.

[0157] 1) Hot technology opportunities. Fig.13As shown in the figure, a total of 11 pairs of hot technology opportunities were identified. This type of technology opportunities are mainly concentrated in the field of battery and electric drive technology, including battery materials and thermal management technology, fast charging and thermal management technology, artificial intelligence technology and battery management system, renewable energy and on-board energy storage technology, power battery and chassis technology, battery management system and powertrain, motor and braking system, powertrain and wheel drive technology, powertrain and motor, high voltage platform and vehicle control system, autonomous driving and vehicle to everything (V2X) technology. The above technology opportunities have shown significant advantages in market demand, technological competitiveness and policy matching, and have mature technology and sufficient policy support to meet market demand, indicating that they have development prospects in improving the performance of electric vehicles, optimizing energy management, driving, etc. Relevant enterprises and technology research and development can give priority to the above opportunities for tracking and investment.

[0158] 2) Market demand-based technology opportunities. There are 7 pairs of technology opportunities of this type, mainly concentrated in the construction of consumer service systems. They cover intelligent service experience and Internet of Vehicles technology, human-computer interaction technology and intelligent service experience, smart cockpit technology and intelligent service experience, brand and performance and human-computer interaction technology, wireless charging and artificial intelligence technology, charging equipment technology and intelligent service experience, vehicle interior and exterior design and intelligent service experience. This is consistent with the core needs of consumers for new energy vehicles mentioned in the report, namely the pursuit of high performance, intelligent technology and a complete service system construction, especially improving battery life and charging speed, intelligent driving assistance functions and high intelligence of vehicle systems. The above-mentioned technology opportunities have great potential in terms of market demand and technological competition, but it is necessary to pay attention to policy trends in a timely manner for adjustment and matching.

[0159] 3) Policy-driven technology opportunities. There are 7 pairs of technology opportunities of this type, mainly concentrated in the fields of infrastructure and hydrogen energy technology, including fast charging technology and smart charging network, integration of smart network connection and charging infrastructure, fast charging and charging infrastructure, charging equipment and charging infrastructure, hydrogen fuel cells and electric motors, hydrogen fuel cells and on-board energy storage technology, and energy recovery systems and hydrogen fuel cells. It is necessary to integrate new energy, new materials, the Internet, big data, artificial intelligence and other transformative technologies, and accelerate the core technology research and development projects of battery technology, smart network connection technology, and basic technology of new energy vehicles (including automotive-grade chips, automotive operating systems, high-efficiency and high-density drive motors, hydrogen fuel cells, etc.). The above-mentioned technology opportunities are not yet mature in terms of technology. However, support in policies, regulations and funds as well as high market demand will help accelerate the research and development and application of these technologies. Relevant companies should pay close attention and actively use policy support to promote technology development and market promotion in these fields.

[0160] 4) Potential technology opportunities. There are 9 pairs of technology opportunities of this type, mainly concentrated in the field of vehicle operating system (VOS), involving comprehensive control and management of vehicles, including hardware and software integration, data processing, safety and user interface. Specifically, it includes the integration of autonomous driving and smart cockpit technology (cockpit integration), autonomous driving and automatic parking system integration (driving and parking integration), smart cockpit and automatic parking system integration (cabin and parking integration), automotive chips and human-computer interaction technology, vehicle control system and smart cockpit, chassis and powertrain, smart cockpit and powertrain, Internet of Vehicles technology and smart cockpit, intelligent service experience and human-computer interaction technology integration. Although the current market demand has not been fully released, the above technology opportunities still show significant advantages in terms of technological competitiveness and policy support. With the maturity of technology and the gradual acceptance of the market, these technologies are expected to usher in broad development prospects in the future. At the same time, the opportunities for future technological development will focus on battery technology, intelligent network technology, low-carbon technology, and the improvement of user experience and after-sales service under the omni-channel customer direct connection mode (DTC). As the core carrier of these technologies, VOS will play a key role in promoting the automotive industry to develop in a more efficient, intelligent and environmentally friendly direction.

[0161] In summary, it can be found that the technological opportunities identified by the present invention are basically consistent with the future technological development direction indicated by relevant policy planning and industry reports, thus verifying the effectiveness of this method.

[0162] (6) Sensitivity analysis

[0163] First, the CRITIC method is used to assign weights to each indicator through Python, and the values ​​of w1, w2, and w3 are 0.396, 0.303, and 0.301 respectively, thus obtaining the comprehensive evaluation formula (12):

[0164] Tod=0.396MRD+0.303TCD+0.301PMD (12)

[0165] To further explore the impact of each indicator on the comprehensive score, assuming that other variables remain unchanged and a certain variable changes within the range of ±10%, a single factor sensitivity analysis was performed. The results are as follows: Fig.14 shown. Fig.14 The results show that the market demand variable has a greater impact on the comprehensive score than the technical competitiveness and policy matching variables, and the market demand is more sensitive to changes in the comprehensive score than the other two variables. It also shows that the trend of new energy vehicle technology integration is greatly affected by the market.

[0166] Next, we selected the top five potential technology opportunities with the highest comprehensive scores for analysis, and increased the weight of each variable by 10% to analyze the specific impact of each indicator on the technology opportunity. Fig.15 shown.

[0167] from Fig.15 It can be observed that the characteristics of different technology opportunities vary significantly, but there are also commonalities. Based on these characteristics, technology opportunities can be further divided into the following application scenarios: typical hot technology, task-oriented technology, and future technology. The specific analysis is as follows:

[0168] 1) Typical hot technologies. Battery materials and thermal management technology (technical opportunity 1) has the highest comprehensive score. When the weights of market demand, technological competitiveness and policy matching indicators are increased by 10%, the comprehensive scores are improved and the differences are not obvious. This reflects that the impact of each indicator on the technology opportunity combination is relatively positive and balanced, and also shows that the technology combination meets market demand, has high market potential, has significant competitive advantages in technological innovation, and has a strong driving effect of policy support. The integration of the two technologies needs to focus on the selection and optimization of materials, carry out research on key core technologies such as positive and negative electrode materials, electrolytes, diaphragms, and membrane electrodes, strengthen the short-board technology of high-strength, lightweight, high-safety, low-cost, and long-life power batteries and fuel cell systems, and the research and development and industrialization of solid-state power battery technology; followed by autonomous driving and vehicle networking (V2X) technology (technical opportunity 2), and the impact of each indicator is basically the same as the former. The integration of the two technologies aims to form an intelligent, low-carbon, safe, and efficient travel ecosystem to promote technological innovation and industrial coordinated development. We need to further strengthen our ability to tackle key core technologies in the areas of high-performance chip research and development, autonomous vehicle computing platforms and operating systems, cloud control platform and communication technology adaptability, and V2G energy optimization management.

[0169] 2) Task-oriented technology. When the weight of market demand and policy matching increased by 10%, the comprehensive score of renewable energy and energy recovery system technology (technical opportunity 3) increased. However, when the weight of technical competitiveness increased by 10%, the comprehensive score decreased instead. The increase in its weight did not bring the expected positive effect. This shows that technological progress and market demand have not been matched synchronously. On the one hand, it may be that the increase in technology investment and complexity has led to rising costs, and the user experience has not been improved synchronously, affecting user acceptance. On the other hand, it may be that policy support has not been fully released on the market demand side. Future technological opportunities can be directed towards the development of high-efficiency solar cells, advanced energy recovery systems, and the coordination of wind power generation and photovoltaic power generation technologies, as well as the construction of wind and solar charging stations, to accelerate the integration of the two technologies. At the same time, it is necessary to control R&D costs, optimize designs to balance technological progress, improve user experience, promote the formulation of relevant technical standards and infrastructure construction, so as to improve the market adaptability and popularity of technology.

[0170] 3) Future technologies. The characteristics of hydrogen fuel cells and hydrogen refueling station infrastructure (Technology Opportunity 4) show that after the market demand weight increases by 10%, the comprehensive score decreases instead, which indicates that the increase in market demand is difficult to be converted into a higher comprehensive score in the short term, suggesting that market demand growth faces challenges in the short term, which may be caused by the low market acceptance of hydrogen energy vehicles and the insufficient number of hydrogen refueling stations. On the contrary, the improvement of technical competitiveness and policy matching can significantly increase the comprehensive score, indicating that technological progress and policy support are key factors in promoting technological integration in this field. Future technical directions need to focus on technological breakthroughs in supporting infrastructure such as hydrogen production, storage, transportation, and hydrogenation. Further improve technologies such as gas hydrogen and liquid hydrogen storage and transportation technology, hydrogen liquefaction, hydrogen compressors, and hydrogen refueling machines. By reducing costs and improving safety, strengthening technical integration, and thus gradually increasing market demand; Finally, V2G and smart charging network technology (Technology Opportunity 5) has similar characteristics to Technology Opportunity 4, but the comprehensive score is slightly lower. This shows that consumers' awareness of V2G and smart charging network technology is still insufficient, and the popularity of related infrastructure (such as two-way charging piles) needs to be improved. To promote the realization of this technological opportunity, future technological innovation, infrastructure construction and policy support are particularly critical. At present, the main difficulty faced by the promotion of V2G technology and the integration of smart charging networks lies in the construction of early charging infrastructure. According to relevant planning guidance, it is necessary to strengthen the energy interaction between new energy vehicles and power grids (V2G) and promote the efficient coordination of new energy vehicles and renewable energy. In addition, how to integrate artificial intelligence technologies such as the Internet of Things, cloud computing, big data and large models to achieve efficient energy conversion and storage, intelligent energy management systems and advanced power grid dispatching strategies is also a key issue that needs to be solved.

[0171] In summary, due to the limitations of the existing technology opportunity identification research, such as insufficient data source coverage, insufficient characterization of technology theme network attributes, and single technology opportunity evaluation dimension. Starting from the perspective of technology integration, this paper is driven by multi-source data and proposes a technology opportunity identification path based on GCN two-stage feature construction for link prediction and multi-dimensional evaluation. This path includes three stages: life cycle division and multi-source data technology theme network construction, GCN-based link prediction model construction, and technology opportunity identification based on multi-dimensional indicator evaluation and sensitivity analysis: literature, patents, and social media data are used as data sources to characterize the theme network in a fine-grained manner; the proposed GCN-based two-stage feature construction link prediction model deeply captures the global semantic information and topological characteristics of the network, effectively improving the prediction accuracy; the multi-dimensional indicator evaluation model is combined with sensitivity analysis, based on the technology itself, to further explore market and policy signals. The empirical results show that 4 types of typical hot spots, task-oriented, policy-oriented and future-oriented technology opportunities in the field of new energy vehicle technology are identified, and 34 technology opportunities are subdivided into application scenarios for 5 technology opportunities.

[0172] See also Fig.16 The present invention also provides a two-stage feature construction and link prediction device based on GCN, which is applied to the above-mentioned two-stage feature construction and link prediction method based on GCN, including:

[0173] An LDA model building unit, used to obtain the life cycle of a preset technical field in multi-source data and build an LDA model of the life cycle, wherein the LDA model includes a document topic model based on probability and statistical methods;

[0174] A prediction model construction unit, used to construct a technical theme co-occurrence network of the preset technical field according to the LDA model, and to construct a two-stage feature construction link prediction model based on GNC according to the technical theme co-occurrence network, wherein the two-stage feature construction link prediction model based on GNC includes node embedding and edge feature aggregation based on GCN;

[0175] The two-stage feature construction and link prediction unit is used to construct a multi-dimensional evaluation system for technology opportunities by combining the prediction model, market demand, technology competitiveness and policy matching, and to determine the technology innovation direction of the multi-dimensional evaluation system for technology opportunities by combining sensitivity analysis to complete the two-stage feature construction and link prediction.

[0176] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A two-stage feature construction and link prediction method based on GCN, characterized in that: The following steps are involved: Acquire the life cycle of a preset technical field in multi-source data, and construct an LDA model of the life cycle, wherein the LDA model includes a document topic model based on probability and statistical methods; Constructing a technical theme co-occurrence network of the preset technical field according to the LDA model, and constructing a two-stage feature construction link prediction model based on GNC according to the technical theme co-occurrence network, wherein the two-stage feature construction link prediction model based on GNC includes node embedding and edge feature aggregation based on GCN; A multidimensional evaluation system for technology opportunities is constructed by combining the prediction model, market demand, technological competitiveness and policy matching, and the technological innovation direction of the multidimensional evaluation system for technology opportunities is determined in combination with sensitivity analysis to complete the two-stage feature construction and link prediction.

2. The two-stage feature construction and link prediction method based on GCN according to claim 1, characterized in that: Obtain the life cycle of new energy technology fields from multi-source data and construct an LDA model of the life cycle, including: The LDA model includes a Bayesian probability model of a three-layer structure of word-topic-document, and explores potential topic structures from documents. The LDA model is used to extract technical topics and keywords to achieve efficient clustering of documents and terms. The process of extracting technical topics by the LDA model includes: Get the hyperparameters α, β and the number of topics K, randomly assign a topic to each word w and initialize the topic assignment matrix z; Iterative updates are performed through Gibbs sampling to calculate the conditional probability of each word in document D being assigned to each topic, and the topics are reallocated. After multiple iterations, the document-topic distribution θ and topic-word distribution are obtained. The perplexity index and consistency score are combined to judge the model performance under different numbers of topics. The number of topics with the lowest perplexity index and the highest consistency score is taken as the optimal number of topics K. The LDA model is trained and the opinions of domain experts are collected to comprehensively evaluate the clustering effect of the LDA model to extract the corresponding technical topics.

3. The two-stage feature construction and link prediction method based on GCN according to claim 1 is characterized in that: The technical topic co-occurrence network in the field of new energy technology is constructed according to the LDA model, including: Phase 1: Construct a feature matrix from the semantic and structural perspectives, and input the feature matrix together with the adjacency matrix with self-loops into GCN for node embedding learning. In the propagation process, global information is captured by layer-by-layer aggregation. The second stage: Use binary operators to construct variable features to enhance the model's edge feature expression ability, combine the edge features generated by GCN embedding with the local similarity index features to form a comprehensive edge feature matrix; The constructed technical topic co-occurrence network defines positive links and negative links. The link that exists between two topics is defined as a positive link, and the link that does not exist between two topics is defined as a negative link. The undersampling method is selected to balance the number of positive links and negative links.

4. The two-stage feature construction and link prediction method based on GCN according to claim 3 is characterized in that: The feature matrix and the adjacency matrix with self-loops added are input into GCN for node embedding learning, including: The construction of node features and the generation of embedded representations are performed through convolution operations to learn the connection information of neighboring nodes. The embedding vector of the node is induced according to the neighborhood attributes of the node. A graph G(V,E) is preset, where V(|V|=n) and E are the sets of nodes and edges respectively. The new k-dimensional node feature matrix is ​​calculated based on the single-layer GCN as follows: Among them, H (l) is the node feature matrix of the lth layer, A is the adjacency matrix with self-loops added, D is the diagonal matrix of node degrees, and W (l) is the learnable weight matrix of the lth layer, σ is the activation function; GCN embedding includes an adjacency matrix and a feature matrix. The adjacency matrix is ​​an n×n square matrix used to represent the connection relationship between nodes in the graph. If there is an edge between two nodes, the corresponding element in the adjacency matrix is ​​represented by 1; if there is no connection between two nodes, the corresponding element in the adjacency matrix is ​​0; The construction of the adjacency matrix includes: Starting from the document-topic probability matrix D output by the LDA model, the matrix P is obtained by filtering low-probability topics by setting the Top-N standard and threshold; The co-occurrence relationship between topics is determined according to the calculation formula (2) of the co-occurrence intensity, and the document-topic probability matrix D is converted into the trotter-topic matrix C; The co-occurrence intensity is converted into a binary relationship to obtain the adjacency matrix U for GCN. When GCN performs feature propagation, the adjacency matrix is ​​updated by adding self-loops according to the characteristics of the node itself, and all elements on the diagonal of the adjacency matrix are set to 1. The calculation formula (2) is: Among them, W ij is the co-occurrence intensity between topic i (i∈(1,2,...,M)) and topic j (j∈(1,2,...,M)), T ai is the topic i in the a(a∈(1,2,...,N))th document, P(T ai ) is the probability value of topic i appearing in document a in the document-topic matrix, P(T aj ) is the probability value of topic j appearing in document a in the document-topic matrix.

5. The two-stage feature construction and link prediction method based on GCN according to claim 4 is characterized in that: The construction process of the feature matrix includes: The Word2Vec model is used to represent words as vectors of a specified length and capture the semantic relationship between words. The capture tools include the CBOW model and the Skip-Gram model. First, use the Skip-Gram model to extract the 32-dimensional feature vector of each technical topic node to obtain the semantic information between nodes; The importance of nodes is evaluated by using multiple indicators that reflect the structural characteristics of nodes, such as degree centrality, closeness centrality, betweenness centrality, clustering coefficient, local clustering coefficient, node betweenness, and node efficiency; Each feature is concatenated to form a node feature matrix and normalized; In the GCN model, two hidden GCNconv layers are used to process graph data. The first hidden layer is followed by a ReLU activation function to increase the nonlinear expression ability of the model. The output of the second hidden layer is directly used to calculate the loss. The loss function uses BCEW, which includes a sigmoid activation function. The sigmoid activation function is used to interpret the model output as probability. The expression of the GCN model is: Z=AReLU(AXW (0) )IN (1) (3) in, is the normalized adjacency matrix, W (0) ∈X N×H is the trainable weight matrix of the first layer, H is the dimension of the hidden layer, W (1) is the node embedding dimension after the first layer of convolution operation, X is the node feature matrix, and Z is the feature matrix of the final output.

6. The two-stage feature construction and link prediction method based on GCN according to claim 3 is characterized in that: The node embedding vector H is obtained based on the node embedding representation obtained in the first stage to construct the edge features. i Based on the representation, the binary operator Hadamard product is applied to embed the two nodes corresponding to each edge into the embedding vector h i 、h j Multiply element by element to get the edge feature matrix h ij , where the node embedding vector H i The dimension is N×D, where N is the number of nodes and the feature matrix h ij The dimension is M×D, where M is the number of node pairs; Use multiple local similarity indicators to construct the feature matrix of the edge. The local similarity indicators are used to capture the similarity and relationship strength between nodes to measure the similarity and integration trend between technologies. The feature matrix E of the edge is calculated and each feature is normalized using the StandardScaler function. The dimension of the feature matrix E is M×F, where F is the number of node pairs with similarity and relationship strength. The edge feature matrix h ij Concatenate with E to form a comprehensive feature matrix E combined .

7. The two-stage feature construction and link prediction method based on GCN according to claim 1, characterized in that: Combining the forecasting model, market demand, technology competitiveness and policy matching, a multi-dimensional evaluation system for technology opportunities is constructed, including: taking newly generated links in the document as technical opportunities and obtaining user comment data of the technical opportunities, and using a data decomposition algorithm to perform functional classification on the user comment data; Performing sentiment analysis on the review data after the functional classification to obtain the number of negative reviews; Calculate the similarity between the technology opportunity combination and the functional module, and complete the user comment-technology demand perception mapping through formula (8): Among them, w j is the similarity in the interval [0,1], r j is the number of negative comments corresponding to the functional module, MRD(·) represents the user demand intensity, and [Topic1,Topic2] represents the technology opportunity combination; According to the document-topic probability distribution output by the LDA model, the topic with the highest corresponding probability in each document is identified, and the patent collection related to Topic 1 and Topic 2 is obtained. The number and quality of patents related to technical topics are used to evaluate the technical competitiveness of potential technical opportunities. The number of patents reflects the output probability of new technology research and development and innovative results, and the quality of patents is reflected by the number of patent citations. By counting the number of patents and the number of citations corresponding to the patent set, the mapping between patents and technological competitiveness is completed through formula (9): Among them, P[Topic1,Topic2] represents the number of patents related to Topic1 and Topic2, Cit[Topic1,Topic2] represents the number of patent citations of Topic1 and Topic2, and TCD represents technological competitiveness.

8. The two-stage feature construction and link prediction method based on GCN according to claim 7, characterized in that: Policy matching is used to measure the degree of attention paid to a specific technology field, the granularity of support, and the impact of the policy environment on technological innovation, which is a key indicator for measuring the fit between the policy environment and technological development. By constructing a mapping relationship between policy texts and technical topics, the granularity of policy support for potential technological opportunities is evaluated, including: Preprocess the policy text and extract feature words, and calculate the TF-IDF value of the feature words; According to the potential technology opportunities, the sum of the TF-IDF values ​​of the feature words related to the technology topic in the policy text is counted to quantify the policy matching degree. The corresponding calculation formula (10) is: Where n is the number of policy texts, KW topici The feature word set representing the technical topic i, TF-IDF (word, P j ) indicates that the feature word word is in the policy text P j The value of TF-IDF in .

9. The two-stage feature construction and link prediction method based on GCN according to claim 8, characterized in that: Combined with sensitivity analysis, the technical innovation direction of the multi-dimensional evaluation system of technical opportunities is determined to complete the two-stage feature construction and link prediction, including: The mean values ​​of the indicators corresponding to market demand, technological competitiveness and policy matching are used as thresholds, and potential technological opportunities are divided into four types to obtain the basis for classification; Construct a three-dimensional indicator evaluation model based on the classification criteria, and visualize various technical opportunities; The CRITIC algorithm is used to weight the indicators corresponding to market demand, technological competitiveness and policy matching, and the comprehensive evaluation score of each pair of potential technological opportunities is calculated by weighted summation. The corresponding formula (11) is: Tod=w1MRD+w2TCD+w3PMD (11) Among them, Tod represents the comprehensive evaluation score; A single-factor sensitivity analysis is adopted, and the weights w1, w2, and w3 of the indicators corresponding to market demand, technological competitiveness, and policy matching are slightly disturbed. The evaluation results before and after the disturbance are compared to identify the indicators that are most sensitive to the results.

10. A two-stage feature construction and link prediction device based on GCN, characterized in that: The two-stage feature construction and link prediction method based on GCN as claimed in any one of claims 1 to 9 comprises: An LDA model building unit, used to obtain the life cycle of a preset technical field in multi-source data and build an LDA model of the life cycle, wherein the LDA model includes a document topic model based on probability and statistical methods; A prediction model construction unit, used to construct a technical theme co-occurrence network of the preset technical field according to the LDA model, and to construct a two-stage feature construction link prediction model based on GNC according to the technical theme co-occurrence network, wherein the two-stage feature construction link prediction model based on GNC includes node embedding and edge feature aggregation based on GCN; The two-stage feature construction and link prediction unit is used to construct a multi-dimensional evaluation system for technology opportunities by combining the prediction model, market demand, technology competitiveness and policy matching, and to determine the technology innovation direction of the multi-dimensional evaluation system for technology opportunities by combining sensitivity analysis to complete the two-stage feature construction and link prediction.

Images